The transformer based approach is seen as a highway to race down in terms of training and building better models with minimal organizational distractions allowed while every company puts its foot down on this highway. Exploratory research into other AI methods is highly discouraged for now and organizations are encouraging transformer based scale out and iteration as a top priority.
As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models. They got someone willing to scale the current method out. Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
I have the view this makes sense too. Current top models are mere (mere in big datacenter terms) trillions of parameters right now. We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model. We'll soon start hitting 10's of trillions and 100's of trillions of parameter models. Unless there's a ceiling which nobody expects they should keep scaling just as they have been. I really do agree that it'd be insane to veer off this highway with exploratory research at this particular point in time. Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
> As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models.
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
Historically, that happens now and then. Examples include color television cameras and fluorescent lamp ballasts.
RCA developed the first good color TV cameras. They were monsters. Originally, electronic TV (as opposed to mechanical scanning) used Zworykin's iconoscope. This needed huge amounts of light. It didn't have a photo-multiplier stage to boost sensitivity, but it did accumulate charge from light over the whole frame time. There was an alternative approach, Farnsworth's image dissector, which also needed huge amounts of light. It had a photo-multiplier stage but only the light from the dot currently being scanned by the scanning beam mattered.
So RCA threw money at the problem. They developed the image orthicon camera tube. It combined the advantages of both approaches, integration over a full frame and a photo-multiplier stage. It was insanely expensive, costing thousands of dollars in the 1950s. It needed about twenty different voltages and multiple independent power supplies. It worked reasonably well.
Then came color. That took three image orthicons (RGB) inside a box the size of a small refrigerator, along with optics to split the incoming image into three colors. Plus
a rack of analog electronics per camera, connected through a large cable. Plus an alignment problem from hell, with a huge number of adjustment knobs.[1]
Scaling worked! From the 1950s to the 1970s, color TV broadcasting was powered by those monsters.
Now, independent of this, other people were working on other approaches. Vidicons, silicon detectors inside camera tubes, and finally fully solid state image sensors. Eventually, in
1983, after thirty years of image orthicon monsters, RCA produced a usable solid state camera. Today, a cell phone camera costs about US$10 in bulk, and it's far better than the monsters.
It took thirty years of low-priority development to get past the giant image orthicon cameras.
Electronic fluorescent lamp ballasts had a different problem. That was a suppressed invention.
The magnetic ballast industry was dominated by Magnetek. Someone invented a solid state replacement, which was smaller and cheaper, and licensed their patent to Magnetek for royalties. Magnetek just sat on it, not wanting to gut their core market. The inventor hired Townsend, Townsend and Crew in Palo Alto and sued. (Read this in the law firm's brag book in their office.) Eventually, the inventor got a hundred million or so, solid state ballasts went into production, and Magnetek seems to be gone. (The company called Magnetek today, which makes crane controls, is an offshoot of a Litton unit, and seems to be unrelated.)
There are also parallels from the steam locomotive industry, the Baldwin Locomotive Works being the best known. Baldwin actually built the first Diesel locomotive. They made a few half-hearted tries at Diesels over several decades, then gave up and went back to steam. Baldwin ceased all locomotive production in 1956. The corporation lingered on for a while, not doing much.
(So it must be those Orthicon cameras which cause the distinctive weirdnesses of early TV images like the reflections from highly-polished surfaces (such as the bells of brass instruments and saxophones) which gleam black? https://www.youtube.com/watch?v=2uTQQKU6ge0&t=111s )
Nice, it got a press mention, with a quote from Crew, one of the law firm's partners.
Suppressed inventions are very rare, but that's one of them.
One of the few others is Airadar, from 1973.[1] Not "AI", "air radar". This
was a phased-array radar for small aircraft. Way ahead of its time. So far
ahead that the USAF was upset about it being available commercial before it
was in use by the military.
Suppressed inventions aren't "very rare" per se. You simply don't hear about those that aren't being rescinded while the inventor is still alive.
According to official data from the U.S. Patent and Trademark Office (USPTO), there were 6,653 active secrecy orders in effect at the end of fiscal year 2026. Roughly 100 to 350 new secrecy orders are added every year. Many remain locked away for decades or more.
If the U.S. government puts a secrecy order on a patent application against the inventor's will, they are legally forbidden from talking about it. The inventor cannot talk about how it works, publish research on it, sell it, or market it. They cannot even discuss the exact classified details with a regular lawyer unless that attorney has the proper security clearance.
For as long as the big companies are committed to focusing on scaling LLMs it does seem reasonable that they don't want their overall efforts led by people like LeCun or Hassabis who don't particularly want to scale LLMs. But it does feel as if there's something a bit extra about many LLM maxis' active hostility to deviationists.
I think Hassabis’ position is far more nuanced than LeCun’s. From what I can see, LeCun believes LLMs are a dead end for AGI while Hassabis position seems to be that some new technique is needed combined with LLMs.
JEPA's method isn't dead, non-generative self-supervised learning that predicts abstract representations (embeddings) of missing information. It is understanding latent structures rather than creating content.
Even if he was wrong about those statements, exploring other types of models isn't necessarily a bad thing. Especially if we can find ones that perform well with less demanding hardware requirements.
Its not a bad thing, but exploratory research as opposed to reaearch that finetunes along a path already paved by exploratory research is something that doesn’t usually happen much in for-profit product-oriented firms when there is such a path available; it happens in academia and corporate labs pick it up once it reaches the point where the basics show commercial promise. CS is a little bit of an exception because the capital cost of basic research isn’t, proportionally, as high, and because the promise-to-product path is often short enough that having done the basic work in house can be a meaningful advantage in getting to market first (though, see Google on much of modern AI for a counterexample).
But even in computing, when there is a promising way forward that the basic reaearch has been done on, its hard to justify side quests.
Those exploratory "side quests" in industry are how we got the transistor, the liquid crystal display, Shannon's information theory, the laser and practical photovoltaic cells, among many other foundational technologies upon which we rely today.
Blue LEDs have been a massive net negative for humanity. Blue light hazard is very real (why blue lasers kill eyesight instantly and green and red often don’t in accidental exposure).
All display technology past CRTs has been a mistake. The vectrex in the 1980s had better refresh rates, color fidelity, beautiful phosphor glow, etc - and we abandoned it for a horrible pixel based future we are now in. We should return to vectors and lasers.
> and corporate labs pick it up once it reaches the point where the basics show commercial promise.
How does that work when the research team moved elsewhere? Re-hire talent? That takes time - the 1 thing you don't have in a fast-moving field like AI. Post sums it up nicely:
"Do the math—how much is it worth to halve training and inference cost given that the industry has investments approaching a trillion dollars?"
How much is these researchers' salaries, compared to what MSFT spends on compute? Chances are it's peanuts in comparison.
Yes but you forget that spending on compute is spending on machines that the company owns while spending on researchers is spending on other apes that trigger the social hierarchies of executives brains and who have the dreaded key word of “rights”
Companies will try to ruthlessly rid themselves of labor as long as those members of labor have agency and rights.
If someone disagrees they can convince me by explaining why a company wouldn’t choose to spend X on a machine they wholly own vs X on an employee they don’t.
There’s some room for argument that employees can do more than machines can but in the current AI zeitgeist arguing that we have machines with equivalent to human intelligence I don’t see how you wouldn’t see this coming if you believed the zeitgeist.
If compute:people were roughly equal spend, I'd agree with that thinking. But we all know AI's 100s of billions $$ spend is on datacenters+GPUs, not people.
In a way this suggests that MS doesn't aim to be a leader in AI <the core tech>. Only in its applications + related services (read: profit to be had there).
Nothing wrong with that btw. But understandable that fundamental researchers then pack up & leave.
> But we all know AI's 100s of billions $$ spend is on datacenters+GPUs, not people.
That’s bolstering my argument. Companies are willing to spend a shit ton on machines. They are not only unwilling to spend on labor but are actively seeking ways to remove labor spend, even though this new technology supposedly makes every unit of labor significantly more productive and thus profitable.
This is a cultural choice not an economic one.
If I see a company laying off people while admitting they’ve maximized a niche and automation lets them increase profitably, or I see them go on a hiring spree because labor is now netting them 10x-100x-1000x-whatever productivity increase they claim, then I will believe them about the value add of AI.
Currently I see companies using AI’s productivity increase as a reason to cut labor and achieve roughly the same productivity.
They aren’t competing, they just want to cut out workers.
> CS is a little bit of an exception because the capital cost of basic research isn’t, proportionally, as high, and because the promise-to-product path is often short enough that having done the basic work in house can be a meaningful advantage in getting to market first (though, see Google on much of modern AI for a counterexample)
...And because they have monopoly money. Basically, research labs are usually vanity spending for corporations with too much money. Xerox parc was funded by a monopoly. Bell labs was funded by a monopoly. Google funds its moonshots with its monopoly.
The issue isn't the exploration itself, it's the opportunity cost. When a figure of Yann LeCun's influence pushes a specific direction, it pulls key talent and resources away from mainstream LLM work. That kind of organizational distraction has a massive cost, and would've stopped Meta from delivering a polished product like Muse sooner.
The mainstream is wrong and in a lot of ways hazardous. He had every right to say he wasn't willing to sign is name on the wrong technology when he new better. When you hire smart people you listen to them. Elsewise you get what you deserve.
He was against LLMs, reinforcement learning, etc. Claiming that those are a dead-end. He was wrong. His own "JEPA" ideas have basically gone nowhere. Meta was right in moving on without him.
Ok, so we're talking about applying connectionist systems/models to the messiness of the real world. And "trust" isn't an on-off switch. The goal is to build systems that are smarter and more general. It's a process. I don't think even Yann LeCun has made any strong claims beyond that. He just thinks that his way of doing deep learning is better.
It is because even if it would pan out, it would mean temporarily losing AI leadership, reducing stock price, losing the ability to hire talent, and by virtue of that losing the opportunity permanently.
LeCun wasn't wrong. He simply had higher standards for results and accountability that LLMs are, by design, incapable of reaching.
Zuckerberg wanted to placate his shareholders with a shiny new thing now.
would've stopped Meta from delivering a polished product like Muse sooner.
Muse is not polished, unless usability and programming standards have fallen so far in the LLM vibe coding era that a pre-alpha state is now considered polished.
You will find brilliant scientists in every discipline who are not the least bit interested in a particular branch of research because they/their lab didn’t invent it. That’s fine for public research but not helpful private research.
Odd. As I see it he was one of the few right about everything. This is also taking his comments about transformers scaling out of context. He specifically said they won’t lead to human level ai and they haven’t, despite exhausting the training data.
AGI will require architecture changes not simply more data.
It is very jagged and some people don't consider that human level AGI - it can solve Maths at the genius level just like Deep Blue could play chess at the grandmaster level. At this point it is simply an opinion and I can see the other side too.
That could be how human intelligence stacks, search retrieval (LLM) with learned approaches to processing that info (harness). Maybe you are insulated from the average person. The frontier models easily surpass their intellectual ability to absorb new information and apply it to novel situations.
The average person will also hallucinate things and make all sorts of causality mistakes without realizing it AND refuse evidence of such whereas models tend to acquiesce to evidence. A model with the right harness will outperform any of these people at their current desk jobs. To me that is general intelligence. Most of us on HN tend to far surpass this general intelligence and likely experience a false consensus or assumed similarity bias.
> This is also taking his comments about transformers scaling out of context
There have been significant improvements in architectural changes, but since "pre-training" isn't being openly published people just lazily attribute it to "scale". Scale of what? Scale
> AGI will require architecture changes not simply more data.
As Sutton claimed, all along[0]. People read what they wanted to read. Unfortunately, they also push against researchers who were trying to explore other architectures and methods. I honestly believe that has held us back. There was always room for people to explore different paths. The problem was, and has always been, preventing or discouraging others from exploring new paths. It's ironic given that one of the most common reasons a paper is rejected is for "lack of novelty"[1]. But if someone does something novel then the reviewer simply compares it to the existing state of the art, which is an unfair comparison[2].
It's silly since people want to move fast, but we're really not good at recognizing what has actually allowed us to move forward in the past.
[1] An essentially meaningless statement. Near meaningless because it is not objective.
[2] You compare a novel technique against a technique that has had hundreds or thousands of researchers optimizing over several years. That's not even remotely fair.
100%. You can both commercialize transformers while still exploring new paths. Having one of the world's top researchers doing productization work is a huge waste of their time.
You also see this whenever news about novel but small models comes out. People are instantly dismissive. "Oh it's small, who cares?" Of course it's small, why waste resources testing it on a larger scale before making sure it works well on a small scale? Would people rather news not come out until it's SOTA model?
Absolutely agree fwiw. Scale out over LeCunns desired exploratory research was absolutely the right call. Called out because it seems like MS is making a similar call.
The author states "I consider the scaling-only focus to be both misdirected and not fun.". I get the point but also totally get the focus on scale out for now.
I empathize with LeCunn but he is a prime example of why scientists often make poor leaders. It’s too tempting to always do the cool and interesting thing instead of drilling down on the boring but reliable thing that actually creates value. There is of course a place for this kind of eternal curiosity and I hope he’s found it, but hyper-competitive industry may not be it.
Ironically, Yann is kind of known in academic circles for being the guy working on boring, unsexy stuff (deep learning) when all the other academics were doing sexy cutting edge SVMs and stuff. That’s probably why he got the Meta job in the first place.
Yeah, being a trendsetter pretty much means you had to be working on something boring and unsexy at first. I can't really remember the last time a major innovation came because someone knew exactly what to invent, everyone agreed it was important, and we all ran full speed ahead on it.
Perhaps the atomic bomb? Not being alarmist, but that seems like one of the few instances in modern history. I also thought of the space race, but even rockets as an innovative concept started centuries beforehand as a tool of war.
For a researcher it is to explore and illuminate. That this exploration and illumination often leads to a landfall of new wealth generation is a happy accident. This ensures that the monied class would bet some of their capital on them, at least the wise ones.
I'm becoming more sympathetic to LeCun's views, but being able to train something that acts the tiniest bit like those who made the training data (authors, professors, researchers, etc) is an absolute juggernaut and will be for some time.
You can't. It's AI booster propaganda. Yann LeCun obviously understands transformers better than random anons on this venture capital fan forum. This means he also understands their limitations, which is too complex for people who just want their preferred billionaire to win.
> "I take an object, I put it on the table, and I push the table... there's no text in the world, I believe, that explains this... your GPT-5000, or whatever it is, it's never gonna learn about this."
Lex fridman podcast 2022
ChatGPT answers this just fine.
He also said they will be “ Writing aids, not much more" and they “Can't plan.”
No one said hes an idiot. He was just clearly wrong about LLM capabilities.
>> "I take an object, I put it on the table, and I push the table... there's no text in the world, I believe, that explains this... your GPT-5000, or whatever it is, it's never gonna learn about this."
As far as I know, those things are still true. Unless there is text in the training material that the text in the prompt can map to, or there is text in training material that describe how to map text in the prompt to text in the training material, LLMs cannot process the prompt.
So may be he is wrong about what could be in the training material. But I don't think he is wrong in the fundamental..
This is sort of what I mention in my original comment. These claims, when pushed, turns out not to hold water...
Yann is smart but obviously not omniscient, I don't think it does the ML community a service to pretend that elders are never incorrect (or even badly incorrect). This type of thing is how we got the cult around Schmidhubering.
They surprised almost everyone, but he remained skeptical well after it was clear we could not put boundaries on the capabilities gained from scaling, and also from new RL techniques. Still, I do not think he is wrong about world models and continuous learning being fundamental to AGI, though the latter issue he was himself dismissive of until relatively recently.
> Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong
You're being uncharitable. He would have been right if Chain of Thought hadn't been pulled out of a hat, and no one saw it coming, or how u reasonably effective it would be.
Today, no one can know if there will be a future step-change improvement to LLMs at the same level as CoT, and conversely, no one can be absolutely certein there won't be - so it shouldn't be a firable offense because neither position obvious; even though it will look like it in hindsight.
The most confusing thing to me about AI and LLM progress in the past 4 years is that I can't clearly see how all novel patterns add up, or what was the right point.
I think adding harness + MCP + chain of thought + specialized data generation resulted in the world where we have now - the one where you don't have to type code manually, but sometimes do still have to think and sometimes don't.
But LLMs without all this tooling and calls under the hood are still kind of... LLMs. If there's a space to hallucinate - they will hallucinate.
CoT wasn't pulled out of a hat. It was obvious to any researcher that it was needed. It just took some time and stronger base models and longer context to actually be implemented. People were already doing "CoT" by telling models to think step by step before answering, primitive harnesses and similar tricks.
Even a nobody like me was playing with a sort of CoT with llama finetunes before o1 came out.
Only this year I'd say we picked all the low hanging fruit that has been obvious for a while that was needed.
I'm not fully up to date on AI research anymore, however I think the obvious stuff has already been implemented, but there's still a lot of room for improvement and scaling for these current approaches.
For verifiable tasks like coding and math, you can keep going, doing more and more RL. You have the AI generate hard problems, then try to solve them and reward correct solutions. We basically have RSI for verifiable tasks, we just need to keep the wheel turning.
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
The idea that someone should be fired just because they were wrong seems to be something that would cause issues if one wanted to think long term.
This approach will make people afraid of being wrong. It will lead them to avoid challenging dogmatic assumptions in large organizations and inevitably lead to group think.
People really shouldn’t be punished for saying wrong things. Long term it’s better for a company if they aren’t punished for that. Helps to let people freely criticize dogmas or propose new ones.
Depends on your role. If youre in charge of research in a hundred billion dollar org you should absolutely be fired if you devote your org to pie in the sky research instead of creating tangible value. If your job is just doing the research obviously you shouldnt be punished if your experiments dont work. Yann was fired because his main job was balancing product development with research and he didnt do that.
None of these second-rate LLMs are creating tangible value. Nobody has made any money developing LLMs yet, and when the capex runway ends, the second-rate models won't see much (if any) use at all, compute is too precious for that.
It's amusing that their grand dream was basically a clone of Roblox, just with better VR tech and far more advanced graphics. But thing is, kids playing Roblox absolutely do not care about graphics.
That at least was a reasonable idea. The Russians had a lot of old Soviet era ICBMs they had no need for anymore (plus, IIRC an agreement with the US to disassemble them), so using their engines as a cheap powerplant until SpaceX nailed reusability made sense.
That failed and SpaceX was then forced to develop its own brand new engines.
Yann LeCun has the syndrome where someone who sees the value of B (non-text modalities, i.e. physical reality) cannot stop themselves from constantly critiquing A (LLMs), even though clearly A has its own advantages.
I think he is a smart guy. But he constantly deploys the crackpot's PR playbook.
It's not so much they _turned_ wrong, they were obviously wrong at the time he made statements, or shortly after. He is just extremely biased against transformers, and it's easy to see why - transformers are already invented, and as a scientist he wants to invent something new.
More specifically:
* he said that autoregressive transformer is fundamentally flawed: if it makes one error, all further generated tokens make no sense; that's, obviously, bullshit: there are many ways to implement backtracking, and there are many papers demonstrating backtracking at a time he did that lecture. And, of course, it just doesn't matter - reasoning model can just write "Oh, that's wrong. Let's try a different thing". - it's just not derailed by one mistake.
* he said that GPT-1000 can't possible have understanding of physical world, when GPT-4 already existing at that time could solve a problem he described.
So it's not like he was bad at predicting things, he's intellectually dishonest.
TPS (and Lean, which it played a role in inspiring) is largely about finding what works empirically and ruthlesslessly optimizing around it.
It does (Lean at least) incorporate controlled experimentation, but tightly controlled with a both a laser focus on what the goal is and tight resource constraints. And Lean (at least several writers on Lean) favors taking multiple paths with a a quick and low risk and a longer-payoff, higher-risk, higher reward path for some efforts—one source used the US Navy’s two-path approach to SLBMs with Polaris and Poseidon as a model for this—but that's different than “maybe there is something down this path” basic research, which is outside of the concern a production model like TPS/Lean and often completely outside the model of a product-oriented firm; basic research is something firms consume and incorporate more than generate, because until the basics have been proven out, the rationally perceived chance of any particular avenue paying off is near zero for any firm.
Ruthlessly optimizing isn't the best goal, because people "optimize" all the quality out, and lack of quality leads to lack of sales and increased cost. Ford/GM may have "optimized" their line for efficiency, but the increase in defects caused more losses. Companies that practice Lean tend to forget this and focus only on speed and cost, leading to enshittification. TPS focuses more on quality and its effects on lowering cost and increasing sales.
Toyota is also a good example of what not to do. Their lack of earlier research investment in EVs has led to being severely behind the market, in addition to their lack of focus on quality causing poorer sales (their quality used to be one of the only reasons to buy them). KIA basically stole their playbook by creating loads of good-quality cheap hybrids and EVs. Toyota could've had that market if the invested more ahead of time in research.
And that's what's happening with AI. While there are plenty of players now, it's the ones that aren't looking ahead that will be left behind. Nothing scales forever; when you hit the ceiling, you're trapped in a box of your own making.
This is a dumb question since I know very little about LLMs and transformers but I keep hearing similar things such as "more GPUs/parameters/compute" leads to a predictable scaling in intelligence.
Is that observed scaling part of some equation or law? Is there a limit to how far that scaling goes before a ceiling or diminishing returns hit? Or is it simply based on more GPUs/parameters/compute plus how much human knowledge is in existence and will scale predictably forever?
It should be noted that these laws aren't, well, laws; "a neural scaling law is an empirical scaling law that describes how neural network performance changes as key factors are scaled up or down.". Notice the word empirical here.
These empirical observations on quantities which aren't "intelligence" since no one has yet defined intelligence.
Data isn't "smart", it doesn't contain intelligence. I would argue that LLMs are functionally intelligent, they have learned to what functionally amounts to thinking by training on text (and images, in some cases). They are capable of performing experiments autonomously and make decisions based on out-of-distribution data. Besides (current) lack of online learning, they are limited by their biases and working memory - but so are humans, and I'd like to see a convincing argument on why these mathematical laws should not also apply to humans, which are arguably intelligent.
Are you smart if you see that cognitive scientists had something figured out, connected it to something in your field of chemistry and made a breakthrough treatment for some disease?
Does any of our existing intelligence build on completely unknown information?
You could argue that any new solution/emergent intelligence is an association of existing information in new ways.
It also brings us to what is brute forcing? Humans are brute forcing stuff most of the time. You will have 10 stupid thoughts for every good thought. 10 stupid ideas for every good idea, you will discard most of them so fast that you won't think of that you did that. It's very rare, if not near impossible, that a non-mathematician with no prior knowledge, goes into a room on day 1 of deciding to think about a conjecture they don't know anything about, and solve it on the blackboard. They need to use their collect associations, formalize their axioms, think of first principles (known information), and associate it.
So to answer your question you need to answer:
1) What does "smarter" refer to? Is it accumulation of information about the universe?(if yes, a database is smart) Is it an efficient, abstracted representation of accumulated information? Is it prediction of the future state of the universe given a learned abstraction?
Or going further, is it solving tasks useful for the human in the world? Is it acting as an agent with it's own goals and interests in the universe? What if the goals are hard for humans to understand?
Does it count as smart if you make an evolutionary approach that spawn 1000 instances and the most fit one survives and spawns 1000 mutations of itself?
2) What does "genuinely new approaches" refer to? Is it fine to use math axioms to prove things? Does interpolation between two existing solutions count as new? At what granularity of concept does novelty happen (Does using two higher abstraction concepts, like combining some new, modern neural network architecture with a database count as a new thing?), even if the underlying lower abstraction of the same information can be expressed as storage, latency and compute, and is not novel?
---
As you can imagine, this is all not really defined or agreed upon. Here are some examples of projects that are set up to drive these ideas:
The Hutter Prize poses that intelligence is defined by compression, so if you manage to compress the English Wikipedia efficiently, you are closer to better intelligence (http://prize.hutter1.net/)
Compute: Intel CPU
The ARC-AGI foundation poses that being able to be goal-oriented and understand higher level abstractions is intelligence: https://arcprize.org/leaderboard
All of this also exists in the context of the compute medium. It could be, that you need a certain capacity of memory/computation to perform whatever we have defined as intelligent tasks. Hutter Prize takes the bet of that to make intelligence, you don't need a lot of computation, an Intel CPU laptop and 50GB RAM should be enough.
A lot of the big labs are instead taking the bet on that you just scale everything, make trillion-sized models and things will fall out of that automatically, for this you will need a very different compute medium (data centers) compared to (a laptop).
'Scaling law' is the prediction that increasing the training data and model size will cause the result to be better. It's not a law, but it was the prediction that OpenAI staked their fledgling startup on to get to the ChatGPT moment.
Your phrasing (_observed_ scaling _part of_ some equation) might indicate you assume / think there’s some theoretical equation that relate intelligence to those parameters and the we observe the data to confirm it. As far as I know, there isn’t - it’s purely empirical curve fitting. Train a model with different amount of compute / data / parameters, measure performance, fit curve.
Yes, the Chinchilla scaling laws from 2022: https://arxiv.org/abs/2203.15556 More parameters directly and mechanically means less prediction error for each token.
(This moves the argument to "is being really good at predicting the next token the same as being intelligent?" which has happily diverted many minds for years now)
The Chinchilla paper is probably the poster child of “paper that aged like milk” in LLM field though. Everyone over the past 4 years have trained with several orders of magnitude more data than the Chinchilla optimal number.
If you don’t know which data has the highest value, though more at it. Chinchilla is probably still true, but with that caveat. And it could be like advertising… you know a lot of it is worthless, but you don’t know which part.
I'm seeing the same pattern at other multinationals with huge R&D departments. Mostly colleagues with academic pedigree in traditional ai research who have been for years in denial that transformers are actually - in combination with recent advanced patterns in software development - the key to ai. it's really simple as that! except they weren't part of the revolution and still don't want to accept the reality that their research has become obsolete.
Is it the key to AI? It probably depends on the squishy definition of AI and what different people want from it. If your definition includes self awareness and sentience I’m not sure it is the key (I’m also not entirely convinced it isn’t, either), but it’s not clear either way in my opinion.
The one thing that is clear is if it can become more cost effective it’s world changing. Even at its subsidized standards now it’s world changing. And it’s reasonably classified as AI IMO, but it’s not self aware.
I think a lot of your general traditional researchers are fundamentally looking to understand ourselves through machines as proxies in a sense. Something self aware that senses, learns, adapts, so on. The current path isn’t there or necessarily clearly the key to achieving that. That may be the disconnect. I think LLMs may give us some hints at that but not the full picture.
Still incredibly powerful for things we once considered nearly impossible.
> Current top models are mere (mere in big datacenter terms) trillions of parameters right now. ... We'll soon start hitting 10's of trillions and 100's of trillions of parameter models.
Chinchilla Law says no, we wont. We're likely well past the point of diminishing returns but MOE hides a fair bit of it.
We already know that power laws apply to scaling so 10x the size is going to result in minor gains - where costs become a real factor (as if they weren't already).
FTA "The core philosophy there is scaling, and aggressively not research."
This is where the real gains for AI come from - good enough and CHEAP is going to win the day here (brute force, mixture of models). It's the engineers displacing the academics - and that will mean a bloodbath for revenue (see DeepInfer for what the end game looks like).
I'm not an AI booster (in a "should" not "can" sense), but in terms of "can" we are going to see next generation hardware (i.e. Vera Rubin GPU architecture) and improvements in model density (i.e. loosely improvements in intelligence per millions of weights), so the models will get bigger, the models will get more dense, and we may see more "expert models" i.e. in the sense that for a while there were human generalists that knew all of science and mathematics until the 1800s, after a certain point we had to specialize since there was too much to know. The models will need to partially specialize after hitting the hardware ceiling.
Bioinspired algorithms may eventually lead to vastly improved intelligence per a weight too.
I don't think we should do this, but I feel doomed in that I suspect that we "can".
> The models will need to partially specialize after hitting the hardware ceiling.
I'm incredibly skeptical of anything that goes "oh, we need to split AI into specialist systems".
I suspect that a lot of the delta between "our best models" and "base models they're trained on" lies in relatively low amounts of parameter changes driven by SFT and RL.
Model capacity is not the limit - the limit is our ability to drive those changes efficiently. Even things like "good research taste" could be represented as tiny in relation to total model capacity, if they could be trained into a model efficiently.
At least for now, those keys haven't been completely unlocked, so parameter scaling is the primary way corporations are achieving increases in ability. If they hit the hardware ceiling, maybe the research direction changes then.
I forget who said it, but a trenchant insight is that current neural networks are generalized function approximators and essentially they are doing through brute force what biology accomplishes elegantly. Our hardware (both in the sense of our brains in terms of elegance and our computer chips in terms of raw throughput) is amazing.
No, parameter scaling still works because it broadly increases sample efficiency. We don't have the raw data to make a Chinchilla optimal 100T, but that doesn't mean there are no gains in building up to 100T.
I don't think "biology" and "elegantly" belongs in the same sentence. Let alone the one you wrote.
I suspect that the "elegant" workings of the human brain are a pile of hacks. Evolution crammed the brain full of patterns that are a good fit for the specific problems it typically solves - the way curve detectors and scale invariance are a good fit for CNNs.
We don't have millions of years of evolution of non-random init, but what we have is enormous datasets and compute at scale.
as fast as possible implies spending the resources that way.
Going close to as fast as possible while saving a stream for the other approaches makes sense as an approach, but let's not ignore that there's a compromise/tradeoff here. Might not matter in orgs with near infinite resources
Also in an exponential competition environment, the difference between 100% commitment and 98% commitment is very possibly not even the difference between a win and second place, but outright lapping by the winner and humiliation at the scale of the gap for second place.
For an example of what this looks like at a more pronounced scale just look at the discourse around European models and how far behind Minstral are perceived to be.
Theoretically one need doesn't preclude the other. The OP gives a detailed argument and his own experience of one need precluding the other in practice.
Both of these approaches require quite a bit of institutional support notably.
Except doing both..is..so..hard when the future payoff of one is more unclear and you are faced with quarterly earnings reports.
Look at what happened to Meta where they squandered a lead, had some mixed results, and reorged to more product certainty and away from fundamental research.
That depends on your definition of scale. Yes! We were to literally convert the whole world and its resources into data centers.
But the whole problem is that these things aren't acting like other technological revolutions like the internet or Google where lots of profit comes from relatively little spend. They scale linearly maybe if you are super smart and good about your design.
Meanwhile, citizens all over the world are fighting data centers tooth and nail because we are hitting a scaling limit in terms of the resources people will allow to be put into this highway project and its equally unwise and untethered from reality to pretend this doesn't matter in a strategic sense.
"citizens all over the world" will not lift a finger for CAFOs or trawl fishing. i really don't think anyone "fighting data centers tooth and nail" is doing so because of a belief in the limits of scaling or any understanding of scarcity.
Im at a company with its own set of machine learning engineers, a few separate organizations internally have them so probably a few hundred MLEs in total.
More and more, the talk of using generative models (e.g. transformers) is coming up to replace existing models in all concepts. Anything from understanding data, image recognition, labeling, ranking, etc.
As someone who is not an MLE and really doesnt understand these models well, it seems like the transformer approach is a "one size fits all" solution that may not be optimal for all problem categories but is good enough that its not worth investing in specialized models to get incremental improvements.
Isn't the thing with transformers is that they work really well, but only if you have a lot of training data? If data is sparse, you need more specialized models
More parameters without significantly more data fails; it loses the ability to generalize and simply memorizes.
And we’ve consumed pretty much all the data, and that part is getting harder as data is now expensive, Balkanized by owners, litigated and charged for. Couple that with model collapse as models eat their own output (not solved for any models, despite more than a decade of the problem), and add that lots of previously human curated datasets (stack overflow style), and it very well may be the models are not getting much better.
Very large models trained with sgd variants don't tend to over fit, especially when using few epochs. See double descent.
The coding/math breakthroughs were caused by training models on the subset of their own outputs which pass a verifier, see reinforcement learning from verified rewards.
That's the historical wisdom, but there are reasons to believe the opposite: that past a certain point more parameters are always better. Basically, that overfitting don't matter, because a large enough model will memorize and build the general solution simultaneously. Then weight decay and SGD noise encourage forgetting of the memorized solution so the general solution takes over.
is scaling working? It is also my understanding that Opus 5.5 is probably a well under 10T parameter model. Probably a similar size class to Fable 5 earlier. And supposedly even GPT-4 was a 1T model. Clearly most of the gains are coming from something other than scaling up the model size. Yes, scaling up RL counts as scaling too, but the fact that labs have pivoted to a different axis of scaling is evidence that it's not just a "straight highway" to improve LLMs.
From the outside it seemed like GPT 4.5 was the apogee of parameter size scaling and since then a huge investment in RLVR became the new scaling axis.
So yeah, LLMs keep getting more powerful/useful but because of an industry wide pivot not a "straight highway" as imagined in the GPT-4 era, when simply going bigger and bigger seemed like a reasonable roadmap for the foreseeable future.
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Where's the money coming for those 10X and 100X ventures?
It's worth it to keep exploiting transformers while doing just a bit of exploration of alternatives because they work so well. Heck, just exploiting transformers more will accelerate exploration of alternatives...
100 trillion parameters doesn't make any sense if your dataset is ~30 trillion tokens. My uneducated guess is 1/10 compression ratio is good and that point to ~300 to ~1000 trillion tokens, which again, doesn't make any practical sense.
Depending on how you count, the human brain is between 100 trillion and 1 quadrillion parameters, which is one guess as to why we're so much more sample efficient at learning than LLMs are.
One key difference between humans and (current) LLMs is that we can actually judge the material we learn from at the time we're learning it, contextualize it and discard garbage when we spot it.
LLMs need an awful lot of handholding all the time, meanwhile humans get pretty good at it on their own.
Unless quantum effects do infact play a role in consciousness. It’s not clear they do by any means, but if it did that could shift the understanding of our own minds and consciousness drastically. Otherwise we are just a dense complex network we can reproduce with enough money.
I mean we literally do this every few seconds already with natural birth processes.
I don't think you even need to go to hypothesized quantum stuff. A single biological neuron in isolation is a surprisingly complex system that resembles a small artificial neural net more than an artificial neuron in many ways. On top of an entire feedback system from diffusion of neurotransmitters which doesn't cleanly map to ANNs.
Trying to estimate equivalent "parameter size" of the brain when we only superficially understand behavior of individual neurons let alone how intelligence emerges from the connectome seems too speculative to be a particularly useful metric.
Same in deepmind too. Bit of a shuffle as well to the frontier AI group, jeff dean and prateek jain headed it now jeff is gone, so prateek has moved to MTV. And most of the different research groups are all in some sense under deepmind, moved to LLMs and in no-publication mode now.
I wouldn't interpret it as some sort of philosophical position. After a few years once the current buildout settles down they will go back to the usual setup. In the meantime, my alma mater is poaching some of the disgruntled publishing-oriented folks from there.
Right now it is stupid to not dedicate all organisational resources towards LLMs, when you are lighting your free cash flow on fire and issuing bonds like there's no tomorrow, you might as well have all your smart people working on it. JEPA can wait a few years.
Addendum: the amount of people in this thread misinterpreting what yann lecun said is insane.
> Hey this Opus 5.5 with the increased parameters seems much better than the last one"
Is this true?
I barely notice any improvements ever since... 4.6? At least for software development.
Opus makes the same sort of mistakes and has the same sort of strengths ever since that point in my perception. My impressions is that increasing the number of parameters clearly has diminishing returns.
In fact, the best improvement I got from Claude was Haiku 5.5. Good cheap workhorse to implement plans.
As in your statement to me indicates that you clearly see the scaling working over longer timespans. You chose a relatively recent 2026 model as your baseline after all, not a 2025 model or earlier. You clearly see that the 2026 model is better than ones in the past.
Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together? If you're not anecdotally seeing improvement between an early 2026 model and one from ~6 months later perhaps just give it another iteration? Personally i do see 5.5 as much better but i can understand not seeing a lot of difference to 4.6 since they literally aren't that far apart in releases.
> Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together?
No, you can't. Because you invalidate your argument by adding later:
> Personally i do see 5.5 as much better
It has to be one or the other.
Either each minor release Anthropic farts is a major improvement from what came before, or I don't see much difference because it has been very little time in between releases. Holding both arguments as true would require me to be excessively apologetic for a company that desires a 2T IPO.
And to be clear, I do think that there is an argument that not much time has passed since Opus 4.6 and Chat GPT 5.3 Codex days.
Part of this conundrum may be because the models reached a good enough point for software development. Maybe walking 80% of the way was easier than the final 20%. I am nowadays more excited about how cheap models can be useful for it (hence my positive remark about Haiku 5.5 and Chinese models in general).
These corporations have an obscene amount of money
This is a false dichotomy if I ever saw one
Why not hedge and just keep an ai skunkworks team working? These companies are big enough that they can have these people far away from the main work without distraction or without starving those teams of bandwidth
> "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model.
Opus is 5T, not 1T.
> We'll soon start hitting 10's of trillions
Fable is already 10T.
> "Foot down, there's a straight highway ahead don't divert". I have the view this makes sense too.
The roadblock they ignored is politics.
Three days ago the states where most of the new datacenters are being built (Texas, Ohio, Georgia and others) elected governors who campaigned on halting datacenter construction. The highway metaphor implies money in => intelligence out. Turns out there's at least one other ingredient, and we just ran out of it.
I'm still convinced the model advancements are largely US firms jingling a set of keys in front of investors/markets/the public to distract while they scramble to build serious patent moats around the harness/orchestration/infrastructure technologies.
I think that we're well into diminishing returns already. The models have improved, but to me the far bigger advances have been in feeding them contexts, building them into loops, etc.
I don't even think much more construction is really necessary in data centers. We can do a lot more with the power/HVAC we have, especially once the ASICs start coming online.
It's like we invented the wheel in 2022, the crowds are fighting over capital to build wheel factories, and savvier tech firms are quietly patenting automobile designs while everyone's focused on making slightly rounder wheels.
When a company pays billions to OpenAI, any activity that does not do a glorious dance in praise and support of that investment is taken to be an act of disloyalty.
This doesn't engage with the article in the precise way that transformer-based improvements and fundamental research that way are not being incorporated into models, at Microsoft AI, which is concerning.
Which many credible, AI researchers believe is already starting up an increasingly steep scaling wall of diminishing returns that eventually ends at a brick wall.
I've been hearing a lot of "credible AI researchers" talk about these diminishing returns of transformer-based architectures since 2023 (at the least), and I'm still waiting to see this promised plateau materialize, because I definitely can't see it in https://metr.org/time-horizons/ or in my personal experience.
Assuming super-linear growth will continue based on the first 3 years of a new technology? Almost every physical, computational and economic system in history, from Dennard scaling in silicon to transonic aviation speed looked super-linear until it hit the inflection point in its eventual S-curve.
A Copernican/Lindy baseline suggests the growth rate in LLM intelligence will peak in around another 2-4 years and I'm leaning on the low side because while the growth of LLM intelligence per dollar and per month are still growing in absolute terms, the growth rate of both seems to already be falling as power-law constraints in compute and data bottlenecks take hold.
There's just not much value to a 'Every technology hit a brick wall eventually' take. Anyone can spout that. Why even bring 'credible AI researchers' if that's all it amounts to. What is relevant is the foresight to predict when and at what capabilities. It doesn't change much of anything if LLMs 'plateau' right after they can replace any human at any job.
This sort of tallies with the claim that every significant technology in the iPhone was initially developed through state funding. Basic research and fundamental inventions are expensive, long term, high risk projects that are generally incompatible with companies driven by quarterly profits and stock price (with some notable exceptions such as pharmaceutical companies).
It seems extremely likely to me that scaling produces some benefits, it will get better at particular things things continue but scaling by itself isn't heading to general intelligence.
I should also say that I use ChatGpt daily and enjoy it for researching algorithms. I tried to set up Claude last week and it/he was so overwhelming as to be unusable. So I think frontier models can be "unwieldy" or something for average use.
Why wouldn't the explore/exploit balance leave at least a little to the explore side? Huge amounts of money are going into AI, why not save a little bit to keep an eye on the things that don't currently look like they're going to scale up?
Like maybe a couple dozen people on an explore team?
I would say I used to believe how the models would scale, now I'm a bit more skeptical. I would be incredibly surprised if oai, anthropic, etc hadn't at least TRIED this yet. They have the compute to do it. They're either sitting on it and saving it for a rainy day, or there is something wrong with it.
We know nothing about their internal models that, for instance, produced revolutionary results in mathematics. The public no longer gets to see what’s at the frontier, and we may never see much of it. Also, the cost to train these is immense. China literally does not have enough chips in circulation to compete. I’m not sure where US lab capabilities tap out but I suspect they are at least partly bound by GPU production as well at this point.
We know that if this immense cost had in fact been spent, and the internal models really were that revolutionary, these companies would be selling access to it for cash instead of taking on mountains of debt.
I agree they would not shut up about it if they had something to reveal, so we probably have a semi-accurate vision of the capabilities if not the implementation details.
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters
Moore's law: Computing power is multiplied by 10 every 6 months? Can't wait for the optimization and rationalization phase, where we'll recover a trillionth of that power to run powerful databases.
>> We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one".
Opus is better (not necessarily "very much") than Sonnet in coding. That's because of a gigantic engineering effort to make it better at coding. That is a regression: we've gone from "sparks of AGI" and systems that were said to be "genearl-purpose" to systems that do one thing well and everything else... well, there's no real measure of how better Opus is from Sonnet in replying to your emails so it's hard to say.
Same goes for OpenAI in maths. Industry wide, there is a regression from general-purpose to special-purpose systems, like in Good, Old-Fashioned AI. LeCun is absolutely right to complain that this is not the path to general intelligence. I don't think his preferred approach, JEPA, is going anywhere either, but at least he's trying. Other luminaries of the field have expressed similar doubts, e.g. Richard Sutton's tweet was in my news feed yesterday saying that brute force is ngmi and algorithmic advances are needed [1]. Schmidhuber I think has said multiple times something along the lines of "there is no intelligence without physical intelligence".
We are very far from that. The recent trend for Vision-Language-Action Transformers (VLA-Transformers) in robotics has conclusively demonstrated that you can't solve the real world by overfitting to a set of discrete tasks, and there is no source of data that we can harness to repeat the feat of LLMs. Algorithmic advances will be necessary, but as the article suggests, these have now be put on hold.
There is a dumbing down, slowing down of research as all resources are spent to try and scale what we have now. What we have now is not enough for intelligence, but that doesn't matter. The industry's goal is not intelligence, it's making more money than anyone has ever made before; in fact more money than all humans put together have ever made before. That's the current race: who will win the money game. Obviously that's not likely to do anything for intelligence.
____________
[1] Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
Do they really? If your experience is with the free version of ChatGPT, I wouldn't be surprised if you're running a MoE with like 8-12B active params and <100B total.
(Just speculation on params, but I find a lot of "LLMs still hallucinate on non-coding problems" come from people using very-cost-optimised free models and drawing the wrong conclusions).
> If your experience is with the free version of ChatGPT,
Even the free version of ChatGPT shouldn't hallucinate, after all it uses a good enough model. And the answer is no, it's a combination of paid and free models coming from OpenAI, Claude and Google's Gemini (with this latter one actually badly miss-identifying a WW2 military dictator a couple of weeks ago, wrote about it here [1] )
What if scaling fails to hold at a certain point? Whether that is a steep or gradual dropoff in returns. What happens to the empires of accumulated compute? How much value do they immediately lose once the returns to scaling ceiling is reached?
> Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
Feels like this is leaving major breakthroughs on the table for the competition. Sure, transformers might get us much farther. But if someone figures out an approach that unlocks a performance boost or better resolution then they are going to kill the competition.
Feels like the American industry is high on all the capitol being shoveled into the furnace without caring how efficient the furnace is. At some point someone is going to make a breakthrough and burst that bubble. This could crater the value as we are currently betting on the low power efficiency of the models as the barrier of entry and point of profit. Of course that would lead to cheaper AI but is that profitable vs local models once (or if) it becomes affordable to do so?
Ignoramus question: does the 100s of trillions of parameters world still exist with a constrained domestic data center build out? Democrats are signaling they are about to go psycho mode on data center contracts.
Like the sister comment, if not in US then in other countries.
US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
> US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
The 'US' isn't a monolithic thing. US market leaders convinced US politicians to open US markets to China, and the lack of labor and environmental regulations was a feature, not a bug in the plan.
This refers specifically to "AI Frontiers", a research group within Microsoft Research that focuses on fundamental advances in AI. There is no end in sight to real AI frontiers.
Perhaps "The End of Microsoft Research's AI Frontiers Group"
would be a less ambiguous title.
The real news is "The End of Microsoft Research". Langford's group is just one of many that have been parted out over the last year or so. The division that was T&R under Kevin Scott was quietly dissolved about a year ago and MSR was reorged under Azure Security with the AzSec leader as hatchetman for the department. MSR itself doesn't have a future beyond a few years at most I reckon. I don't suspect Kevin Scott will be around much longer either since his department keeps getting reduced and is fully eligible for VEO.
Langford has been at MSR a very long time and has done amazing research doing that time. He's bounced around a bit at MSR during his tenure and it's a shame to lose him. AI Frontiers wasn't around that long.
Much of this is Microsoft performing a very large and uncommunicated strategy update. All updates of MSR reorgs are just that "MSR is becoming more lean and focused" when it's obvious this isn't the case. This gets into the conversation that many Microsoft employees tend to have these days which is "Why does culture no longer matter at Microsoft and why did Satya turn into another heartless CEO". I won't speculate there, but the place isn't nearly as fun as it used to be.
simply owning most of OpenAI via a convertible note on very cheap inference sold to it could turn out to be a bigger deal than the rest of what microsoft does.
Yeah, I went down the rabbit hole scratching my head the whole time. This is about the group at Microsoft Research that was disbanded or reorged into what is now Microsoft Research.
To the author: 100% need this kind of thinking. It’s a balancing act. Historically though, dedicated labs haven’t had the best track record at commercializing their tech. The skillset differs vastly from running a product-focused business.
That said, even within this narrow scope - there are areas ripe for research.
Some of my particular interests include:
(1) Latent thinking/reasoning, which could arguably be more efficient but also opaque is one
(2) Separation of “intelligence” from knowledge - ie. keep the ability to reason like a frontier model and make tool calls (calling the ‘oracle’ or searching the web) but remove all parameters associated with niche knowledge/facts.
These are a couple off the top of my head that are very much aligned to scaling.
That statement is broad enough to be unfalsifiable, but if you narrow it down just a bit to scaling LLMs, maybe there is an end in sight. At least financial markets are skeptical of there being no end in sight.
If scaling LLMs really was the road to God in a box, then compute demand would be infinite, and nobody building to meet that demand would have a junk debt rating. Did those money people just not get it?
It makes me sad because there are so many unexplored areas of AI research, even on text generation. Billions are invested in almost one idea, which very realistically isn't the most optimal way to do text generation, but a lot of companies are afraid of taking risks. I still wouldn't put $1b on JEPA but I prefer a world where we can try things like that.
LLMs have very favourable exploration<>consolidation dynamics. An LLM alternative would need to display fundamental improvements at relatively large scales which is extremely challenging for new research.
> there’s a breakdown of the basic deal in MSR: do good work and good things happen
This really hit home for me. This is the contract that held for a *long* time at Microsoft and has been broken and rebroken with current leadership. It’s culture-destroying.
He didn't say they were not doing transformers research. In fact, he specifically mentioned a type of transformer they were working on. And labs like Qwen and DeepSeek (at least) have shown how incredible results can be when you innovate in the area of LLMs.
I think the real point of the post is that office politics and poor management got in the way of innovation due to infighting and poor execution in transfer of technology from research to product.
To me, all signs point to US over investment in hardware capacity to support the "frontier" when in reality the next major breakthroughs are likely radical optimizations that end in increased consumer training and inference efficiency.
I'm not sure how that, combined with open models inching up on frontier intelligence, is compatible with these massive AI IPO valuations?
It seems like closed model AI demand has a lot of contingencies that are ultimately fragile and possibly short sighted.
Yeah, I'm speaking to "scaling vs research" and where investment is going. Which is relevant to the article. Accelerating scale requires demand, closed frontier demand is what I'm questioning.
And gp's comment is relevant to the particular problems the guy (John Langford) cites in his comment about leaving. IE, that research first winds being only about scaling and then winds-up ignored because "just scale" can be an easier answer.
If they make more efficient models won’t people just use the big data centres to make and run even bigger more powerful models than they can currently support? Or massive swarms of agents to look at all different avenues of a problem?
Everyone wants radically more efficient AI. That doesn’t mean it’s coming any time soon. And when models are twice as efficient, guess what happens next. We try making one twice as big for the same cost.
Companies with billions of $, stopping R&D on the most revolutionizing technology/ breakthrough is something I can’t understand. I do understand there’s an AI Race to win, and that is with transformer and current LLMs architectures, but if research could find a better, more efficient approach they would re-shape this race, lead it, or even dominate it.
The problem is the "if". There is no guarantee that there's a "better, more efficient approach" to be found.
So far, the attempts to replace the transformer LLM with something else have yielded either "the same transformer LLM but with this one tweak that makes it 10% better" or "a radically different architecture that's a bitch to work with, and performs, best case, about the same as the transformer LLM".
By now, I'm firmly in the camp of "the architecture is good enough". Rifle is fine. Transformer is fine. Get better at the transformer before you look at anything else.
I think it’s more likely that some future LLM model finds a more efficient architecture than a human does, at this point. So yeah I’d go all in on LLMs too.
Me neither. When revenues are growing every year, a fixed chunk of researchers' salaries (maybe growing a few % each year) doesn't really hurt, but R&D offers a nice upside.
Or if you are Microsoft, you throw in the towel on leading in any given industry or technology, and you're happy to be third best at everything and leverage b2b dealmaking to stay relevant.
More like, a large company has to be rational, as the article points out, throwing money at research for uncertain returns is not as straightforward as throwing it at scaling for known returns.
Startups are allowed to be irrational because investors know upfront that what their dollars are getting is a high variance outcome. People buy a stake in Microsoft when they want what they perceive to be a safe investment.
> More like, a large company has to be rational, as the article points out, throwing money at research for uncertain returns is not as straightforward as throwing it at scaling for known returns.
It's arguable whether scaling will give the expected results to justify the investment. Currently, it is a bet, just like funding alternative research is a bet.
But is it a race to zero? Maybe AI will make very expensive things (eg software engineering) very cheap, but there will be enough competition that the margins are tiny and don't actually justify putting crazy dollar amounts into research
That's exactly what's happening. Software -> 0 marginal cost (obviously still need engineers to actually make the architecture). LLMs -> Price of Energy / Amortized Cost of GPUs. "AI Labs" -> Owned by Creditors
From an engineer, who’s devastated by this, i have to admit/raise my concern: why do we still need engineers to make architecture? Why is everyone saying that’s a must? For me the toughest problem was coding a huge repo/pkgs, architecture should be much of an easier task to do. I’ve seen it do sys design, scaled architecture better than 10X engineers. With good planning, context and prospective info -> it can produce extremely good architecture even for complex systems.
Almost as fascinating is that the article references Digg as the source - genuinely hadn't heard of Digg for maybe a decade? I thought the DVD encryption fiasco + reddit had ended it?
The headline makes good bait for AI skeptics. But this is a team of 50 people with a grandiose name.
The more interesting idea is that MS has abandoned AI research entirely, presumably because their zillions invested in other companies are a better bet in that area.
If MS abandoned a promising idea, pursuing it outside makes perfect sense. I am sure there are investors willing to fund this. This could even be an opening to attack the moat of the current AI frontier model companies.
It is notable that with the advancements of 24/7 coding agents with basic research capabilities the premise:
> a clever idea pursued by one person it was likely to arrive late and unconvincing because one person could not move fast enough and the baselines which one person could assemble just were not strong enough to really show that an idea was worth taking flight
may no longer be true. The limiting factors are: frontier model token and compute budgets, researcher skill, drive, foresight and imagination, not the ability to work in teams. Of course, working in teams still helps, and benefits from internal use, peer review and promotion of the research line are certainly there, but compared to some time prior to 2026, it's pretty much not a factor.
On the flip side, frontier models are so good and relatively cheap now that it's possible to pursue your pet research project on your own as long as resources are not a constraint.
I think the dichotomy of open-research vs. commercially-driven innovation might be easier to coordinate, with three stages, not two:
1. Open research: Great creative freedom, with great resource constraints. Lots of ideas, lots of things tried, at insignificant cost relative to commercial work.
2. Results filtering, vetting, regularization: Opportunistic selection of apparently successful work from stage 1, for tentative testing at increasing scale. The best ideas end up with moderate budgets that verify benefits, which are reworked for compatibility before any handoff to stage 3.
3. Commercial efforts: aimed strictly at improving quality and economics of current revenue sources.
Direct negotiation between #1 and #3 involves too much dissonance over globally complementary but locally incompatible values.
#2 can be a relatively small, high payoff group, keeping #1 and #3 from butting heads, while facilitating them both.
It's mentioned in the article as "post-inflection" which means when Mustafa's startup was acquihired, and Mustafa became CEO of msft AI. (Microsoft likes to call the heads of various orgs "CEO".)
Mustafa has since been demoted from head of MAI (which includes applications) to just head of the superintelligence team. (Of course not externally broadcasted as a demotion.)
Real research at most big companies is dead. Unfortunately, what has replaced it is mostly venture-backed startups, where the timeline to a return is still expected to be quite short.
As a result, you end up with companies doing basic research under the guise of developing a product, announcing timelines which are never going to be met. Fusion energy and quantum computing startups are some of the most egregious examples of this.
Remember Bell labs, XEROX parc? Ask yourself why that existed and why research came out of them.
Monopoly.
So - many people don’t want to admit this but monopolies can be good from an R&D perspective, but without strong oversight by regulators - money will instead spent on R&D projects that specifically raise entry barriers.
I can show you many more examples of monopolies which did not produce PARC or Bell Labs than those which did. It's possible such research happened in spite of dominating the market, not because of it. Anyway, you cannot have both monopolies and strong regulators. These are incompatible situations. Monopolies occur because regulation doesn't exist or is captured.
I meant competent, but regulator strength in the way you're implying at its limit becomes monopolistic itself. So when I speak of regulation strength with respect to monopoly, I am speaking to the ability of a regulator to effectively both understand their duty and carry it out.
Agreed. I used to be an industrial researcher at one of the last companies that didn’t require their researchers to think of short-term productization goals. However, things changed at my former company, and I ended up leaving for a tenure-track position at a community college.
I don’t know what industrial research in CS is like in the era of agentic coding. On one hand I suppose building prototypes is faster than ever, but on the other hand I genuinely don’t know how well LLMs do with harder programming tasks. It’s one thing to have an agent work on a CRUD app, but I wonder how effective agents are at handling things like parallelism, distributed computing, load-balancing, and other topics that are harder to write than the average CRUD app. I am also aware of efforts such as Sakana AI’s automated scientific discovery research, and I wonder if many of the incremental, short-term research topics many companies pursue can be automated.
I’m a bit pessimistic about the state of research these days, from funding to publish-or-perish to the effects of generative AI (automation, AI slop publications, AI slop reviews). Then again, I’m a bit of a curmudgeon, and maybe things are better than I think.
I have been shorting most of the Quantum Computing startups for about a year. They have been printing. Missed out on Fusion! (?) These were one of few examples of the market leaving money on the table. Others examples have been Nikola. Usually the scammyiest companies don't IPO.
I'm about 90% certain that there are numerous huge inefficiencies in these big transformer models. Inefficiencies so big that eliminating them would make it possible to run the equivalent of the latest frontier models on your phone.
But nobody wants to look for them too hard because centralized AI as a service is where all the money is. So everybody keeps building bigger rather than smarter.
If you value the freedom to pursue your own research direction and share your work freely, then choose academia. All the perks of working in an industry research role are provided in exchange for this freedom.
If your research aligns with the incentives of industry, you can get the best of both worlds in industry. This was the case for machine learning researchers in the post-AlexNet, pre-ChatGPT era. The good times can't last forever, though.
>If you value the freedom to pursue your own research direction and share your work freely, then choose academia
Perhaps it is more free, but my understanding has always been that professional academia still restricts freedom of research. You have to apply for funding, convince the board that your direction will be fruitul.
An archaeologist can't just dig wherever. They compete with their peers for a limitted funding pool. If the school's board is more interested in the bronze age than the paleolithic, you're probably not going to get your dino dig approved.
The same is true for CS, I assume. You can't just "i would like spend university resources developing brainfuck 2.0" because the other teams are more likely to get their funding over you getting yours.
Both corporations and university's have finite budgets, both with bottom lines and margins. The uni wants their research done in a popular field that will attract students because students generate income that funds research that attracts students that generate income that funds research that...
Academic research is free in the same sense as running your own business. You can try whatever you want if you can find the funding, but you still have to deal with people and regulations.
Universities are mostly just administrative organizations. They have little control over your research, they rarely fund it, and they are usually not even interested in it.
That’s not the issue here, it’s not like AIF was doing purely academic work.
They just think it makes sense to explore things, rather than just rinse and repeat with more cash.
The good news for John is that when 99.99% of the competitors are doing all the same, it’s when the rewards for looking around can be the biggest. Those are the real moments for disruption, and companies disregard them at their peril.
>
If you value the freedom to pursue your own research direction and share your work freely, then choose academia.
The working environment in academia is often precarious :-( , so such people often have no alternatives to going into industry, even though they still deeply appreciate and are influenced by the values of academia.
Hmm... So strange. It seems obvious(?) that both research and scale are needed, and that one should never have one starve the other. In fact research can influence the scale needed, and vice versa.
I guess when money is concerned people will always seek out the the most reliable and predictable path - at the risk of being potentially left behind if it turned out to the wrong (or at least suboptimal) path.
This seems to track with a recent blog post from MAI regarding their approach to pre-training their first Frontier level model which seems promising but does seem to have a focus on scaling.
(In the context of Microsoft Research) Man, the history of how large corps fumble major wins should be studied.
I remember during my thesis how Microsoft Academic (a Google scholar/Scopus alternative) was ridiculously good at its job.
I'm talking, "input the key words and filter out the wrong fields and you've got 20 highly relevant papers that no other search term has found" level.
...Of course, they proceeded to close it down.
I can't even express in words how mad and disappointed I was and still am at Microsoft for this alone. It felt the closest to a genie or an oracle in being amazing. (The old Swype vs modern Gboard is probably a close example for those who'd used both.) It's just... crazy, frustrating, and crazy frustrating.
I used to work on the search index and query processing for Microsoft Academic. It was a great project, perhaps the most interesting thing I've ever worked on.
Ironically now by collaborating with AI one person can absolutely move fast enough. I used to work in orgs like that and am now vastly more productive than I ever was back then. It could take me months to convince a peer of an idea that now only takes minutes to explain to AI.
Executives chasing short-term returns. Google did 20 years of AI research, even creating custom chips, just so we could get to this point. Without long-term strategy Microsoft will perpetually be the also-ran. (which I guess is fine with shareholders?)
I'll play devil's advocate. Of course John Langford thinks what he and his group were doing held a lot of promise, and the work was fun and challenging and super important. But in the end you also need to live in the present and deliver. In the words of Jerry Maguire, "Show me the money". He had years to show Microsoft the money, and in the same time other people at other firms showed their bosses the money, big time.
this is not true. even with ideas with provable/strong scalability, there was still a lack of resources (especially GPUs) to validate such ideas at scale. ultimately, "showing the money" still requires some initial faith from those higher up who control the resources.
Yes, leave it to the Chinese to advance the state of the art. America throws money at it, the rest brains. Because they don't have the HW.
Same happened to computer chess in the 80ies. Clever small SW was leading the tournaments, until Ken Thompson could persuade IBM to throw hardware at it, Big Blue. This win was shortsighted. HW didn't scale and nobody profited from it. SW did, esp. reinforcement learning.
They've offered voluntary redundancy to a quite large proportion of the staff (a quarter?) so they'll also soon be losing anyone with employment prospects elsewhere.
I see a lot of news about waves of layoffs in large tech companies. Is MS the only shitty place now? I think all of them are, maybe except OpenAI & Anthropic, since they're building something new and are kind of on the "good" side right now.
Yep, from what I hear and see, it's like this everywhere. It actually started with the Twitter layoffs and has spread across the industry. At first, it seemed like a reaction to the ZIRP-era overhiring, but it's kept going long after correcting for it, and with AI has been cemented as the modus operandi for many companies.
From the outside, Microsoft seems to actually be one of the better places, and is making at least token efforts to reduce headcount in more amicable ways, as exemplified the buyout offer.
But in general, the strategy seems to be overload people until they burn out and quit, because hey, AI will replace them anyway.
but Mircosoft AI model is growing greatly and recently beat the nanobanana in generate video domain. I can't image the situation of them are bad as you explained.
It seems as if Microsoft is losing developers now.
The whole "let's go into AI" is not really compatible
with oldschool software jobs. AI drains humans dry.
I mean we are already seeing the hype cycle coming to
a close, with the stock market showing this, but key
people leaving kind of means that Microsoft is losing
its profile here. Add to this that many don't like Win11
and you see a huge paradigm shift here - and tons of
associated problems.
"In practice, there are substantial issues because
the philosophy of the ex-Inflection leadership of
MAI is incompatible."
AI is killing people here - not necessarily literally
(though we have to see how much economic damage AI
has caused) but figuratively speaking. The counter-wave
is coming here.
> If the research products were truly innovative and useful, he wouldn't need to force Microsoft to adopt them.
That's just patently not true. History is littered with products that were very innovative and useful (and often far superior to competing products), but were canned, bought by competetors, or shelved in other way - often because they threatened the profits being made by some entity (though that's not the only reason).
Yeah, I agree. What he describes reflects the exact challenges I've experienced literally everywhere. He is basically saying work/life is hard so he looking for a better playground. I do not know his contributions to AI so maybe he deserves to be better listened to but I find that never reflects corpo life.
Autogen is open source. I don't know the availability of other projects.
Generally engineers in megacorps aren't shy of picking up the latest AI tools. I suppose Microsoft may have barriers around this though, curious to discover more details.
This is true in every large company right now.
The transformer based approach is seen as a highway to race down in terms of training and building better models with minimal organizational distractions allowed while every company puts its foot down on this highway. Exploratory research into other AI methods is highly discouraged for now and organizations are encouraging transformer based scale out and iteration as a top priority.
As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models. They got someone willing to scale the current method out. Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
I have the view this makes sense too. Current top models are mere (mere in big datacenter terms) trillions of parameters right now. We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model. We'll soon start hitting 10's of trillions and 100's of trillions of parameter models. Unless there's a ceiling which nobody expects they should keep scaling just as they have been. I really do agree that it'd be insane to veer off this highway with exploratory research at this particular point in time. Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
> As an example Meta literally fired it's previously head of AI who held views that transformers won't scale. He wanted to veer off the highway and heavily fund exploratory new types of models.
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
Historically, that happens now and then. Examples include color television cameras and fluorescent lamp ballasts.
RCA developed the first good color TV cameras. They were monsters. Originally, electronic TV (as opposed to mechanical scanning) used Zworykin's iconoscope. This needed huge amounts of light. It didn't have a photo-multiplier stage to boost sensitivity, but it did accumulate charge from light over the whole frame time. There was an alternative approach, Farnsworth's image dissector, which also needed huge amounts of light. It had a photo-multiplier stage but only the light from the dot currently being scanned by the scanning beam mattered.
So RCA threw money at the problem. They developed the image orthicon camera tube. It combined the advantages of both approaches, integration over a full frame and a photo-multiplier stage. It was insanely expensive, costing thousands of dollars in the 1950s. It needed about twenty different voltages and multiple independent power supplies. It worked reasonably well.
Then came color. That took three image orthicons (RGB) inside a box the size of a small refrigerator, along with optics to split the incoming image into three colors. Plus a rack of analog electronics per camera, connected through a large cable. Plus an alignment problem from hell, with a huge number of adjustment knobs.[1]
Scaling worked! From the 1950s to the 1970s, color TV broadcasting was powered by those monsters.
Now, independent of this, other people were working on other approaches. Vidicons, silicon detectors inside camera tubes, and finally fully solid state image sensors. Eventually, in 1983, after thirty years of image orthicon monsters, RCA produced a usable solid state camera. Today, a cell phone camera costs about US$10 in bulk, and it's far better than the monsters.
It took thirty years of low-priority development to get past the giant image orthicon cameras.
Electronic fluorescent lamp ballasts had a different problem. That was a suppressed invention. The magnetic ballast industry was dominated by Magnetek. Someone invented a solid state replacement, which was smaller and cheaper, and licensed their patent to Magnetek for royalties. Magnetek just sat on it, not wanting to gut their core market. The inventor hired Townsend, Townsend and Crew in Palo Alto and sued. (Read this in the law firm's brag book in their office.) Eventually, the inventor got a hundred million or so, solid state ballasts went into production, and Magnetek seems to be gone. (The company called Magnetek today, which makes crane controls, is an offshoot of a Litton unit, and seems to be unrelated.)
There are also parallels from the steam locomotive industry, the Baldwin Locomotive Works being the best known. Baldwin actually built the first Diesel locomotive. They made a few half-hearted tries at Diesels over several decades, then gave up and went back to steam. Baldwin ceased all locomotive production in 1956. The corporation lingered on for a while, not doing much.
[1] https://www.tvcameramuseum.org/rca/tk40a/pix/rcatk40aart.pdf
(So it must be those Orthicon cameras which cause the distinctive weirdnesses of early TV images like the reflections from highly-polished surfaces (such as the bells of brass instruments and saxophones) which gleam black? https://www.youtube.com/watch?v=2uTQQKU6ge0&t=111s )
Appears so:
http://www.electricalfun.com/electroncafe/Image_Orthicon_tub...
(Thanks!)
Its anywhere you have finite resources, time, people, knowledge etc and have to take a call on how to split them - Explore-Exploit Tradeoff.
The LA Times has an article on the Magnetek verdict: https://www.latimes.com/archives/la-xpm-1990-01-10-fi-282-st...
Nice, it got a press mention, with a quote from Crew, one of the law firm's partners.
Suppressed inventions are very rare, but that's one of them.
One of the few others is Airadar, from 1973.[1] Not "AI", "air radar". This was a phased-array radar for small aircraft. Way ahead of its time. So far ahead that the USAF was upset about it being available commercial before it was in use by the military.
[1] https://books.google.com/books?id=NWzlTqj0gQ4C&pg=PA64&lr=&r...
Suppressed inventions aren't "very rare" per se. You simply don't hear about those that aren't being rescinded while the inventor is still alive.
According to official data from the U.S. Patent and Trademark Office (USPTO), there were 6,653 active secrecy orders in effect at the end of fiscal year 2026. Roughly 100 to 350 new secrecy orders are added every year. Many remain locked away for decades or more.
If the U.S. government puts a secrecy order on a patent application against the inventor's will, they are legally forbidden from talking about it. The inventor cannot talk about how it works, publish research on it, sell it, or market it. They cannot even discuss the exact classified details with a regular lawyer unless that attorney has the proper security clearance.
[dead]
Why are you Claude
For as long as the big companies are committed to focusing on scaling LLMs it does seem reasonable that they don't want their overall efforts led by people like LeCun or Hassabis who don't particularly want to scale LLMs. But it does feel as if there's something a bit extra about many LLM maxis' active hostility to deviationists.
I think Hassabis’ position is far more nuanced than LeCun’s. From what I can see, LeCun believes LLMs are a dead end for AGI while Hassabis position seems to be that some new technique is needed combined with LLMs.
I am not sure if LLMs will stay eventually, but I am totally sure that they will be used to develop new types of models
JEPA's method isn't dead, non-generative self-supervised learning that predicts abstract representations (embeddings) of missing information. It is understanding latent structures rather than creating content.
Even if he was wrong about those statements, exploring other types of models isn't necessarily a bad thing. Especially if we can find ones that perform well with less demanding hardware requirements.
Its not a bad thing, but exploratory research as opposed to reaearch that finetunes along a path already paved by exploratory research is something that doesn’t usually happen much in for-profit product-oriented firms when there is such a path available; it happens in academia and corporate labs pick it up once it reaches the point where the basics show commercial promise. CS is a little bit of an exception because the capital cost of basic research isn’t, proportionally, as high, and because the promise-to-product path is often short enough that having done the basic work in house can be a meaningful advantage in getting to market first (though, see Google on much of modern AI for a counterexample).
But even in computing, when there is a promising way forward that the basic reaearch has been done on, its hard to justify side quests.
Those exploratory "side quests" in industry are how we got the transistor, the liquid crystal display, Shannon's information theory, the laser and practical photovoltaic cells, among many other foundational technologies upon which we rely today.
Blue LEDs too, which have transformed lighting and displays.
Blue LEDs have been a massive net negative for humanity. Blue light hazard is very real (why blue lasers kill eyesight instantly and green and red often don’t in accidental exposure).
All display technology past CRTs has been a mistake. The vectrex in the 1980s had better refresh rates, color fidelity, beautiful phosphor glow, etc - and we abandoned it for a horrible pixel based future we are now in. We should return to vectors and lasers.
Laser printers were a side project that almost got the inventor fired.
And then they made back 200x the operating expenses of the lab (Xerox PARC) they came from over its entire lifetime.
Jet engines, too.
> and corporate labs pick it up once it reaches the point where the basics show commercial promise.
How does that work when the research team moved elsewhere? Re-hire talent? That takes time - the 1 thing you don't have in a fast-moving field like AI. Post sums it up nicely:
"Do the math—how much is it worth to halve training and inference cost given that the industry has investments approaching a trillion dollars?"
How much is these researchers' salaries, compared to what MSFT spends on compute? Chances are it's peanuts in comparison.
Yes but you forget that spending on compute is spending on machines that the company owns while spending on researchers is spending on other apes that trigger the social hierarchies of executives brains and who have the dreaded key word of “rights”
Companies will try to ruthlessly rid themselves of labor as long as those members of labor have agency and rights.
If someone disagrees they can convince me by explaining why a company wouldn’t choose to spend X on a machine they wholly own vs X on an employee they don’t.
There’s some room for argument that employees can do more than machines can but in the current AI zeitgeist arguing that we have machines with equivalent to human intelligence I don’t see how you wouldn’t see this coming if you believed the zeitgeist.
If compute:people were roughly equal spend, I'd agree with that thinking. But we all know AI's 100s of billions $$ spend is on datacenters+GPUs, not people.
In a way this suggests that MS doesn't aim to be a leader in AI <the core tech>. Only in its applications + related services (read: profit to be had there).
Nothing wrong with that btw. But understandable that fundamental researchers then pack up & leave.
> But we all know AI's 100s of billions $$ spend is on datacenters+GPUs, not people.
That’s bolstering my argument. Companies are willing to spend a shit ton on machines. They are not only unwilling to spend on labor but are actively seeking ways to remove labor spend, even though this new technology supposedly makes every unit of labor significantly more productive and thus profitable.
This is a cultural choice not an economic one.
If I see a company laying off people while admitting they’ve maximized a niche and automation lets them increase profitably, or I see them go on a hiring spree because labor is now netting them 10x-100x-1000x-whatever productivity increase they claim, then I will believe them about the value add of AI.
Currently I see companies using AI’s productivity increase as a reason to cut labor and achieve roughly the same productivity.
They aren’t competing, they just want to cut out workers.
> CS is a little bit of an exception because the capital cost of basic research isn’t, proportionally, as high, and because the promise-to-product path is often short enough that having done the basic work in house can be a meaningful advantage in getting to market first (though, see Google on much of modern AI for a counterexample)
...And because they have monopoly money. Basically, research labs are usually vanity spending for corporations with too much money. Xerox parc was funded by a monopoly. Bell labs was funded by a monopoly. Google funds its moonshots with its monopoly.
Bell and Xerox became powerful because they invented game-changing technologies. They were fundamentally research-based organizations from the start.
> it happens in academia and corporate labs pick it up once it reaches the point where the basics show commercial promise.
Ah, well, good thing we've almost entirely defunded academic science.
Remind me again where the transformer comes from?
Or residual networks / connections.
It’s a simple exploration/exploitation trade off. Always explore, never exploit is probably not the optimal strategy every time.
It seems to me that everyone is in "always exploit" mode.
Seems weird when it's become apparent that LLMs aren't really special sauce outside of just having the compute.
The right idea usually comes from a wrong one.
The issue isn't the exploration itself, it's the opportunity cost. When a figure of Yann LeCun's influence pushes a specific direction, it pulls key talent and resources away from mainstream LLM work. That kind of organizational distraction has a massive cost, and would've stopped Meta from delivering a polished product like Muse sooner.
The mainstream is wrong and in a lot of ways hazardous. He had every right to say he wasn't willing to sign is name on the wrong technology when he new better. When you hire smart people you listen to them. Elsewise you get what you deserve.
>when he new better
He was against LLMs, reinforcement learning, etc. Claiming that those are a dead-end. He was wrong. His own "JEPA" ideas have basically gone nowhere. Meta was right in moving on without him.
I'm not sure he was wrong. No one has solved the fundamental problem with LLMs, which is that they cannot be trusted.
Ok, so we're talking about applying connectionist systems/models to the messiness of the real world. And "trust" isn't an on-off switch. The goal is to build systems that are smarter and more general. It's a process. I don't think even Yann LeCun has made any strong claims beyond that. He just thinks that his way of doing deep learning is better.
Most humans can not be trusted either, that's why you have processes in place to test and review work
Sure. And we made efforts over millenia to ensure that.
Let me know when LLM's get those. Otherwise I trust it as much as a human under anarchic rule. Probably less because I know exactly who controls it.
It's not a distraction if the research pans out and then they become the market leader for the next several years or more.
The job of the CEO is to assess the research risk correctly. Zuck made the right bet here to replace him.
The guy that named his company after his own diversion?
It's his call for sure, but rather funny as an argument.
The fact that someone made a mistake before doesn't mean that they will always make wrong calls.
not as a matter of certainty, but it certainly increases the bayesian prior probability of their calls being wrong
Maybe he is keeping the name as a constant reminder of the mistaken. Like Berkshire Hathaway: Warren Buffett says Berkshire Hathaway is the “dumbest” stock he ever bought. https://www.cnbc.com/2010/10/18/warren-buffett-buying-berksh...
Oh yeah? How's the metaverse doing?
Ok yeah but it’s not a certain bet. Resources are limited and you gotta allocate them efficiently.
10% chance to be market leader doesn’t sound as good as 60% second place to them for example
This is a prime example of why large companies have a problem innovating.
Once you get close to the top, the risks don't look worth it.
> It's not a distraction if the research pans out
It is because even if it would pan out, it would mean temporarily losing AI leadership, reducing stock price, losing the ability to hire talent, and by virtue of that losing the opportunity permanently.
LeCun wasn't wrong. He simply had higher standards for results and accountability that LLMs are, by design, incapable of reaching.
Zuckerberg wanted to placate his shareholders with a shiny new thing now.
would've stopped Meta from delivering a polished product like Muse sooner.
Muse is not polished, unless usability and programming standards have fallen so far in the LLM vibe coding era that a pre-alpha state is now considered polished.
You will find brilliant scientists in every discipline who are not the least bit interested in a particular branch of research because they/their lab didn’t invent it. That’s fine for public research but not helpful private research.
Compared to the user buying a Mac mini and following a tutorial off Twitter to install some random shit they found online?
Muse is imperfect, as is all software, but calling it pre-alpha is a bit much.
Odd. As I see it he was one of the few right about everything. This is also taking his comments about transformers scaling out of context. He specifically said they won’t lead to human level ai and they haven’t, despite exhausting the training data.
AGI will require architecture changes not simply more data.
I'm quite confused by what you mean human level AI here. Maybe we haven't hit Einstein but we blew past the average person a long time ago.
It is very jagged and some people don't consider that human level AGI - it can solve Maths at the genius level just like Deep Blue could play chess at the grandmaster level. At this point it is simply an opinion and I can see the other side too.
Sorry, but what? We haven't done anything beyond strapping a large language model to a "harness".
At no point do I ever confuse a human being with a Claude session.
Even if I'm somehow fooled or confused for a minute, the context window / hallucinations / Claudeish will immediately give it away.
Sure, we passed the turing test for chatbots but we are very far away from AGI.
That could be how human intelligence stacks, search retrieval (LLM) with learned approaches to processing that info (harness). Maybe you are insulated from the average person. The frontier models easily surpass their intellectual ability to absorb new information and apply it to novel situations.
The average person will also hallucinate things and make all sorts of causality mistakes without realizing it AND refuse evidence of such whereas models tend to acquiesce to evidence. A model with the right harness will outperform any of these people at their current desk jobs. To me that is general intelligence. Most of us on HN tend to far surpass this general intelligence and likely experience a false consensus or assumed similarity bias.
Sorry but I stopped reading about halfway through. This is an insane worldview and you will be impossible to converse with.
> but we blew past the average person a long time ago.
We did?
With this perspective on LLMs, you are presumably often "quite confused" if you interact with them to a significant extent.
It's silly since people want to move fast, but we're really not good at recognizing what has actually allowed us to move forward in the past.
[0] He speaks about Bitter Lesson directly at 9:57 but the earlier part gives better context: https://www.youtube.com/watch?v=21EYKqUsPfg
[1] An essentially meaningless statement. Near meaningless because it is not objective.
[2] You compare a novel technique against a technique that has had hundreds or thousands of researchers optimizing over several years. That's not even remotely fair.
100%. You can both commercialize transformers while still exploring new paths. Having one of the world's top researchers doing productization work is a huge waste of their time.
You also see this whenever news about novel but small models comes out. People are instantly dismissive. "Oh it's small, who cares?" Of course it's small, why waste resources testing it on a larger scale before making sure it works well on a small scale? Would people rather news not come out until it's SOTA model?
All statements he made about world models did not already turn out to be wrong. He may be right in the long term.
LLMs work well for search problems with the help of a strong verifier.
As it turns out, every problem is ultimately a search problem. But not every domain has an easily available verifier.
Absolutely agree fwiw. Scale out over LeCunns desired exploratory research was absolutely the right call. Called out because it seems like MS is making a similar call.
The author states "I consider the scaling-only focus to be both misdirected and not fun.". I get the point but also totally get the focus on scale out for now.
I empathize with LeCunn but he is a prime example of why scientists often make poor leaders. It’s too tempting to always do the cool and interesting thing instead of drilling down on the boring but reliable thing that actually creates value. There is of course a place for this kind of eternal curiosity and I hope he’s found it, but hyper-competitive industry may not be it.
Ironically, Yann is kind of known in academic circles for being the guy working on boring, unsexy stuff (deep learning) when all the other academics were doing sexy cutting edge SVMs and stuff. That’s probably why he got the Meta job in the first place.
Yeah, being a trendsetter pretty much means you had to be working on something boring and unsexy at first. I can't really remember the last time a major innovation came because someone knew exactly what to invent, everyone agreed it was important, and we all ran full speed ahead on it.
Perhaps the atomic bomb? Not being alarmist, but that seems like one of the few instances in modern history. I also thought of the space race, but even rockets as an innovative concept started centuries beforehand as a tool of war.
They have different jobs.
For a corporation it is to make money.
For a researcher it is to explore and illuminate. That this exploration and illumination often leads to a landfall of new wealth generation is a happy accident. This ensures that the monied class would bet some of their capital on them, at least the wise ones.
I'm becoming more sympathetic to LeCun's views, but being able to train something that acts the tiniest bit like those who made the training data (authors, professors, researchers, etc) is an absolute juggernaut and will be for some time.
>LLMs won't be able to do this or that) turned out to be wrong
Where can I read more about this? I have seen people make this claim for a while, and when I looked deep, it used to turn out to be wrong.
You can't. It's AI booster propaganda. Yann LeCun obviously understands transformers better than random anons on this venture capital fan forum. This means he also understands their limitations, which is too complex for people who just want their preferred billionaire to win.
> "I take an object, I put it on the table, and I push the table... there's no text in the world, I believe, that explains this... your GPT-5000, or whatever it is, it's never gonna learn about this."
Lex fridman podcast 2022
ChatGPT answers this just fine.
He also said they will be “ Writing aids, not much more" and they “Can't plan.”
No one said hes an idiot. He was just clearly wrong about LLM capabilities.
>> "I take an object, I put it on the table, and I push the table... there's no text in the world, I believe, that explains this... your GPT-5000, or whatever it is, it's never gonna learn about this."
As far as I know, those things are still true. Unless there is text in the training material that the text in the prompt can map to, or there is text in training material that describe how to map text in the prompt to text in the training material, LLMs cannot process the prompt.
So may be he is wrong about what could be in the training material. But I don't think he is wrong in the fundamental..
This is sort of what I mention in my original comment. These claims, when pushed, turns out not to hold water...
Yann is smart but obviously not omniscient, I don't think it does the ML community a service to pretend that elders are never incorrect (or even badly incorrect). This type of thing is how we got the cult around Schmidhubering.
They surprised almost everyone, but he remained skeptical well after it was clear we could not put boundaries on the capabilities gained from scaling, and also from new RL techniques. Still, I do not think he is wrong about world models and continuous learning being fundamental to AGI, though the latter issue he was himself dismissive of until relatively recently.
> Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong
You're being uncharitable. He would have been right if Chain of Thought hadn't been pulled out of a hat, and no one saw it coming, or how u reasonably effective it would be.
Today, no one can know if there will be a future step-change improvement to LLMs at the same level as CoT, and conversely, no one can be absolutely certein there won't be - so it shouldn't be a firable offense because neither position obvious; even though it will look like it in hindsight.
The most confusing thing to me about AI and LLM progress in the past 4 years is that I can't clearly see how all novel patterns add up, or what was the right point.
I think adding harness + MCP + chain of thought + specialized data generation resulted in the world where we have now - the one where you don't have to type code manually, but sometimes do still have to think and sometimes don't.
But LLMs without all this tooling and calls under the hood are still kind of... LLMs. If there's a space to hallucinate - they will hallucinate.
CoT wasn't pulled out of a hat. It was obvious to any researcher that it was needed. It just took some time and stronger base models and longer context to actually be implemented. People were already doing "CoT" by telling models to think step by step before answering, primitive harnesses and similar tricks.
Even a nobody like me was playing with a sort of CoT with llama finetunes before o1 came out.
Only this year I'd say we picked all the low hanging fruit that has been obvious for a while that was needed.
Hi, I am curious, what do you think are the parts that are obvious but still work in progress? Maybe continuous multimodal input?
I'm not fully up to date on AI research anymore, however I think the obvious stuff has already been implemented, but there's still a lot of room for improvement and scaling for these current approaches.
For verifiable tasks like coding and math, you can keep going, doing more and more RL. You have the AI generate hard problems, then try to solve them and reward correct solutions. We basically have RSI for verifiable tasks, we just need to keep the wheel turning.
TTT
I would take it from other angle.
Meta AI is actually frontier lab now with LLMs. Muse Wonderfull is usable with multi agent orchestration (where other labs struggle).
I do not think that would happen with pure research and without single focus!
In the short term. In the long term, his research on learning will be the next revolution.
Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong. They were right to fire him.
The idea that someone should be fired just because they were wrong seems to be something that would cause issues if one wanted to think long term.
This approach will make people afraid of being wrong. It will lead them to avoid challenging dogmatic assumptions in large organizations and inevitably lead to group think.
People really shouldn’t be punished for saying wrong things. Long term it’s better for a company if they aren’t punished for that. Helps to let people freely criticize dogmas or propose new ones.
Depends on your role. If youre in charge of research in a hundred billion dollar org you should absolutely be fired if you devote your org to pie in the sky research instead of creating tangible value. If your job is just doing the research obviously you shouldnt be punished if your experiments dont work. Yann was fired because his main job was balancing product development with research and he didnt do that.
None of these second-rate LLMs are creating tangible value. Nobody has made any money developing LLMs yet, and when the capex runway ends, the second-rate models won't see much (if any) use at all, compute is too precious for that.
If what you say is true then they would have abandoned LLMs. You're saying that it would have benefited them?
It's Meta, lol. They wasted 80 billion $ on Zuck's fever dream called Metaverse.
It's amusing that their grand dream was basically a clone of Roblox, just with better VR tech and far more advanced graphics. But thing is, kids playing Roblox absolutely do not care about graphics.
Well, don’t care about graphics yet. Unclear what will happen once those gamers hit 20, 30 years old.
Musk initially started his rocket company by going to Russia and trying to buy Russian rockets. The Russians spat on his shoe.
The rest is history.
That at least was a reasonable idea. The Russians had a lot of old Soviet era ICBMs they had no need for anymore (plus, IIRC an agreement with the US to disassemble them), so using their engines as a cheap powerplant until SpaceX nailed reusability made sense.
That failed and SpaceX was then forced to develop its own brand new engines.
Exactly.
Yann LeCun has the syndrome where someone who sees the value of B (non-text modalities, i.e. physical reality) cannot stop themselves from constantly critiquing A (LLMs), even though clearly A has its own advantages.
I think he is a smart guy. But he constantly deploys the crackpot's PR playbook.
It's not so much they _turned_ wrong, they were obviously wrong at the time he made statements, or shortly after. He is just extremely biased against transformers, and it's easy to see why - transformers are already invented, and as a scientist he wants to invent something new.
More specifically:
* he said that autoregressive transformer is fundamentally flawed: if it makes one error, all further generated tokens make no sense; that's, obviously, bullshit: there are many ways to implement backtracking, and there are many papers demonstrating backtracking at a time he did that lecture. And, of course, it just doesn't matter - reasoning model can just write "Oh, that's wrong. Let's try a different thing". - it's just not derailed by one mistake.
* he said that GPT-1000 can't possible have understanding of physical world, when GPT-4 already existing at that time could solve a problem he described.
So it's not like he was bad at predicting things, he's intellectually dishonest.
Yeah that slide of him "proving" autoregressive is flawed was really the nail in the coffin. I stopped listening to him after that.
Yann LeCun was not the "head of AI", he just ran a different research group.
Umm no. There’s the old factory way to build things like Ford model T; and then there’s the f*king Toyota method!
Which is superior, the big f*king monstrosity or the elegant on-demand, small batches, reduce waste?
Meta will go in the shitter and Deepseek will make the small batches.
TPS (and Lean, which it played a role in inspiring) is largely about finding what works empirically and ruthlesslessly optimizing around it.
It does (Lean at least) incorporate controlled experimentation, but tightly controlled with a both a laser focus on what the goal is and tight resource constraints. And Lean (at least several writers on Lean) favors taking multiple paths with a a quick and low risk and a longer-payoff, higher-risk, higher reward path for some efforts—one source used the US Navy’s two-path approach to SLBMs with Polaris and Poseidon as a model for this—but that's different than “maybe there is something down this path” basic research, which is outside of the concern a production model like TPS/Lean and often completely outside the model of a product-oriented firm; basic research is something firms consume and incorporate more than generate, because until the basics have been proven out, the rationally perceived chance of any particular avenue paying off is near zero for any firm.
Ruthlessly optimizing isn't the best goal, because people "optimize" all the quality out, and lack of quality leads to lack of sales and increased cost. Ford/GM may have "optimized" their line for efficiency, but the increase in defects caused more losses. Companies that practice Lean tend to forget this and focus only on speed and cost, leading to enshittification. TPS focuses more on quality and its effects on lowering cost and increasing sales.
Toyota is also a good example of what not to do. Their lack of earlier research investment in EVs has led to being severely behind the market, in addition to their lack of focus on quality causing poorer sales (their quality used to be one of the only reasons to buy them). KIA basically stole their playbook by creating loads of good-quality cheap hybrids and EVs. Toyota could've had that market if the invested more ahead of time in research.
And that's what's happening with AI. While there are plenty of players now, it's the ones that aren't looking ahead that will be left behind. Nothing scales forever; when you hit the ceiling, you're trapped in a box of your own making.
[dead]
This is a dumb question since I know very little about LLMs and transformers but I keep hearing similar things such as "more GPUs/parameters/compute" leads to a predictable scaling in intelligence.
Is that observed scaling part of some equation or law? Is there a limit to how far that scaling goes before a ceiling or diminishing returns hit? Or is it simply based on more GPUs/parameters/compute plus how much human knowledge is in existence and will scale predictably forever?
There are some neural scaling laws: https://en.wikipedia.org/wiki/Neural_scaling_law
And you can in practice often see a large network follow an expected training trajectory for loss/entropy.
In practice data is getting more sparse than compute, so you are usually overtraining using more tokens and being suboptimal.
It should be noted that these laws aren't, well, laws; "a neural scaling law is an empirical scaling law that describes how neural network performance changes as key factors are scaled up or down.". Notice the word empirical here.
These empirical observations on quantities which aren't "intelligence" since no one has yet defined intelligence.
Another dumb question: can the model be smarter than the data?
I wonder about this a lot.
There are all these publications from OpenAI and Anthropic that are solving complex math & science problems.
Are they just "brute forcing" it? Able to try a lot more "known" or reasonable approaches given existing knowledge?
Or are they showing "emergent intelligence" and able to try genuinely new approaches that can't be expected from the knowledge in their dataset?
> Another dumb question: can the model be smarter than the data?
There's a mathematical sense in which it cannot be: the data-processing inequality (https://en.wikipedia.org/wiki/Data_processing_inequality). The extent to which that sense is totally counterintuitive to how things normally work in ML nowadays has led to investigating different foundations for information theory: https://arxiv.org/abs/2601.03220.
Data isn't "smart", it doesn't contain intelligence. I would argue that LLMs are functionally intelligent, they have learned to what functionally amounts to thinking by training on text (and images, in some cases). They are capable of performing experiments autonomously and make decisions based on out-of-distribution data. Besides (current) lack of online learning, they are limited by their biases and working memory - but so are humans, and I'd like to see a convincing argument on why these mathematical laws should not also apply to humans, which are arguably intelligent.
"Smarter" is ill-defined.
Are you smart if you see that cognitive scientists had something figured out, connected it to something in your field of chemistry and made a breakthrough treatment for some disease?
Does any of our existing intelligence build on completely unknown information? You could argue that any new solution/emergent intelligence is an association of existing information in new ways.
It also brings us to what is brute forcing? Humans are brute forcing stuff most of the time. You will have 10 stupid thoughts for every good thought. 10 stupid ideas for every good idea, you will discard most of them so fast that you won't think of that you did that. It's very rare, if not near impossible, that a non-mathematician with no prior knowledge, goes into a room on day 1 of deciding to think about a conjecture they don't know anything about, and solve it on the blackboard. They need to use their collect associations, formalize their axioms, think of first principles (known information), and associate it.
So to answer your question you need to answer:
1) What does "smarter" refer to? Is it accumulation of information about the universe?(if yes, a database is smart) Is it an efficient, abstracted representation of accumulated information? Is it prediction of the future state of the universe given a learned abstraction?
Or going further, is it solving tasks useful for the human in the world? Is it acting as an agent with it's own goals and interests in the universe? What if the goals are hard for humans to understand?
Does it count as smart if you make an evolutionary approach that spawn 1000 instances and the most fit one survives and spawns 1000 mutations of itself?
2) What does "genuinely new approaches" refer to? Is it fine to use math axioms to prove things? Does interpolation between two existing solutions count as new? At what granularity of concept does novelty happen (Does using two higher abstraction concepts, like combining some new, modern neural network architecture with a database count as a new thing?), even if the underlying lower abstraction of the same information can be expressed as storage, latency and compute, and is not novel?
---
As you can imagine, this is all not really defined or agreed upon. Here are some examples of projects that are set up to drive these ideas:
The Hutter Prize poses that intelligence is defined by compression, so if you manage to compress the English Wikipedia efficiently, you are closer to better intelligence (http://prize.hutter1.net/)
Compute: Intel CPU
The ARC-AGI foundation poses that being able to be goal-oriented and understand higher level abstractions is intelligence: https://arcprize.org/leaderboard
Compute: NVIDIA T40s
Modded NanoGPT competition poses that intelligence is minimizing entropy/loss over a dataset in the least amount of steps: https://github.com/kellerjordan/modded-nanogpt
Compute: 8x NVIDIA H100
---
All of this also exists in the context of the compute medium. It could be, that you need a certain capacity of memory/computation to perform whatever we have defined as intelligent tasks. Hutter Prize takes the bet of that to make intelligence, you don't need a lot of computation, an Intel CPU laptop and 50GB RAM should be enough.
A lot of the big labs are instead taking the bet on that you just scale everything, make trillion-sized models and things will fall out of that automatically, for this you will need a very different compute medium (data centers) compared to (a laptop).
[dead]
'Scaling law' is the prediction that increasing the training data and model size will cause the result to be better. It's not a law, but it was the prediction that OpenAI staked their fledgling startup on to get to the ChatGPT moment.
Your phrasing (_observed_ scaling _part of_ some equation) might indicate you assume / think there’s some theoretical equation that relate intelligence to those parameters and the we observe the data to confirm it. As far as I know, there isn’t - it’s purely empirical curve fitting. Train a model with different amount of compute / data / parameters, measure performance, fit curve.
Yes, the Chinchilla scaling laws from 2022: https://arxiv.org/abs/2203.15556 More parameters directly and mechanically means less prediction error for each token.
(This moves the argument to "is being really good at predicting the next token the same as being intelligent?" which has happily diverted many minds for years now)
From your link:
> for every doubling of model size the number of training tokens should also be doubled
Well, I don't think many claim that increasing the model size while also increasing the training tokens doesn't make sense.
The problem is how to get the training tokens...
Interestingly enough labs do train with 100s or even 1000s times more tokens than what's proposed by the paper as an optimal value.
And the answer to your question is basically just “synthetic data”.
The Chinchilla paper is probably the poster child of “paper that aged like milk” in LLM field though. Everyone over the past 4 years have trained with several orders of magnitude more data than the Chinchilla optimal number.
If you don’t know which data has the highest value, though more at it. Chinchilla is probably still true, but with that caveat. And it could be like advertising… you know a lot of it is worthless, but you don’t know which part.
I'm seeing the same pattern at other multinationals with huge R&D departments. Mostly colleagues with academic pedigree in traditional ai research who have been for years in denial that transformers are actually - in combination with recent advanced patterns in software development - the key to ai. it's really simple as that! except they weren't part of the revolution and still don't want to accept the reality that their research has become obsolete.
Is it the key to AI? It probably depends on the squishy definition of AI and what different people want from it. If your definition includes self awareness and sentience I’m not sure it is the key (I’m also not entirely convinced it isn’t, either), but it’s not clear either way in my opinion.
The one thing that is clear is if it can become more cost effective it’s world changing. Even at its subsidized standards now it’s world changing. And it’s reasonably classified as AI IMO, but it’s not self aware.
I think a lot of your general traditional researchers are fundamentally looking to understand ourselves through machines as proxies in a sense. Something self aware that senses, learns, adapts, so on. The current path isn’t there or necessarily clearly the key to achieving that. That may be the disconnect. I think LLMs may give us some hints at that but not the full picture.
Still incredibly powerful for things we once considered nearly impossible.
> still don't want to accept the reality that their research has become obsolete.
I see it that most of legacy research is not obsolete but can be significantly amplified by AI.
> Current top models are mere (mere in big datacenter terms) trillions of parameters right now. ... We'll soon start hitting 10's of trillions and 100's of trillions of parameter models.
Chinchilla Law says no, we wont. We're likely well past the point of diminishing returns but MOE hides a fair bit of it.
We already know that power laws apply to scaling so 10x the size is going to result in minor gains - where costs become a real factor (as if they weren't already).
FTA "The core philosophy there is scaling, and aggressively not research."
This is where the real gains for AI come from - good enough and CHEAP is going to win the day here (brute force, mixture of models). It's the engineers displacing the academics - and that will mean a bloodbath for revenue (see DeepInfer for what the end game looks like).
I'm not an AI booster (in a "should" not "can" sense), but in terms of "can" we are going to see next generation hardware (i.e. Vera Rubin GPU architecture) and improvements in model density (i.e. loosely improvements in intelligence per millions of weights), so the models will get bigger, the models will get more dense, and we may see more "expert models" i.e. in the sense that for a while there were human generalists that knew all of science and mathematics until the 1800s, after a certain point we had to specialize since there was too much to know. The models will need to partially specialize after hitting the hardware ceiling.
Bioinspired algorithms may eventually lead to vastly improved intelligence per a weight too.
I don't think we should do this, but I feel doomed in that I suspect that we "can".
> The models will need to partially specialize after hitting the hardware ceiling.
I'm incredibly skeptical of anything that goes "oh, we need to split AI into specialist systems".
I suspect that a lot of the delta between "our best models" and "base models they're trained on" lies in relatively low amounts of parameter changes driven by SFT and RL.
Model capacity is not the limit - the limit is our ability to drive those changes efficiently. Even things like "good research taste" could be represented as tiny in relation to total model capacity, if they could be trained into a model efficiently.
At least for now, those keys haven't been completely unlocked, so parameter scaling is the primary way corporations are achieving increases in ability. If they hit the hardware ceiling, maybe the research direction changes then.
I forget who said it, but a trenchant insight is that current neural networks are generalized function approximators and essentially they are doing through brute force what biology accomplishes elegantly. Our hardware (both in the sense of our brains in terms of elegance and our computer chips in terms of raw throughput) is amazing.
> so parameter scaling is the primary way corporations are achieving increases in ability
Thinking tokens have nothing to do with parameter count and where some of the biggest gains have come from.
We're really past the point of scaling parameters being viable. We simply lack the content to make "bigger" better at this point.
No, parameter scaling still works because it broadly increases sample efficiency. We don't have the raw data to make a Chinchilla optimal 100T, but that doesn't mean there are no gains in building up to 100T.
I don't think "biology" and "elegantly" belongs in the same sentence. Let alone the one you wrote.
I suspect that the "elegant" workings of the human brain are a pile of hacks. Evolution crammed the brain full of patterns that are a good fit for the specific problems it typically solves - the way curve detectors and scale invariance are a good fit for CNNs.
We don't have millions of years of evolution of non-random init, but what we have is enormous datasets and compute at scale.
The brain is a pile of hacks, but it's still the most advanced thinking machine on the planet and it works on barely any power.
LLMs are very competitive on power use at inference time.
Sure, frontier LLMs require timeslots on massive server arrays. But they also get things done a lot faster than humans do. The net favors the LLM.
What makes sense is to do both.
Race down the transformer highway as fast as possible and explore other AI methods.
One need not preclude the other.
as fast as possible implies spending the resources that way.
Going close to as fast as possible while saving a stream for the other approaches makes sense as an approach, but let's not ignore that there's a compromise/tradeoff here. Might not matter in orgs with near infinite resources
Also in an exponential competition environment, the difference between 100% commitment and 98% commitment is very possibly not even the difference between a win and second place, but outright lapping by the winner and humiliation at the scale of the gap for second place.
For an example of what this looks like at a more pronounced scale just look at the discourse around European models and how far behind Minstral are perceived to be.
My understanding is that the resources are limited and even AI labs with all the funding in the world don't have enough hardware.
Theoretically one need doesn't preclude the other. The OP gives a detailed argument and his own experience of one need precluding the other in practice.
Both of these approaches require quite a bit of institutional support notably.
Except doing both..is..so..hard when the future payoff of one is more unclear and you are faced with quarterly earnings reports.
Look at what happened to Meta where they squandered a lead, had some mixed results, and reorged to more product certainty and away from fundamental research.
it literally does when gpu's are scarce.
That depends on your definition of scale. Yes! We were to literally convert the whole world and its resources into data centers.
But the whole problem is that these things aren't acting like other technological revolutions like the internet or Google where lots of profit comes from relatively little spend. They scale linearly maybe if you are super smart and good about your design.
Meanwhile, citizens all over the world are fighting data centers tooth and nail because we are hitting a scaling limit in terms of the resources people will allow to be put into this highway project and its equally unwise and untethered from reality to pretend this doesn't matter in a strategic sense.
"citizens all over the world" will not lift a finger for CAFOs or trawl fishing. i really don't think anyone "fighting data centers tooth and nail" is doing so because of a belief in the limits of scaling or any understanding of scarcity.
Im at a company with its own set of machine learning engineers, a few separate organizations internally have them so probably a few hundred MLEs in total.
More and more, the talk of using generative models (e.g. transformers) is coming up to replace existing models in all concepts. Anything from understanding data, image recognition, labeling, ranking, etc.
As someone who is not an MLE and really doesnt understand these models well, it seems like the transformer approach is a "one size fits all" solution that may not be optimal for all problem categories but is good enough that its not worth investing in specialized models to get incremental improvements.
Isn't the thing with transformers is that they work really well, but only if you have a lot of training data? If data is sparse, you need more specialized models
> If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Already happening: https://news.ycombinator.com/item?id=50397994
Obviously we'll see how it pans out, but other approaches feel a bit like trying to ignore the bitter lesson.
More parameters without significantly more data fails; it loses the ability to generalize and simply memorizes.
And we’ve consumed pretty much all the data, and that part is getting harder as data is now expensive, Balkanized by owners, litigated and charged for. Couple that with model collapse as models eat their own output (not solved for any models, despite more than a decade of the problem), and add that lots of previously human curated datasets (stack overflow style), and it very well may be the models are not getting much better.
Very large models trained with sgd variants don't tend to over fit, especially when using few epochs. See double descent.
The coding/math breakthroughs were caused by training models on the subset of their own outputs which pass a verifier, see reinforcement learning from verified rewards.
More data or parameters can break double descent; it's not a magic free ride, e.g.,
https://arxiv.org/pdf/1912.02292
There's plenty more. Read Google scholar
That's the historical wisdom, but there are reasons to believe the opposite: that past a certain point more parameters are always better. Basically, that overfitting don't matter, because a large enough model will memorize and build the general solution simultaneously. Then weight decay and SGD noise encourage forgetting of the memorized solution so the general solution takes over.
https://youtu.be/lKoJJxjUfdw is a really excellent lecture that changed how I think about it.
This is what I thought a year ago, too. But how does it explain what has happened since?
is scaling working? It is also my understanding that Opus 5.5 is probably a well under 10T parameter model. Probably a similar size class to Fable 5 earlier. And supposedly even GPT-4 was a 1T model. Clearly most of the gains are coming from something other than scaling up the model size. Yes, scaling up RL counts as scaling too, but the fact that labs have pivoted to a different axis of scaling is evidence that it's not just a "straight highway" to improve LLMs.
From the outside it seemed like GPT 4.5 was the apogee of parameter size scaling and since then a huge investment in RLVR became the new scaling axis.
So yeah, LLMs keep getting more powerful/useful but because of an industry wide pivot not a "straight highway" as imagined in the GPT-4 era, when simply going bigger and bigger seemed like a reasonable roadmap for the foreseeable future.
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters. If the scaling continues to hold we'll be asking the AI to do the exploratory research for us soon enough.
Where's the money coming for those 10X and 100X ventures?
From mass layoffs of tech company employees.
There’s a good example of a model that won’t scale.
It's worth it to keep exploiting transformers while doing just a bit of exploration of alternatives because they work so well. Heck, just exploiting transformers more will accelerate exploration of alternatives...
100 trillion parameters doesn't make any sense if your dataset is ~30 trillion tokens. My uneducated guess is 1/10 compression ratio is good and that point to ~300 to ~1000 trillion tokens, which again, doesn't make any practical sense.
Depending on how you count, the human brain is between 100 trillion and 1 quadrillion parameters, which is one guess as to why we're so much more sample efficient at learning than LLMs are.
Synapses and LLM parameters are not equivalent.
In Silicon Valley they are.
One key difference between humans and (current) LLMs is that we can actually judge the material we learn from at the time we're learning it, contextualize it and discard garbage when we spot it.
LLMs need an awful lot of handholding all the time, meanwhile humans get pretty good at it on their own.
Unless quantum effects do infact play a role in consciousness. It’s not clear they do by any means, but if it did that could shift the understanding of our own minds and consciousness drastically. Otherwise we are just a dense complex network we can reproduce with enough money.
I mean we literally do this every few seconds already with natural birth processes.
I don't think you even need to go to hypothesized quantum stuff. A single biological neuron in isolation is a surprisingly complex system that resembles a small artificial neural net more than an artificial neuron in many ways. On top of an entire feedback system from diffusion of neurotransmitters which doesn't cleanly map to ANNs.
Trying to estimate equivalent "parameter size" of the brain when we only superficially understand behavior of individual neurons let alone how intelligence emerges from the connectome seems too speculative to be a particularly useful metric.
The extra parameters are useful for optimization. Deep-learning relies on highly singular solutions.
Yup, we havent seen any limits yet. We have discovered a general learning algorithm. We are in for big things
Same in deepmind too. Bit of a shuffle as well to the frontier AI group, jeff dean and prateek jain headed it now jeff is gone, so prateek has moved to MTV. And most of the different research groups are all in some sense under deepmind, moved to LLMs and in no-publication mode now.
I wouldn't interpret it as some sort of philosophical position. After a few years once the current buildout settles down they will go back to the usual setup. In the meantime, my alma mater is poaching some of the disgruntled publishing-oriented folks from there.
Right now it is stupid to not dedicate all organisational resources towards LLMs, when you are lighting your free cash flow on fire and issuing bonds like there's no tomorrow, you might as well have all your smart people working on it. JEPA can wait a few years.
Addendum: the amount of people in this thread misinterpreting what yann lecun said is insane.
> Hey this Opus 5.5 with the increased parameters seems much better than the last one"
Is this true?
I barely notice any improvements ever since... 4.6? At least for software development.
Opus makes the same sort of mistakes and has the same sort of strengths ever since that point in my perception. My impressions is that increasing the number of parameters clearly has diminishing returns.
In fact, the best improvement I got from Claude was Haiku 5.5. Good cheap workhorse to implement plans.
Well 4.6 was released Feb this year.
As in your statement to me indicates that you clearly see the scaling working over longer timespans. You chose a relatively recent 2026 model as your baseline after all, not a 2025 model or earlier. You clearly see that the 2026 model is better than ones in the past.
Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together? If you're not anecdotally seeing improvement between an early 2026 model and one from ~6 months later perhaps just give it another iteration? Personally i do see 5.5 as much better but i can understand not seeing a lot of difference to 4.6 since they literally aren't that far apart in releases.
> Might i suggest that you perhaps just don't see the improvement as clearly when comparing iterations that are fairly close together?
No, you can't. Because you invalidate your argument by adding later:
> Personally i do see 5.5 as much better
It has to be one or the other.
Either each minor release Anthropic farts is a major improvement from what came before, or I don't see much difference because it has been very little time in between releases. Holding both arguments as true would require me to be excessively apologetic for a company that desires a 2T IPO.
And to be clear, I do think that there is an argument that not much time has passed since Opus 4.6 and Chat GPT 5.3 Codex days.
Part of this conundrum may be because the models reached a good enough point for software development. Maybe walking 80% of the way was easier than the final 20%. I am nowadays more excited about how cheap models can be useful for it (hence my positive remark about Haiku 5.5 and Chinese models in general).
These corporations have an obscene amount of money
This is a false dichotomy if I ever saw one
Why not hedge and just keep an ai skunkworks team working? These companies are big enough that they can have these people far away from the main work without distraction or without starving those teams of bandwidth
The companies would rather have all the skunkworks people working on making the LLMs better.
> "Hey this Opus 5.5 with the increased parameters seems much better than the last one". Well that's a trillion parameter model.
Opus is 5T, not 1T.
> We'll soon start hitting 10's of trillions
Fable is already 10T.
> "Foot down, there's a straight highway ahead don't divert". I have the view this makes sense too.
The roadblock they ignored is politics.
Three days ago the states where most of the new datacenters are being built (Texas, Ohio, Georgia and others) elected governors who campaigned on halting datacenter construction. The highway metaphor implies money in => intelligence out. Turns out there's at least one other ingredient, and we just ran out of it.
I'm still convinced the model advancements are largely US firms jingling a set of keys in front of investors/markets/the public to distract while they scramble to build serious patent moats around the harness/orchestration/infrastructure technologies.
I think that we're well into diminishing returns already. The models have improved, but to me the far bigger advances have been in feeding them contexts, building them into loops, etc.
I don't even think much more construction is really necessary in data centers. We can do a lot more with the power/HVAC we have, especially once the ASICs start coming online.
It's like we invented the wheel in 2022, the crowds are fighting over capital to build wheel factories, and savvier tech firms are quietly patenting automobile designs while everyone's focused on making slightly rounder wheels.
When a company pays billions to OpenAI, any activity that does not do a glorious dance in praise and support of that investment is taken to be an act of disloyalty.
This doesn't engage with the article in the precise way that transformer-based improvements and fundamental research that way are not being incorporated into models, at Microsoft AI, which is concerning.
> there's a straight highway ahead
Which many credible, AI researchers believe is already starting up an increasingly steep scaling wall of diminishing returns that eventually ends at a brick wall.
I've been hearing a lot of "credible AI researchers" talk about these diminishing returns of transformer-based architectures since 2023 (at the least), and I'm still waiting to see this promised plateau materialize, because I definitely can't see it in https://metr.org/time-horizons/ or in my personal experience.
Assuming super-linear growth will continue based on the first 3 years of a new technology? Almost every physical, computational and economic system in history, from Dennard scaling in silicon to transonic aviation speed looked super-linear until it hit the inflection point in its eventual S-curve.
A Copernican/Lindy baseline suggests the growth rate in LLM intelligence will peak in around another 2-4 years and I'm leaning on the low side because while the growth of LLM intelligence per dollar and per month are still growing in absolute terms, the growth rate of both seems to already be falling as power-law constraints in compute and data bottlenecks take hold.
There's just not much value to a 'Every technology hit a brick wall eventually' take. Anyone can spout that. Why even bring 'credible AI researchers' if that's all it amounts to. What is relevant is the foresight to predict when and at what capabilities. It doesn't change much of anything if LLMs 'plateau' right after they can replace any human at any job.
This sort of tallies with the claim that every significant technology in the iPhone was initially developed through state funding. Basic research and fundamental inventions are expensive, long term, high risk projects that are generally incompatible with companies driven by quarterly profits and stock price (with some notable exceptions such as pharmaceutical companies).
It seems extremely likely to me that scaling produces some benefits, it will get better at particular things things continue but scaling by itself isn't heading to general intelligence.
I should also say that I use ChatGpt daily and enjoy it for researching algorithms. I tried to set up Claude last week and it/he was so overwhelming as to be unusable. So I think frontier models can be "unwieldy" or something for average use.
> it/he
It
Why wouldn't the explore/exploit balance leave at least a little to the explore side? Huge amounts of money are going into AI, why not save a little bit to keep an eye on the things that don't currently look like they're going to scale up?
Like maybe a couple dozen people on an explore team?
I would say I used to believe how the models would scale, now I'm a bit more skeptical. I would be incredibly surprised if oai, anthropic, etc hadn't at least TRIED this yet. They have the compute to do it. They're either sitting on it and saving it for a rainy day, or there is something wrong with it.
We know nothing about their internal models that, for instance, produced revolutionary results in mathematics. The public no longer gets to see what’s at the frontier, and we may never see much of it. Also, the cost to train these is immense. China literally does not have enough chips in circulation to compete. I’m not sure where US lab capabilities tap out but I suspect they are at least partly bound by GPU production as well at this point.
We know that if this immense cost had in fact been spent, and the internal models really were that revolutionary, these companies would be selling access to it for cash instead of taking on mountains of debt.
We definitely know that.
I agree they would not shut up about it if they had something to reveal, so we probably have a semi-accurate vision of the capabilities if not the implementation details.
> Foot down, get to 10 trillion parameters, see if the scaling holds, if so off to 100 trillion parameters
Moore's law: Computing power is multiplied by 10 every 6 months? Can't wait for the optimization and rationalization phase, where we'll recover a trillionth of that power to run powerful databases.
>> We definitely haven't hit the ceiling and everyone's noticing "Hey this Opus 5.5 with the increased parameters seems much better than the last one".
Opus is better (not necessarily "very much") than Sonnet in coding. That's because of a gigantic engineering effort to make it better at coding. That is a regression: we've gone from "sparks of AGI" and systems that were said to be "genearl-purpose" to systems that do one thing well and everything else... well, there's no real measure of how better Opus is from Sonnet in replying to your emails so it's hard to say.
Same goes for OpenAI in maths. Industry wide, there is a regression from general-purpose to special-purpose systems, like in Good, Old-Fashioned AI. LeCun is absolutely right to complain that this is not the path to general intelligence. I don't think his preferred approach, JEPA, is going anywhere either, but at least he's trying. Other luminaries of the field have expressed similar doubts, e.g. Richard Sutton's tweet was in my news feed yesterday saying that brute force is ngmi and algorithmic advances are needed [1]. Schmidhuber I think has said multiple times something along the lines of "there is no intelligence without physical intelligence".
We are very far from that. The recent trend for Vision-Language-Action Transformers (VLA-Transformers) in robotics has conclusively demonstrated that you can't solve the real world by overfitting to a set of discrete tasks, and there is no source of data that we can harness to repeat the feat of LLMs. Algorithmic advances will be necessary, but as the article suggests, these have now be put on hold.
There is a dumbing down, slowing down of research as all resources are spent to try and scale what we have now. What we have now is not enough for intelligence, but that doesn't matter. The industry's goal is not intelligence, it's making more money than anyone has ever made before; in fact more money than all humans put together have ever made before. That's the current race: who will win the money game. Obviously that's not likely to do anything for intelligence.
____________
[1] Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
https://x.com/RichardSSutton/status/2119142310894334262?a=ry...
> Hey this Opus 5.5 with the increased parameters seems much better than the last one
The LLMs still hallucinate way too much when there's no computer code involved, but the computer programmers in here won't be aware of that.
Do they really? If your experience is with the free version of ChatGPT, I wouldn't be surprised if you're running a MoE with like 8-12B active params and <100B total.
(Just speculation on params, but I find a lot of "LLMs still hallucinate on non-coding problems" come from people using very-cost-optimised free models and drawing the wrong conclusions).
> If your experience is with the free version of ChatGPT,
Even the free version of ChatGPT shouldn't hallucinate, after all it uses a good enough model. And the answer is no, it's a combination of paid and free models coming from OpenAI, Claude and Google's Gemini (with this latter one actually badly miss-identifying a WW2 military dictator a couple of weeks ago, wrote about it here [1] )
[1] https://news.ycombinator.com/reply?id=50256321&goto=threads%...
What if scaling fails to hold at a certain point? Whether that is a steep or gradual dropoff in returns. What happens to the empires of accumulated compute? How much value do they immediately lose once the returns to scaling ceiling is reached?
If/when this happens the value generation will be focused on application and the training compute will be shifted towards inference.
At this point I think it'll be a lot easier to find out what happens then and what to do next if you're using the frontier AI
Aren't there scaling laws with corpus size and parameter count?
> Essentially the messaging it "Foot down, there's a straight highway ahead don't divert".
Feels like this is leaving major breakthroughs on the table for the competition. Sure, transformers might get us much farther. But if someone figures out an approach that unlocks a performance boost or better resolution then they are going to kill the competition.
Feels like the American industry is high on all the capitol being shoveled into the furnace without caring how efficient the furnace is. At some point someone is going to make a breakthrough and burst that bubble. This could crater the value as we are currently betting on the low power efficiency of the models as the barrier of entry and point of profit. Of course that would lead to cheaper AI but is that profitable vs local models once (or if) it becomes affordable to do so?
Are jev models / methods on this straight highway?
Get the transformer to help accelerate research. That's how you get a new architecture.
Anthropic, which is more open about this has shown the numbers, it works.
Ignoramus question: does the 100s of trillions of parameters world still exist with a constrained domestic data center build out? Democrats are signaling they are about to go psycho mode on data center contracts.
If not in the US then in other countries.
Like the sister comment, if not in US then in other countries.
US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
We are repeating the mistake again.
> US rightfully placed a bunch of labor/environmental regulations around manufacturing which moved to China but stupidly didn't place restrictions on goods manufactured with severe labor and environmental violations.
The 'US' isn't a monolithic thing. US market leaders convinced US politicians to open US markets to China, and the lack of labor and environmental regulations was a feature, not a bug in the plan.
> The 'US' isn't a monolithic thing
Yeah, we all know that. What next? You are going to tell me "US politicians" aren't a monolithic thing? Or that we are made up of atoms?
Who voted for the politicians? Who clamored for cheaper goods and fell for the lies?
This refers specifically to "AI Frontiers", a research group within Microsoft Research that focuses on fundamental advances in AI. There is no end in sight to real AI frontiers.
Perhaps "The End of Microsoft Research's AI Frontiers Group" would be a less ambiguous title.
The real news is "The End of Microsoft Research". Langford's group is just one of many that have been parted out over the last year or so. The division that was T&R under Kevin Scott was quietly dissolved about a year ago and MSR was reorged under Azure Security with the AzSec leader as hatchetman for the department. MSR itself doesn't have a future beyond a few years at most I reckon. I don't suspect Kevin Scott will be around much longer either since his department keeps getting reduced and is fully eligible for VEO.
Langford has been at MSR a very long time and has done amazing research doing that time. He's bounced around a bit at MSR during his tenure and it's a shame to lose him. AI Frontiers wasn't around that long.
Much of this is Microsoft performing a very large and uncommunicated strategy update. All updates of MSR reorgs are just that "MSR is becoming more lean and focused" when it's obvious this isn't the case. This gets into the conversation that many Microsoft employees tend to have these days which is "Why does culture no longer matter at Microsoft and why did Satya turn into another heartless CEO". I won't speculate there, but the place isn't nearly as fun as it used to be.
simply owning most of OpenAI via a convertible note on very cheap inference sold to it could turn out to be a bigger deal than the rest of what microsoft does.
Yeah, I went down the rabbit hole scratching my head the whole time. This is about the group at Microsoft Research that was disbanded or reorged into what is now Microsoft Research.
To the author: 100% need this kind of thinking. It’s a balancing act. Historically though, dedicated labs haven’t had the best track record at commercializing their tech. The skillset differs vastly from running a product-focused business.
That said, even within this narrow scope - there are areas ripe for research.
Some of my particular interests include:
(1) Latent thinking/reasoning, which could arguably be more efficient but also opaque is one
(2) Separation of “intelligence” from knowledge - ie. keep the ability to reason like a frontier model and make tool calls (calling the ‘oracle’ or searching the web) but remove all parameters associated with niche knowledge/facts.
These are a couple off the top of my head that are very much aligned to scaling.
Ok we'll use that above. Thanks!
> There is no end in sight to real AI frontiers.
That statement is broad enough to be unfalsifiable, but if you narrow it down just a bit to scaling LLMs, maybe there is an end in sight. At least financial markets are skeptical of there being no end in sight.
If scaling LLMs really was the road to God in a box, then compute demand would be infinite, and nobody building to meet that demand would have a junk debt rating. Did those money people just not get it?
I see this playing the same way in the other big companies that have a pressure to deliver (cough *value for shareholders*)
It makes me sad because there are so many unexplored areas of AI research, even on text generation. Billions are invested in almost one idea, which very realistically isn't the most optimal way to do text generation, but a lot of companies are afraid of taking risks. I still wouldn't put $1b on JEPA but I prefer a world where we can try things like that.
It's a lot easier to explore these other areas with the help of LLMs.
LLMs have very favourable exploration<>consolidation dynamics. An LLM alternative would need to display fundamental improvements at relatively large scales which is extremely challenging for new research.
> there’s a breakdown of the basic deal in MSR: do good work and good things happen
This really hit home for me. This is the contract that held for a *long* time at Microsoft and has been broken and rebroken with current leadership. It’s culture-destroying.
He didn't say they were not doing transformers research. In fact, he specifically mentioned a type of transformer they were working on. And labs like Qwen and DeepSeek (at least) have shown how incredible results can be when you innovate in the area of LLMs.
I think the real point of the post is that office politics and poor management got in the way of innovation due to infighting and poor execution in transfer of technology from research to product.
To me, all signs point to US over investment in hardware capacity to support the "frontier" when in reality the next major breakthroughs are likely radical optimizations that end in increased consumer training and inference efficiency.
I'm not sure how that, combined with open models inching up on frontier intelligence, is compatible with these massive AI IPO valuations?
It seems like closed model AI demand has a lot of contingencies that are ultimately fragile and possibly short sighted.
The article is about a guy leaving Microsoft and the disbanding of an internal team called "AI Frontiers".
Yeah, I'm speaking to "scaling vs research" and where investment is going. Which is relevant to the article. Accelerating scale requires demand, closed frontier demand is what I'm questioning.
And gp's comment is relevant to the particular problems the guy (John Langford) cites in his comment about leaving. IE, that research first winds being only about scaling and then winds-up ignored because "just scale" can be an easier answer.
If they make more efficient models won’t people just use the big data centres to make and run even bigger more powerful models than they can currently support? Or massive swarms of agents to look at all different avenues of a problem?
Do you think humanity is just going to run out of problems that need the smartest intelligence?
If so, then yeah, increasing efficiency 100x means we hit that point much faster.
If not, then it just means we make improvements 100x faster.
Everyone wants radically more efficient AI. That doesn’t mean it’s coming any time soon. And when models are twice as efficient, guess what happens next. We try making one twice as big for the same cost.
Companies with billions of $, stopping R&D on the most revolutionizing technology/ breakthrough is something I can’t understand. I do understand there’s an AI Race to win, and that is with transformer and current LLMs architectures, but if research could find a better, more efficient approach they would re-shape this race, lead it, or even dominate it.
The problem is the "if". There is no guarantee that there's a "better, more efficient approach" to be found.
So far, the attempts to replace the transformer LLM with something else have yielded either "the same transformer LLM but with this one tweak that makes it 10% better" or "a radically different architecture that's a bitch to work with, and performs, best case, about the same as the transformer LLM".
By now, I'm firmly in the camp of "the architecture is good enough". Rifle is fine. Transformer is fine. Get better at the transformer before you look at anything else.
And getting the transformer better may make finding the next thing easier
I think it’s more likely that some future LLM model finds a more efficient architecture than a human does, at this point. So yeah I’d go all in on LLMs too.
They are losing tons of money on it. Simple as.
Me neither. When revenues are growing every year, a fixed chunk of researchers' salaries (maybe growing a few % each year) doesn't really hurt, but R&D offers a nice upside.
Or if you are Microsoft, you throw in the towel on leading in any given industry or technology, and you're happy to be third best at everything and leverage b2b dealmaking to stay relevant.
Big corporations, by nature, are not rational in the decisions they make.
That's why they'll always be room for startups.
Ah yes, startups, bastions of rational thinking.
More like, a large company has to be rational, as the article points out, throwing money at research for uncertain returns is not as straightforward as throwing it at scaling for known returns.
Startups are allowed to be irrational because investors know upfront that what their dollars are getting is a high variance outcome. People buy a stake in Microsoft when they want what they perceive to be a safe investment.
> More like, a large company has to be rational, as the article points out, throwing money at research for uncertain returns is not as straightforward as throwing it at scaling for known returns.
It's arguable whether scaling will give the expected results to justify the investment. Currently, it is a bet, just like funding alternative research is a bet.
But is it a race to zero? Maybe AI will make very expensive things (eg software engineering) very cheap, but there will be enough competition that the margins are tiny and don't actually justify putting crazy dollar amounts into research
That's exactly what's happening. Software -> 0 marginal cost (obviously still need engineers to actually make the architecture). LLMs -> Price of Energy / Amortized Cost of GPUs. "AI Labs" -> Owned by Creditors
From an engineer, who’s devastated by this, i have to admit/raise my concern: why do we still need engineers to make architecture? Why is everyone saying that’s a must? For me the toughest problem was coding a huge repo/pkgs, architecture should be much of an easier task to do. I’ve seen it do sys design, scaled architecture better than 10X engineers. With good planning, context and prospective info -> it can produce extremely good architecture even for complex systems.
Im not sure I would say architecture specifically. Just anything on the input and output.
It cant lmao (skill issue)
Almost as fascinating is that the article references Digg as the source - genuinely hadn't heard of Digg for maybe a decade? I thought the DVD encryption fiasco + reddit had ended it?
Digg was briefly rebooted but once again died mere months after launch. Now they're peddling AI scams and the like.
It got rebooted a couple years back. One of the founders of reddit was involved too lol.
https://en.wikipedia.org/wiki/Digg#2025%E2%80%93present:_The...
Seems like it quickly pivoted quickly to cash in on "AI" hype.
The headline makes good bait for AI skeptics. But this is a team of 50 people with a grandiose name.
The more interesting idea is that MS has abandoned AI research entirely, presumably because their zillions invested in other companies are a better bet in that area.
If MS abandoned a promising idea, pursuing it outside makes perfect sense. I am sure there are investors willing to fund this. This could even be an opening to attack the moat of the current AI frontier model companies.
Curse of money strikes again. $4tn market cap company thinks team of 50 researchers too expensive.
Yeah but to have research be actually useful, you kinda have to have a full-on frontier lab.
Or be content to publish academic papers, which is what this team was doing.
It is notable that with the advancements of 24/7 coding agents with basic research capabilities the premise:
> a clever idea pursued by one person it was likely to arrive late and unconvincing because one person could not move fast enough and the baselines which one person could assemble just were not strong enough to really show that an idea was worth taking flight
may no longer be true. The limiting factors are: frontier model token and compute budgets, researcher skill, drive, foresight and imagination, not the ability to work in teams. Of course, working in teams still helps, and benefits from internal use, peer review and promotion of the research line are certainly there, but compared to some time prior to 2026, it's pretty much not a factor.
Relevant: https://www.quantamagazine.org/when-chatgpt-broke-an-entire-...
On the flip side, frontier models are so good and relatively cheap now that it's possible to pursue your pet research project on your own as long as resources are not a constraint.
Great link, thanks for sharing. I missed that when it came out. I'd love to read something similar about other fields, like chess AI up to 1996.
John was the creator of Vowpal Wabbit, which used to be a popular ML library: https://en.wikipedia.org/wiki/Vowpal_Wabbit
It was my introduction to feature hashing.
I think the dichotomy of open-research vs. commercially-driven innovation might be easier to coordinate, with three stages, not two:
1. Open research: Great creative freedom, with great resource constraints. Lots of ideas, lots of things tried, at insignificant cost relative to commercial work.
2. Results filtering, vetting, regularization: Opportunistic selection of apparently successful work from stage 1, for tentative testing at increasing scale. The best ideas end up with moderate budgets that verify benefits, which are reworked for compatibility before any handoff to stage 3.
3. Commercial efforts: aimed strictly at improving quality and economics of current revenue sources.
Direct negotiation between #1 and #3 involves too much dissonance over globally complementary but locally incompatible values.
#2 can be a relatively small, high payoff group, keeping #1 and #3 from butting heads, while facilitating them both.
The Microsoft AI (MAI) part stood out to me. I bet the politics of being in Mustafa's group were not fun.
Who's this Mustafa and what does a MSR team getting disbanded have do do with it?
It's mentioned in the article as "post-inflection" which means when Mustafa's startup was acquihired, and Mustafa became CEO of msft AI. (Microsoft likes to call the heads of various orgs "CEO".)
Mustafa has since been demoted from head of MAI (which includes applications) to just head of the superintelligence team. (Of course not externally broadcasted as a demotion.)
> Microsoft likes to call the heads of various orgs "CEO"
Hah this is as funny as banks calling their junior employees VP of something something
A quick Google search tells me, the comment bove is talking about Mustafa Suleyman, the CEO of Microsoft AI.
Some more discussion on source: https://news.ycombinator.com/item?id=50406589
Real research at most big companies is dead. Unfortunately, what has replaced it is mostly venture-backed startups, where the timeline to a return is still expected to be quite short.
As a result, you end up with companies doing basic research under the guise of developing a product, announcing timelines which are never going to be met. Fusion energy and quantum computing startups are some of the most egregious examples of this.
There’s actually a bizarre irony.
Remember Bell labs, XEROX parc? Ask yourself why that existed and why research came out of them.
Monopoly.
So - many people don’t want to admit this but monopolies can be good from an R&D perspective, but without strong oversight by regulators - money will instead spent on R&D projects that specifically raise entry barriers.
I can show you many more examples of monopolies which did not produce PARC or Bell Labs than those which did. It's possible such research happened in spite of dominating the market, not because of it. Anyway, you cannot have both monopolies and strong regulators. These are incompatible situations. Monopolies occur because regulation doesn't exist or is captured.
> Anyway, you cannot have both monopolies and strong regulators...
> Monopolies occur because regulation doesn't exist or is captured.
IMO, the stronger the regulator, the stronger the incentive to capture it.
Unless by "strong" you mean "competent"?
I meant competent, but regulator strength in the way you're implying at its limit becomes monopolistic itself. So when I speak of regulation strength with respect to monopoly, I am speaking to the ability of a regulator to effectively both understand their duty and carry it out.
Agreed. I used to be an industrial researcher at one of the last companies that didn’t require their researchers to think of short-term productization goals. However, things changed at my former company, and I ended up leaving for a tenure-track position at a community college.
I don’t know what industrial research in CS is like in the era of agentic coding. On one hand I suppose building prototypes is faster than ever, but on the other hand I genuinely don’t know how well LLMs do with harder programming tasks. It’s one thing to have an agent work on a CRUD app, but I wonder how effective agents are at handling things like parallelism, distributed computing, load-balancing, and other topics that are harder to write than the average CRUD app. I am also aware of efforts such as Sakana AI’s automated scientific discovery research, and I wonder if many of the incremental, short-term research topics many companies pursue can be automated.
I’m a bit pessimistic about the state of research these days, from funding to publish-or-perish to the effects of generative AI (automation, AI slop publications, AI slop reviews). Then again, I’m a bit of a curmudgeon, and maybe things are better than I think.
I have been shorting most of the Quantum Computing startups for about a year. They have been printing. Missed out on Fusion! (?) These were one of few examples of the market leaving money on the table. Others examples have been Nikola. Usually the scammyiest companies don't IPO.
Lack of focus and long term planning at MS
This is the canary in the coal mine. The canary is dead. The AI Winter is coming.
This scaling approach is unsustainable, it is generating diminishing returns and ballooning costs. It is already blowing up.
Did you read the article? it literally showcases research published 2 weeks ago on how to make *better models*
I'm about 90% certain that there are numerous huge inefficiencies in these big transformer models. Inefficiencies so big that eliminating them would make it possible to run the equivalent of the latest frontier models on your phone.
But nobody wants to look for them too hard because centralized AI as a service is where all the money is. So everybody keeps building bigger rather than smarter.
If you value the freedom to pursue your own research direction and share your work freely, then choose academia. All the perks of working in an industry research role are provided in exchange for this freedom.
If your research aligns with the incentives of industry, you can get the best of both worlds in industry. This was the case for machine learning researchers in the post-AlexNet, pre-ChatGPT era. The good times can't last forever, though.
>If you value the freedom to pursue your own research direction and share your work freely, then choose academia
Perhaps it is more free, but my understanding has always been that professional academia still restricts freedom of research. You have to apply for funding, convince the board that your direction will be fruitul.
An archaeologist can't just dig wherever. They compete with their peers for a limitted funding pool. If the school's board is more interested in the bronze age than the paleolithic, you're probably not going to get your dino dig approved.
The same is true for CS, I assume. You can't just "i would like spend university resources developing brainfuck 2.0" because the other teams are more likely to get their funding over you getting yours.
Both corporations and university's have finite budgets, both with bottom lines and margins. The uni wants their research done in a popular field that will attract students because students generate income that funds research that attracts students that generate income that funds research that...
Academic research is free in the same sense as running your own business. You can try whatever you want if you can find the funding, but you still have to deal with people and regulations.
Universities are mostly just administrative organizations. They have little control over your research, they rarely fund it, and they are usually not even interested in it.
That’s not the issue here, it’s not like AIF was doing purely academic work.
They just think it makes sense to explore things, rather than just rinse and repeat with more cash.
The good news for John is that when 99.99% of the competitors are doing all the same, it’s when the rewards for looking around can be the biggest. Those are the real moments for disruption, and companies disregard them at their peril.
> If you value the freedom to pursue your own research direction and share your work freely, then choose academia.
The working environment in academia is often precarious :-( , so such people often have no alternatives to going into industry, even though they still deeply appreciate and are influenced by the values of academia.
I largely agree. For US though, academia has not been a good place lately.
Wait until you hear where academia gets its funding (hint: it is not no strings attached)
Hmm... So strange. It seems obvious(?) that both research and scale are needed, and that one should never have one starve the other. In fact research can influence the scale needed, and vice versa.
I guess when money is concerned people will always seek out the the most reliable and predictable path - at the risk of being potentially left behind if it turned out to the wrong (or at least suboptimal) path.
This seems to track with a recent blog post from MAI regarding their approach to pre-training their first Frontier level model which seems promising but does seem to have a focus on scaling.
https://microsoft.ai/news/frontier-pre-training-at-microsoft...
(In the context of Microsoft Research) Man, the history of how large corps fumble major wins should be studied.
I remember during my thesis how Microsoft Academic (a Google scholar/Scopus alternative) was ridiculously good at its job.
I'm talking, "input the key words and filter out the wrong fields and you've got 20 highly relevant papers that no other search term has found" level.
...Of course, they proceeded to close it down.
I can't even express in words how mad and disappointed I was and still am at Microsoft for this alone. It felt the closest to a genie or an oracle in being amazing. (The old Swype vs modern Gboard is probably a close example for those who'd used both.) It's just... crazy, frustrating, and crazy frustrating.
I used to work on the search index and query processing for Microsoft Academic. It was a great project, perhaps the most interesting thing I've ever worked on.
It's still available here (separate from the papers that Microsoft Academic collected and indexed): https://www.microsoft.com/en-us/research/project/knowledge-e...
I haven't worked on this since 2016 so I don't know if it still works.
John Langford was the creator of "VW" ("Vee Dub", or "Vowpal Wabbit"), an ML framework that was quite efficient back in the day ( ~ 20 years ago ).
https://vowpalwabbit.org/
Ironically now by collaborating with AI one person can absolutely move fast enough. I used to work in orgs like that and am now vastly more productive than I ever was back then. It could take me months to convince a peer of an idea that now only takes minutes to explain to AI.
Well I never thought I’d see Digg used as a primary source
This is certainly an interesting timeline in this section of the light cone
Executives chasing short-term returns. Google did 20 years of AI research, even creating custom chips, just so we could get to this point. Without long-term strategy Microsoft will perpetually be the also-ran. (which I guess is fine with shareholders?)
I'll play devil's advocate. Of course John Langford thinks what he and his group were doing held a lot of promise, and the work was fun and challenging and super important. But in the end you also need to live in the present and deliver. In the words of Jerry Maguire, "Show me the money". He had years to show Microsoft the money, and in the same time other people at other firms showed their bosses the money, big time.
this is not true. even with ideas with provable/strong scalability, there was still a lack of resources (especially GPUs) to validate such ideas at scale. ultimately, "showing the money" still requires some initial faith from those higher up who control the resources.
Part of the job of a researcher anywhere in the world is to convince the people who hold the purse that their idea is worth investing in.
Yes, leave it to the Chinese to advance the state of the art. America throws money at it, the rest brains. Because they don't have the HW.
Same happened to computer chess in the 80ies. Clever small SW was leading the tournaments, until Ken Thompson could persuade IBM to throw hardware at it, Big Blue. This win was shortsighted. HW didn't scale and nobody profited from it. SW did, esp. reinforcement learning.
New AI lab incoming?
Yeah, in India.
That's a name I haven't heard in a while
Is this the same group that did ResNet? Bad news
> scaling is straightforward
I remember IllyaS saying scaling doesnt work anymore so we are just RLcrazy. Is that not true?
Microsoft is a really shitty place to work right now. There are a hundred stories like this all over the company like this.
They've offered voluntary redundancy to a quite large proportion of the staff (a quarter?) so they'll also soon be losing anyone with employment prospects elsewhere.
I see a lot of news about waves of layoffs in large tech companies. Is MS the only shitty place now? I think all of them are, maybe except OpenAI & Anthropic, since they're building something new and are kind of on the "good" side right now.
Yep, from what I hear and see, it's like this everywhere. It actually started with the Twitter layoffs and has spread across the industry. At first, it seemed like a reaction to the ZIRP-era overhiring, but it's kept going long after correcting for it, and with AI has been cemented as the modus operandi for many companies.
From the outside, Microsoft seems to actually be one of the better places, and is making at least token efforts to reduce headcount in more amicable ways, as exemplified the buyout offer.
But in general, the strategy seems to be overload people until they burn out and quit, because hey, AI will replace them anyway.
Something tells me the only “good thing” in those two companies is the pay.
They look more like complete organizational shit shows.
Mr. S Altman just said almost all Startups will die within a couple years. This is sad.
Microsoft's now shown researchers their good work isn't safe.
but Mircosoft AI model is growing greatly and recently beat the nanobanana in generate video domain. I can't image the situation of them are bad as you explained.
Not what i expected, but good read nonetheless
It seems as if Microsoft is losing developers now.
The whole "let's go into AI" is not really compatible with oldschool software jobs. AI drains humans dry.
I mean we are already seeing the hype cycle coming to a close, with the stock market showing this, but key people leaving kind of means that Microsoft is losing its profile here. Add to this that many don't like Win11 and you see a huge paradigm shift here - and tons of associated problems.
"In practice, there are substantial issues because the philosophy of the ex-Inflection leadership of MAI is incompatible."
AI is killing people here - not necessarily literally (though we have to see how much economic damage AI has caused) but figuratively speaking. The counter-wave is coming here.
This was expected. Despite ill-advised views, AI was indeed NFTs in slow motion. We're finally returning to normalcy.
More here
https://archive.is/AaY5O
(Original https://hunch.net/?p=13764006 )
This should be the main link
Almost inclined to agree, but the headline of Langford's post needs some additional context.
Appreciate this, much needed context.
Ok let's put that link above!
Hugged to death like it’s 2008 and you’ve just lit your LAMP stack.
Thanks for the heads-up. I added an archive.is link to give it a breather.
[flagged]
[flagged]
[flagged]
[dead]
His complaints sound childish to me. If the research products were truly innovative and useful, he wouldn't need to force Microsoft to adopt them.
> If the research products were truly innovative and useful, he wouldn't need to force Microsoft to adopt them.
That's just patently not true. History is littered with products that were very innovative and useful (and often far superior to competing products), but were canned, bought by competetors, or shelved in other way - often because they threatened the profits being made by some entity (though that's not the only reason).
Yeah, I agree. What he describes reflects the exact challenges I've experienced literally everywhere. He is basically saying work/life is hard so he looking for a better playground. I do not know his contributions to AI so maybe he deserves to be better listened to but I find that never reflects corpo life.
You think incumbent megacorps are all quick at adopting new, useful technologies?
Some are, some aren't. Some try, but botch it.
Autogen is open source. I don't know the availability of other projects. Generally engineers in megacorps aren't shy of picking up the latest AI tools. I suppose Microsoft may have barriers around this though, curious to discover more details.
Edit: Fara is open weight as well. Seems like a great match for https://www.microsoft.com/en-us/windows-365/agents ?
[flagged]