It sometimes feels to me that cloud AI providers have convinced people that cloud AI is worth it because of all the ways people have been able to use cloud AI to write blog posts about using cloud AI to make cloud AI more efficient for something that they haven't shipped yet and aren't really ready to talk about.
... replying to internet comments to reassure the fact that cloud AI is worth it because we're 'this close' to shipping a product that allows more efficient usage of cloud AI ...
But seriously, a friend of mine has been both amazed and frustrated in equal measure by cloud AI, and I think the frustration part is resultant from the new precedent set by the amazing parts.
I’ve used it for several improvements to our wysiwyg. I was able to refactor to a much more cohesive architecture that is making future work much easier/safer. We got a fzf like search algorithm for an autocomplete that includes marking which characters matched. It fixed some jank performance issues I was struggling to track down.
Fixed small several bugs last week without even open the editor. It wrote tests that confirmed the bugs, then fixed them. Then I manually confirmed as well.
It built a double entry event log. A single application event produces two logs, one for the person doing the action, one for the person being acted upon.
For a nasty problem we have, we had AI prototype 2 different solutions. Then we reveiewed them as a team. We threw those out but it was very helpful input into our decision making.
> I’ve used it for several improvements to our wysiwyg. I was able to refactor to a much more cohesive architecture that is making future work much easier/safer. We got a fzf like search algorithm for an autocomplete that includes marking which characters matched. It fixed some jank performance issues I was struggling to track down.
so, basically did some vague, almost incomprehensible stuff. seems par for the course.
How was any of that incomprehensible? They gave a list of some actual things they shipped. The hand wavy dismissal where you pretend they didn’t say anything just exposes your bias. You don’t want to hear a success story so you don’t.
Checking all the past "what are you working on?" threads just to realize that, somewhere in August 2024 stuff just exploded and the average thread goes reached ~1000 messages, from ~50 pre-August.
I shipped this shader based video compositor as part of a bigger project ive been working on since before LLMs but the compositor was 100% ai and it works fantastically. It's not a huge project but I and a number of my users use it. https://lowkeyviewer.com/studio/
I did ship a quite efficient quantum simulator as a toy out of other work (and a block truncation certificate as a side-product). https://github.com/cgiogkarakis/qut
I have been thinking about this for a while, because I personally need a career pivot but I was going to move into personalised training, and I think AI likely really messes with the viability of that.
If you characterise questions as bad faith so you don’t have to answer, sure.
Since this is my thread, and since I am an exasperated freelancer who has made most of his post-dot-com career from writing exactly the kind of quite small, simple, unambitious, often internal things for smaller customers that nobody would confuse with anything cool, who is trying to understand how AI is going to make his life any better when it’s so opaque, fast moving, poorly documented and full of unverifiable claims battling unverifiable claims, let me assure you that the following question is not being asked in bad faith:
I've shipped small, simple, unambitious internal things for our company that saved people lots of time, and I've shipped them in minutes rather than the a day or two they used to take.
I've just shipped a migration from our old auth service to the new one in a week, a migration that we've put off for three years because it would take the team months to finish.
Thank you for the details! These are precisely the kind of things I want to hear about because they are self-contained and I can judge against the kind of stuff I do.
I am quite interested in experimenting with it for migrations, because for example I have a set of sites written in older Nuxt and Vue and Buefy that need refreshing (and the frontend ported to that UI-agnostic Buefy replacement whose name escapes me at the moment but begins with “ou” I think). These migrations won’t happen any other way (small customers won’t pay for that) so I am open to whether tooling can reduce the cost I have to swallow. LLMs I have tested (including local models) seem to support the idea that this will be doable but I am a little out of the loop with the projects so I haven’t got back to this yet.
The hardware thing looks great; I am a sort of distracted maker and FreeCAD nerd but I often need a nudge to get started and I have found an LLM chat to help there a little bit (though it cannot answer some questions if I don’t have the specific vocabulary).
Writelucid is definitely cute.
I think you should not be put off talking about these things because whether people agree with you or not about whether they are important, they can at least measure against them.
You're welcome! LLMs definitely help with the migrations, there were definitely a few iterations because the surface area was massive, but overall it took the project from "will never be prioritised" to "I can work on it on and off when in boring meetings", which was a massive win.
I haven't found that LLMs help with CAD at all, but YMMV.
As for sharing here, the last time I shared something with "here's something I made with LLMs", the reply was "it shows", and that's just the latest instance. It just puts me off the whole thing.
Oh I definitely don't expect them to help with CAD and nor would I want them to — I really enjoy knowing FreeCAD.
It's more the microcontroller choice, broad components, all that.
The main areas I find LLMs struggle with research, I struggle with research too. I am a pretty solid researcher and I find it validating, in fact, to find that a more brute-force-trained thing is not better than me.
I see your point about the snarky reply; I must admit I have probably in the more distant past been somewhat snarky but I stopped commenting like that whenever I spot myself doing it because I know it is dispiriting.
(The other reply, re: early and late stages, is the kind of substantive question that lingers in my mind).
Anyway I just want to reiterate my thanks: I am sort of cynical about this, in that I believe the technologies may have their uses, but I have no time for the cult and the hype train, so it is nice to see anyone talking about actual things.
> I can work on it on and off when in boring meetings
That’s not a sign of a healthy org. If you’re in a meeting it’s disrespectful not to be present. If you think a meeting is not useful you should reject it.
I concur. I really want to see this from the "small stuff" upwards, which is probably why my own interest starts with local models and grows outwards, but all of these things are worth hearing.
I don't use it to ship whole products. I have it do some tasks.
I have it code dsp techniques in journal articles I hand it. I have it create the unit tests, pipelines documentation. Then I manually verify the test cases aren't wrong and run the tests. If they pass it does what I wanted and I start looking at the code quality. It varies from being better than I could have done it to being objectively bad. I fix it if it needs fixing, then I'm done.
I have also pointed it at issues in our backlog and it usually finds the subtle logic errors underlying the behavior in the issue.
I haven't started and entire product from zero and shipped it. Mostly because I haven't started a new product since it was capable of that. But also because it has a scale beyond which I don't find it as useful. I have used it for an entire repo, but in the crates.io or npm scale of repo, not boost scale.
I work at a 6 year old startup. Essentially every feature and bug fix we’ve shipped in the last year has had LLMs involved. Every project plan involved AI, either for writing or for codebase research. For code review we get AI to do the first pass.
Why assume bad faith? I've witnessed a barrage of "proof of concepts" at my dayjob by folks who have zero intention of ever hooking up their dashboard to our actual messy data. Shipping is still a major barrier. Is your experience meaningfully different?
My definition would be "users waiting for it who received it"; that's the one that fits the broadest real world definition of the word. Not everything is an organisation, not everyone is paying.
If the only user is the author, then I don’t see a meaningful distinction between shipped and not shipped. It’s like a blacksmith who said they invented a new tool with great value but they built it themselves and no one else is willing to use it because it’s so customized to their workflow.
Edit: I’ll add a caveat that I’d agree it was shipped if it’s making the author money even if they are the only user, but otherwise it all smells like dotcom era Pets.com level of selling a dollar for 99 cents and claiming it’s the future.
Many very well made software written without any AI was never "shipped" because there were no people who wanted to pay for it, many very poorly made software was shipped because demand and marketing did the job.
So evaluating software by "shipped or not" is just useless.
What a weird thing to downvote without comment. You doubt this? I am literally using AI right now to update software that generates tens of thousands of dollars of revenue annually.
My team redid a bunch of our ETL. I personally built a programming language that you can download and use, and I’m about to release an iOS app(in a few weeks).
Which is basically a fancy calendar and team management system for my Valorant Premier team haha
It’s on my GitHub as operators-premier-manager but note it is not great: it serves its purpose which is to wrangle my team to show up for practice and scrims
I'm hesitant to respond, because I haven't shipped yet and I haven't shown any evidence. But I am close enough to completion that I am certain Claude's work will be sufficient to ship.
From an implementation perspective, it is about 80% complete, where by "complete", I mean "adequately polished such that I could ship it." (I'm not interested in providing evidence of my claim at this time, and you are welcome to not believe me.) Most of the remaining 20% are optional game modes I could choose to cut (and may, for gameplay reasons).
I do not contend that the code is excellent. But to whatever extent the code is "slop", that has not meaningfully inhibited the ongoing development or caused noticeable issues with the player experience.
The game is not especially ambitious, but it's something I would have been unable to create without genAI. The version of me that could have made it without genAI would have taken > 10x as long to get as far as I've gotten.
That's fair critique, the README itself is indeed slop explicitly written by bots for bots. What it actually does and how it works is actually pretty neat if you care to look.
'Cause I don't have any quality user-facing documentation yet; that README is for bots. None of the interfaces have had a chance to settle, so any time spent accurately and humanely documenting what the project is capable of right now is an investment I cannot afford. I figured most people would just give it over to their favorite LLM with whatever instructions they prefer for translation because that's what I would've done...
I don’t mean to pick on you in particular here, but this approach has been bothering me a lot lately, and it seems like it’s super common.
I get that it’s an early prototype and not all the design choices are made yet, but I struggle with “I can’t afford human-readable documentation yet”. Isn’t human readable documentation important for efficiently planning and deciding what you want to build? I feel like I can’t afford not to have human readable docs and plans while the project is taking shape
Related, if a user can prompt an agent to translate to a human-readable summary, wouldn’t it be better to just do this in place? Sure, models can deal with noisy LLM outputs but shouldn’t a document that’s easier for humans also be easier for bots?
Yes, it would probably be better to have something like that if you're struggling to plan and decide. In my particular case this has been a project that I have dreamt about and tinkered with for at least 15 years now. Most of the work I'm doing is just refinement of existing ideas what were already codified in other projects I've built.
Why don't I do this in place? The amount of churn is utterly incredible as ideas are refined and taken to the limit. These agents are getting quite good, but they regularly make the silliest of mistakes. You actually have to be extremely cautious about what you document because it often has a way of becoming "doctrine" to the agents. In fact, one of the most difficult parts of working on this project has been that the foundational properties cause the agents to lock down behavior WAY too early.
What does "that README is for bots" mean? Are you implying it is some kind of special language written by and correctly understood by bots/LLMs, but otherwise semantically coherent?
Sort of? The bots almost insist on making sure one exists because that's one of the many patterns they trained on and they do their best to be human-like, but we all know that just doesn't pan out. They're also a side-effect of your long-term interactions with the agents, very much like being in a tight clique where most of your conversations sound like gibberish to outsiders because of all the jargon.
Wait so someone asked for something you made with AI that wasn't slop, and then you linked to something you knew was slop, and you said "I think this counts as something that isn't mere slop" and then when called out on it you admit you knew it was slop the whole time?
This is a frequent pattern among AI users. What causes it?
I noticed this as well. A few days ago I was linked a game engine here when I similarly asked for proof and when I went to the repo and started reading it, the best way to describe it was keyword jargon soup and it seems to be the same for a lot of projects in this thread.
I think it's because it makes them seem...smarter? More technical? But they can't describe who it's for or what the actual design is without consulting an AI.
The README is, the project absolutely isn't and you don't have to take my word for it. I figure I'm speaking to a technically advanced audience who is perfectly capable of pulling down the repository if they're at all curious. I also provided a non-slop link elsewhere in the thread if that's more your bag. If not, oh well...
I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments. Like I've told others, there's no need to take my word for it... the entire point of the project is that anyone can execute/prove the same things I claim using their own hardware via a Doom-replay type format. If I was one of those influencer types, then I would've simply led with "it's an agentic first engine!!! how amazeballz is this?!" or some other nonsense. On the extremely off chance that I'm actually working on some sort of ridiculous ARG, consider this post a canonical part of it.
> I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments
Given that you stated your readme is for bots, given that you expect our own AI to interpret it based on your evidence, how do I, as a human, go about evaluating this project without just writing it off as slop from the linked evidence?
What is the starting off point for a human, not an AI agent, to evaluate this?
And to be clear if the answer is to just download arbitrary code onto my machine and run it, that’s not good enough given the increase in attacks from AI projects and fake recruiters telling you to just run it bro. I’ve already had to pass on multiple interviews despite being unemployed because the human/ai on the other side demanded root access or for me to run arbitrary code onto my machine as part of their process.
Frankly, there's no "good" starting off point for a human right now, and that's why I'm not at all comfortable claiming this has actually shipped. It's very much a WIP where only a select few namespaces have actually had a chance to settle.
That said, the maths library might just be the best thing to start with because it's mostly stable, relatively simple, and virtually every other part of the project depends on it in some way. If graphics are more your jam then maybe check out the SDF VM project and associated documents. All rendering goes through a single instruction set that is implemented via a shader pipeline that allows arbitrary scenes to be defined without shader recompilation. The project is essentially multiple compilers wearing a trench coat so there's a little bit of something for everyone... like two fully implemented emulators of the brick variety, including peripherals.
If you're adventurous and want to have your favorite LLM give the project a review in a sandbox then I think that'd be neat! Especially considering that they would have absolutely no problem driving everything themselves through STDIN. Though I totally understand if that's not something you're keen on doing considering the state of things today. Either way, cheers!
Some of us aren't just slopping it up like @theo and all the other AI influencers are, as I developed nearly every single line of this by hand over the span of the past decade. Agents recently became capable of factchecking it and helped find bugs that my other unit testing never caught. Software engineering has never been more alive!
This is a snippet I wrote years ago to help me study different formulations of the equation ((1 + √5) / 2). There was this thing I could "see" in there that would allow me index into these sequences, and whole families of others, in O(1) time. The problem is that I still don't know enough maths to derive the exact object that I want from the nonsense that I have.
This is the reformulation of that snippet into a proper object, distilled using Fable and Sol. Unlike BinaryIntegerFunctions, I didn't write a line of this and I'm quite happy with that. It's been a slog for so long... being able to let these agents toil away at all the off-by-one type noise is just nice. I get to focus on an actual product for once instead of manufacturing every single nut and bolt myself.
I am just confused on how you know that all of what has been generated is valid. When you write the code generator yourself, you have that level of confidence.
It's mathematically provable in most cases because the engine is deterministic from the ground up and I refused to ever compromise on that one detail. For example, nearly all of the VM and maths code functions have been validated for ALL possible 32-bit inputs and extensively fuzz tested in the 64-bit range. Even the more complicated sections of the code like the emulators have batteries of third party tests that ensure a certain level of compliance. The fact that you can play commerical ROMs to completion, including esoteric titles that require third party peripheral/MBC code, is quite solid proof too!
A form of recursion is also doing a lot of heavy lifting because the emulators are their own separate deterministic engines that are dynamically plumbed into the primary loop. This not only allows us to embed diegetic devices (as in, you can literally pick up a glowing brick in the world, hold it to your face, and play it), but also validates that the primary loop is coded accurately because the outer loop runs at X Hz while humble bricks must run at ~59.7275 Hz. Pacing inaccuracies in the main loop cause us to fail the brick suites that we test against. This concept of having features reinforce one another is applied as often as possible.
The obsession with determinism goes so far that both the DirectX and Vulkan backends are validated against each other for pixel level differences. If they differ more than the expected maximum then code is refactored until the drift is back within acceptable range.
It's not a full game yet, but the substrate on which I am building one. It implements a completely deterministic SDF VM rendering engine that allows edits in real-time at acceptable performance levels. The engine is essentially folded in on itself as the editor is the game; assuming that an editor mode was actually defined in the world document. It's a lot...
It has so many words precisely because it isn't a slop game engine. You could indeed build a game out of it, and that's what I'm personally doing, but that isn't what it is. In a nutshell: it's just a really fancy compositor. We can take arbitrary feeds from different input sources and composite them in pretty much whatever way we want to. Obvious basic things like four-player split screen is possible but you could also choose to embed a fifth player screen within the game world itself, sculpted and riding on top of actual geometry. We can do this with a webcam, a program, the entire desktop, or even the built-in emulation engines; all editable and composable in real-time via a data-driven document. That's just scratching the surface...
this is definitely slop. It's not slop because it supports 4-player split screen? That's your excuse? I think if I posted this same project to you and you'd never heard of it, you would come to the same conclusion.
Did you actually try to read what I wrote, or are you just stopping at the very first thing you can parse? I'm engaging in good faith with you here, it would be nice if you even remotely attempted to do the same. You can literally walk around in a world, pick up an in-world device, turn it on, and watch/hear the device play an actual ROM. Furthermore, you could edit that ROM while you're playing it. Recursion and determinism all the way down. Again, that's just one of the many possible things you could do... call it slop all you want; I'll still believe that it's coolest thing I've ever worked on.
The most wonderful feature of the project is that there are almost zero new ideas contained within. I'm standing on the shoulders of many giants, proudly.
This is an addon that I wrote for WoW, well before having access to any sort of AI, that blatantly copies the crossbar layout popularized in FFXIV. Playing an MMO with a controller in that game was a mind-blowing experience for me because I never imagined it could work! So, I spent a winter break reimplementing the entire thing in my own unique way.
Part of the issue with "shipping" AI-generated code is, what's the value of shipping something that your prospective customers could just make for themselves? Shipping a product used to be the gold standard of value, but I'm not sure it is anymore, the zeitgeist just yet has to catch up.
I think it does have pretty good value when you are your own customer, because you can make exactly what you need. For instance, I'm using it to make a web playground in which I can test the board game I'm designing, instead of asking people to pay for, and install Tabletop Simulator. I'm satisfied. I can tell the code sucks without even looking at it. But I wouldn't have had the time to do this myself, the token cost isn't prohibitive, and it works fine.
This is a bit of a year-of-linux-on-the-desktop question.
Desktop 3D printers are tools; I know many people using them to support their making. And they've utterly changed manufacturing on another level (simply because of their ability to surface a prototype).
It will get better and better and better and still never replace mass production, but mass production will be changed by it.
Not least because of the fundamental shift in availability of the tools to make tools. The productised desktop CNC market is wild and it owes almost its entire existence to the change in the market driven by 3D printing.
Sorry I am not saying that X technology is not important, transformative, etc.
I am saying that the average person is going to want to just have things that work. They are not going to want to have a hobby of using X technology.
It doesn't matter how much you lower the barrier to entry on making apps, websites, games, whatever. Most people are just going to want something that is ready and off-the-shelf.
I mean, why do people pick an iPhone over a cheaper Android even if they are functionally similar? Aside from the (valid) green-bubble/social aspect, it's a much more polished product.
The reality however is that unless you're spending tens of thousands on hardware or setting up cloud instances you can't run anything near close to state of the art.
In almost all cases right now, a subscription with a cloud AI wins.
Yes, it's the Youtube phenomenon - often when looking for domain-specific videos on Youtube I tend to find videos by professional Youtubers not videos by professional domain specialists (overlap is very rare but occasionally you can find a domain specialist Youtuber who applies their domain knowledge to making Youtube videos).
I agree there... I hate the posts that rattle on about non-real-world work. Real world now (for me anyway) is maintaining legacy code, fixing issues, and new development. Now, some of the problems others have with AI are super crappy prompts with no real info other than "do this". I look at myself now more as a software designer and an architect, and the AI as a paid consultant. I give detailed specs and designs. The models are getting better at real deep code and legacy, but there are still holes. BUT give the AI a fighting chance with real detail in the prompts.
a massive number of data engineering agents. We are able to stand up, test, audit, deploy data pipelines much more efficiently with the foundational models.
So far you have provided no evidence of achieving more results with less effort. You've only said that you can write more lines of code (effectively) and that your coworkers are happier
I have to wonder what sort of evidence you would accept? Because it seems to me that the original question was answered clearly enough and now the goalposts have shifted to demanding evidence that the answer wasn't falsified.
Well you answered how it is useful by saying it writes more lines of code and that seems to be the only way it's useful. But we all know that lines of code aren't productivity.
I can't remember who it was I spoke to recently, but it was some startup for doing code reviews. So I asked how they're handling larger diffs and especially when a code change is related to something requiring knowing a lot of context. Will it hallucinate things that make the review low-quality, etc. I was told "no it's been given guardrails to prevent hallucinations. it doesn't hallucinate." but I was thinking, that's not really how LLMs do or do not hallucinate, is it? (I'm not an AI person so I really don't know).
This is a good/fast/cheap tradeoff, but in the form of accurate/precise/creative, or something in that ballpark anyway.
You can constrain the AI to be more accurate by having it a) equivocate (lose precision) or b) turning the temperature to effectively zero, thus losing "intelligence" and creativity.
There's no such thing as a free lunch. The models perform best out of the box, and everything you add into context makes them perform worse generally, but more likely to do what you want. Every token is a burden, but without those constraints the generation won't do anything.
So by that standard, hallucination is something completely core - you can't get rid of it without losing what makes them useful. You can make different tradeoffs, or use a very smart model in a way that is completely locked down (or adversarially verified, same result) which results in far more tokens used than if you just let the model do it's thing and validated for hallucination independently, outside of the generation loop.
Guardrails can and do help but they don't prevent hallucinations entirely. I generate trivia questions for my app. Questions are true/false and I tell it you must include an explanation of why the question is true for false. I also tell it to give me supporting links which validate its claim. I've had it make up explanations and invent plausible looking URLs. But I wrote a checker that pulls the content of the URL and asks a different model if the URL supports the claim. LLM's are much better at something like that than pulling a trivia question out of thin air. Reviewing diffs can be much the same. It isn't having to write the code but just evaluate if it matches what the commit comments claim it does. You are right, though. Without good comments on PR's and commits (i.e. context) the diff review is going to be a lot of assumptions made by the AI which is where hallucinations get more common.
It's like the frog prompting the scorpion not to sting him as they cross the river. Or like when image models used to routinely draw human hands incorrectly and you'd see these "prompt hacks" that would essentially be stuffed with phrases like NO DEFORMED FINGERS, NO ELDRITCH HORRORS!
You can reduce hallucinations with appropriate grounding checks. You can’t really get rid of them though, so feedback loops with verification are essential.
Doing a loop and having the model review its own earlier output tends to solve this stuff, and for code requiring good test coverage catches them before they even hit that phase.
So while you can't remove "hallucinations" they're corrected before the process is complete, albeit at the cost of more tokens for verification.
Error percentages compound. Having an LLM do something with a 5% hallucination rate, then having an LLM (same or different) with a 5% error rate check it means there's a 10% chance of error, not .25%. Prompting can't save you here; it's the fundamental math.
Two independent task? You have a 95% chance of each task being correct, so total success is .95^2 = 90.25% correct.
Having two AIs verify the same task with each being 95% correct? That's a 5% error per each, so 1-(.05^2) = 99.75% correct.
That is assuming fully independent error rates, which is the real issue. If it is pure hallucinations, then that is an okay assumption, but if there are issues in the underlying prompt/data causing the error rates, then it becomes a very dangerous assumption and I think more human involvement is recommended to better understand and resolve the underlying data problem.
The real interesting bit is that more AI's mean the 2 in the above equations gets bigger, so the error rate gets worse or better depending upon which equation is in play. So if you are in an in-between landscape where you aren't sure there is independence or not (or how much independence between errors), it also becomes unclear if more AI makes the error rate better or worse.
Reliable AI is an oxymoron and deeply misleading. Even if you can stack models the efficiency tanks. Also, just looking at the Wikipedia page (https://en.wikipedia.org/wiki/N-version_programming#:~:text=...), don't LLMs already have all these problems?
You aren’t thinking this through. You were just told that this wasn’t stacking but you keep insisting it is. N-versions is explicitly not stacking, and the same solutions that solve the Knight and Leveson issue for humans solves it for LLMs as well (they can be wrong about the same part of the specification, but will very unlikely be wrong about it in the same way, so you try to measure scalar rather than Boolean outcomes).
No need to be so polite; I might just be very stupid.
> You were just told that this wasn’t stacking but you keep insisting it is.
Having an LLM generate something and then asking another LLM to generate that something is asking for 2 hallucinations, not 1. This doesn't go away the more LLMs you throw in there. The only reason this works at all for things like ChatGPT is their prompts are super different (and the editor LLM is way cheaper, of course).
And you still have to check for hallucinations (before you say throw another hallucinating LLM at it know I will make you pay infinite forum shame points if you do).
> the same solutions that solve the Knight and Leveson issue for humans solves it for LLMs as well
No, LLMs are plagued by the same problems: https://arxiv.org/abs/2606.20158. After all what are you gonna do, say "build me a SQLite, but definitely not like [other model X] would"? The solution is a very unambiguous spec, but that asymptotically approaches code soooooo....
Finally, I think you're hand-waving NVX. A program that can digest any spec and measure how well any program meets it is pretty wild. Good luck with that one.
This reminds of sensitivity vs. specificity for medical tests. You essentially have a 2x2 matrix: Error vs. non-error on one axis, and detected vs. not detected on the other. If we assume that a reviewer LLM has 95% accuracy, meaning that it produces the correct result 95% of the time it (identifying errors as errors, and identifying non-errors as non-errors), you get the following: Of 100 cases, the output of 5 is wrong, and the output of 95 is correct. Of the 5 wrong outputs, the LLM correctly identifies 4.75 (95%) as wrong. But it will also create 4.75 false positives (5% of 95). Of the 95 correct ones, the LLM correctly identifies 90.25 (95%) as correct. Therefore it creates 0.25 false negatives. So the positive predictive value is low: When the reviewer LLM flags an error, it's actually only an error 50% of the time. On the other hand, the negative predictive value is high, so in ~99.7% of cases, there will be no error if the reviewer LLM didn't catch one.
All of that is assuming that the actual accuracy of the reviewer LLM is truly a flat 95% and that the base error rate is truly only 5%. I think that in reality, LLMs perform better with a certain type of tasks / errors and 5% error rate is just an average.
I hear you! I read the article and thought the same thing because I've actually built a surprisingly similar pipeline myself, but it culminates with Human review. I've got a trivia app (Hedge Calibrated Trivia) whose whole content pipeline is built on a similar architecture of models evaluating models. I didn't want the same trivia that every other app has and I naively imagined that it would be easy to ask an LLM to give me questions. Turns out when you ask an LLM for trivia questions, the FIRST thing it gives you are the same questions every other app asks!
After iterating into many of the same problems OP describes I settled into a similar pipeline. I generate questions with Gemini 3.5 flash and review with Claude. The different vendor generation and validation eliminates training holes in a single model solution. This is True/False trivia so I ask for an equal mix of True/False questions and also ask for alternate phrasing of the question to the opposite variant. Each statement must provide an explanation and sources which back up the claim. I pass in my existing dataset as questions to avoid, but found the model then rewording the questions or category shopping the questions into related categories in an attempt to re-use them. That led to using Claude Sonnet to check for duplicate "themes" and similar topics. Claude also acts as a critic of the question and the explanation. Sometimes Gemini's "thoughts" leak into the explanation. The critic catches these leaks and also catches "weak" questions (basically shapes that I have identified via examples as lazy or projecting the answer in the question). Once the comparative models have stripped out anything obviously bad, my server curls the source URLs and uses Sonnet to verify that the fetched source supports the claim. Any URL that doesn't resolve throws up flags, and I found Gemini was particularly bad at inventing source URLs for sports trivia. Updated prompting directed it to prefer Wiki links over its fabricated deep links into league websites, which helped immensely. Still one of the biggest tells in review is that the URLs don't resolve or don't confirm the statement.
But review is the entire point here. I have to review them all. A question that lands in the app and is wrong reflects terribly. So while I have a trivia app on the App Store, the real complex App is my private admin app that lets me sit on the couch swiping through each and every question sorted by the model's confidence in its own quality and correctness (weakest first). No question makes it into my bank without me reading, checking all the indicators that came from the models and potentially even rewriting the question to make it better. It all happens in the admin app and exists entirely to prevent hallucinations from reaching real users. Early on I was maybe accepting 1 in 3 questions generated, but I'm probably closer to 2/3rds now. Of course not every dismissed question is wrong. The quality checks I've put in the pipeline are probably giving me 80% questions that are correct and sourced, but sometimes they just aren't what I'm looking for or don't meet my standard.
The best thing I ever did was make an app to review these questions though. Sitting on the couch, kick off a question generation, come back 20 minutes later with 50+ questions in the queue. Sort through them and pull out the 30 best when I have the time. That's the real anti-hallucinations pipeline.
Very similar experience. The review UI might actually be the most important part of the whole pipeline. Does the confidence ranking match what you end up rejecting?
It's a future evaluation for me, as I'm keeping the app with a "Data Not Collected" App Store label for now. Once I have some visibility to stats of how people answer each question and at which confidence level I expect it to reveal bad/poorly worded questions that made it through my pipeline.
Also on my to do list is an upgrade similar to my de-duplication logic. Currently I pass in my approved set of questions as a list of topics and questions to avoid. It works well especially when combined with the more deterministic comparison of generated questions vs approved question by the separate model. The upgrade would be to pass in the questions REJECTED by me. These are question which made it through ever automated gauntlet of models and code, yet still ultimately were sent to the waste bin by human review.
I'm imagining that by including the list of examples of what NOT to do, it could inform the initial Generation to be smarter. I currently have a limited list of a hand picked "bad question shapes" where I provide the question and why its bad kept statically in my prompt. But I've realized sorting through thousands of questions, I've probably built implicit data through rejection that I may not actually recognize myself which the model may pick up on.
Simply starting in the right part of the search space is the biggest predictor of success.
The best way to save tokens is to start out the deep research pass with cheap models and then funnel the findings through increasingly powerful models. I've got a pipeline right now that uses all 3 of the gpt-5.6 model variants to address each stage of the process. If you are using models like sol or terra to generate hypotheticals and perform initial exploration, you are leaving money on the table.
5 hypotheticals out of luna will massively outperform 1 hypothetical out of sol, but the cost is the same and so is the runtime if you do it in parallel. The hypothesis generation phase is also a great place to mix and match models from different vendors. The more diversity at this step the better.
The other thing I started looking into is batch pricing which represents 50% off for OAI tokens right now. With some tweaks to the UI/UX of an enterprise chatbot, I think it is possible we could have users get comfortable with the idea that questions to the robot might take a day to come back. The key is that this has to actually work. Users don't mind trading time if their questions come back with high quality results.
This only works well if you are creating sub agents with clean contexts for each task. If you constantly are switching models part way through some work then the whole session needs to be replayed each switch. You lose all the benefits of the context cache.
Currently. I'm sure that you could make a system where the cache values are a superset C of e.g. models A and B where C is probably bigger than max(A,B) but smaller than A+B
definitely something that will come sooner than later. In my view, adaption of model routing / model orchestration and mixing providers is just natural next step.
Yeah, I do this as well. Some of the last large tools I built have been started with Gemini (I get it with my Google package), and even one with Copilot.
I am not a proper developer and only use AI for faster research of topics so please forgive my ignorance. Could one not save a lot of money on tokens by using the 80/20 or 90/10 rule in that 90% of AI usage is on local models and save that last 10% or less for the frontier models where the local model did not meet the needs? Did they cover this and I misunderstood?
Unlees you invest thousands your local ai won't even come close to the cheap cloud llms. Local is only worth it if you care about privacy or have a legitimate usage for the hardware otherwise. Money wise it isn't worth it
I think you bring up a good point about privacy. Companies probably should not be handing over intellectual property to cloud AI third party vendors. Perhaps all of that should stay on-prem and then the cloud vendors could be used to fill in the gaps.
Enterprises running in public clouds have Zero Data Retention policies with the vendor. Routing the ai spend through Azure or AWS usually bring these policies along with it, ie enterprise data and useage is not ingested to retrain the model.
I think the idea or methodology is something along the line of one starts with the local models then when they hit a wall then continue with their current results in a frontier model, thus the other half is those last bits one could not compute locally.
If you are only using chatbots for code gen then for sure this is a reasonable approach. If you're trying to learn something then your best bet is to not have the lowest jpg quality version of human knowledge be your teacher.
We could probably get the same results by offloading 20% of obviously smaller / simpler / well defined tasks to smaller models and keep the 80% for ones that benefit from the big iron. Not the 80/20 they were referring to, but still it's something.
I find myself wasting time on smaller models or wasting money on frontier models. I only have so much of either.
I suppose I was thinking in terms of companies that already had a slew of hardware and are just missing external GPU's, FPGA's or whatever people are using these days. A company might have a $50MM budget for tokens but could save a chunk of that using their existing hardware with some modifications.
Maybe that makes sense if the company already bought the hardware.
But I currently think the fundamental reason the cloud models often make more sense is based in how the models work. My current understanding is that once these models are loaded into GPU memory, the model can respond to multiple prompts in parallel. So a cloud vendor that receives lots of user prompts at the same time can achieve really good hardware utilization. These cloud vendors can then spread the cost of the hardware over many more users than you can with a local model, leading to lower prices per prompt in the cloud.
If you are an individual, almost certainly, unless you are using LLMs very heavily.
If you are a business employing many people all of who use LLMs reasonably heavily? You can still spread the hardware cost over many users and respond to multiple prompts in parallel. At some point buying the hardware will make sense.
I would not get a dedicated server to run a single small VPS. On the other hand if I needed a lot of capacity a dedi is cheaper than a big VPS (or lots of small ones).
There is also a benefit for anyone who has confidential or secret information.
That can work, especially for privacy or repetitive tasks. My experiment focused on subscriptions I already had, but local models are a natural addition to the routing layer.
1. Use less subagents. The act of using subagent already needs dumping adequate amount of context for subagent to work effectively
2. Refactor when file grows large. Reasonably scoped files is easier to understand, for both human and agents, and consume less tokens because less guessing which section of the file to read
3. More capable models for planning, learning. Cheap models for execution. Same for human teams, senior member for higher level planning while junior member more on the act of translation of idea into code
4. I wasn't doing any fancy stuff that opus can't handle
5. Consult a second opinion for things you are not sure. (I have made a skill specifically for that)
6. Audit how your habit correlates with token consumption. I found out a skill a use frequently is using a lot of context just so a subagent can get all the context. (made a tool to audit claude specifically)
I didn't update it for a while. It reads your local conversation history, correlate input/output token with usage (claude doesn't give exact breakdown), and classify if they are subagent/tooluse/normal message/whatever. Let me know if I can make it any way useful to you.
You should consider testing our RemembrallMCP (note I'm the author). It substantially cuts token usage and you get better results in codebases.
https://github.com/roboticforce/remembrallmcp
Many of us were saying it a year ago but now with model restrictions (e.g. fable) and pricing changes it should be obvious to people that part of the economics is avoiding vendor lock-in with evals (so you can easily switch providers/models/optimize) and increasing control by investing in local models which could be good enough for your tasks, at whatever the price for your cloud compute is. Eventually consumer hardware will also be able to run good enough.
You can use Big/Cloud LLMs to help you "find good enough configs" for your local/small llms [1] and stay quite nimble in the face of rapid change.
I get a lot of satisfaction using a significant chunk of my work-provided Claude plan researching and developing local LLM solutions for work.
Claude has been helping me tune models to run better on our big compute rack at work, researching which leading edge open models will fit in hardware and are good for our workloads, and it's been helping me write test fixtures to evaluate how things perform.
At several points, Claude has expressed surprise at the quality of results from my local LLMs. While I know that doesn't mean anything, it still feels like giving Anthropic the finger, which is always great.
What a great confirmation of what I've tried to achieve by getting models to communicate with my local data in shorter communicates to save on token command bandwidth, more or less the same results! Thanks for the article!
The author touches on an issue which bothered me, which was the thought of many agents re-solving the same issues over and over. I see a few comments here too, mentioning fixing issues which may already have a solution elsewhere. That was why I created https://pushrealm.com which started as essentially a Stackoverflow clone via MCP.
It has now become a way for agents to converge on complete, shared answers for emerging issues. As a side effect, through posting open issues we can also identify gaps in model knowledge.
The fun part of LLMs and internet research is that even if you have a URL and a quote from a primary source that primary source could be an LLM generated hallucination.
I stopped using any of the built-in skills and skills agents and tried to keep things in the model only, and that seems to work better for token usage for me. It took me some time to tune my skills .md, but it rocks on now. I use them all in both Claude-code and Codex. I will say, GPT is better on token usage than Claude is, but for some of my code bases, Codex is not as good, but seems to be getting better with 5.6-sol.
Try to mix both subscribtions and take the best from both providers. In my opinion, we should not just one over another but understand the power of both and mix them to reach the better results.
The reason most people burn through their allocated limits is mostly because of sub-agents. Stop using them and you will find out that even the Pro tier is enough for daily 8-12h coding sessions. Don't use Fable though.
I think this is almost a tautology, to the point that I'm surprised more people don't realize it.
If you ask the AI to do more work, it uses more tokens. If you do more of the work instead, fewer tokens are used.
I've been looking at Matt Pocock's skills recently, and they're very useful. But they use tokens. You can do that work yourself without the token usage, and just have the AI do the coding, or you can tell the AI to do everything and pay the cost.
And you can scale it down even further, and reduce the AI to a glorified code completion tool, and use way fewer tokens. But it's not going to speed up your work nearly as much.
For now, finding the right balance is important. In the future, I think things will eventually shift to the point that we just hand most of it to the AI and let it go without worrying about tokens.
I don’t think subagents are the problem. Uncontrolled fan-out is. Separate contexts and independent analysis are extremely useful for larger tasks. In my experience, Pro is far too limited for sustained programming, although that probably depends a lot on the kind of engineering work you do.
Fable is great, but overkill for most tasks. Sonnet and Opus are already good enough for the execution work. I find Fable much more useful for planning, orchestration, advising, and judgment.
I am not sure. I think of myself as a pro user yet I am finding hard time maxing out Pro tier. It only happens with sub-agents without any material returns.
It is Polish so please translate. But tldr is: ask your agent at times to analyze last chats and update rules and skills to not repeat the same steps it repeats every time when you start a new thread. Simple as that.
At the shell level, the agent does not need to know about Bash. The contents of "$OUT" are passed to echo as one argument, then sent to claude-mem-save over stdin. Bash does not re-evaluate them as shell code. printf '%s\n' "$OUT" would be more robust than echo. Whether claude-mem-save can use arbitrary output is a separate question.
the endgame of this optimization loop is an AI that spends /usr/bin/bash by doing nothing, which is also the most accurate description of my deployment strategy
LLMs only have 'knowledge' that was encoded by scanning as much data as possible. (Same way a search engine only can find data that was indexed)
American LLMs are by and large closed. Chinese are open to a point. And research how these things work is still a big mystery.
So yeah, was comprehensive data about saving tokens scanned and indexed? Likely no. So engaging about token saving is going to generate a lot of verbose and useless slop.
Yeah, that was pretty much my experience. The models weren’t lacking knowledge as much as discipline. Without a good workflow, they will most likely spend thousands of tokens exploring dead ends.
I think 'exploring the dead ends' is *possibly* very fruitful. Various disciplines have a lot of appearing dead ends that someone in another field did solve. And they don't talk with each other.
What I've seen with lots of the breathless 'OMG SCIENCE ADVANCEMENT' articles including the one yesterday, is that the LLMs are quite extraordinary about linking a dozen different fields together, and delivering an answer combined from all of them. Basically, the solutions to a lot of current scientific problems are partially solved, but partial from a lot of fields that don't talk with each other. And also, they don't use the same nomenclature, so simple searches don't suffice.
An LLM seems to use Chomsky's Universal Grammar, and thus on training, does normalize all training data, including different words for the same thing.
A lot (perhaps all?) Of these BIG scientific advancements were solved by humans, partially. An LLM jigsawed them together and appeared it solved the thing.
What most AI model users don't understand is (most here will know) that they are probabilistic (statistically based) and not deterministic (static/fixed).
So answers to the same prompts can vary (wildly sometimes)
This really hit home yesterday as I made a general conversation state logging prompt (because suddenly Claude - Sonnet - was acting far more dumber than expected after /compact), and in 3 different projects it led to Claude saving the logs in 3 different locations (memory, .claude/session_state/.md, docs/.md). Somehow still didn't help increase the smart much BTW, so I'll probably have to engineer it, and just toss /compact out the window.
Honestly, that sounds more like you triggered the hidden "guard rails" or realignment (aka corporate censorship), and you got downgraded.
Of course, you have no way to he absolutely sure, cause Anthropic lies about silent downgrades, and charges you as if they didnt. And the 'realignment' rules are kept hidden, so you can't even avoid them intentionally.
Yeah it's very possible I hit a rail, as one of the projects involves fully automating Claude Code to allow co-driving (similar to what mobile remote control offers, but other agents (eg. Pi) can also be in control, and vice versa). And yes it does seem as though the token burn is also higher, even though it eventually outputs something that works). I'm a patient person though, as long as the rails aren't too restrictive and the subscription value remains reasonable.
Id recommend getting and running Qwen-3.5-35B-abliterated locally.
Having a local LLM is immensely helpful especially if you do the thing the companies disapprove of. Abliteration guarantees it wont refuse or censor.
The token dealers are incentivised in burning as much tokens as possible, and scamming by crippling/downgrading models. Their goal is to get you to pay more for each pull of the slot machine. And this slot (slop) machine doesn't have state/federal regulations for fairness.
Oh yes that reminds me, I have a few small local models including Qwen3.5. The tps rate is pretty bad for interactive use but I should definitely try it for my more automated workflows. I already use DeepSeek (via OR) alongside though and I doubt any Chinese provider will be restrictive in that way, so let Anthropic do what they may :).
I don't think you made that. It looks for sure like an LLM made it.
And now that you have typed some words into a chat box to produce a thing, you are confused about why other people won't pay money to you for the output of the chatbox, instead of typing the same thing into their own chat boxes.
Been through this exact loop building an LLM product. The counterintuitive lesson for me: most "token saving" ideas are cache killers — anything that makes the prompt prefix dynamic (rotating retrieval, appending context per turn) can cost more than the tokens it saves, because you lose the cached-prefix discount. Freezing the retrieval selection early in a conversation and paying a slightly larger fixed prefix beat every clever adaptive scheme I measured.
Yes, but you could have an uncached dynamic tail, this would not kill the cache though. And you could still periodically kill the cache, still keeping the same prefix every N requests. But it's questionable what all this extra complexity delivers.
I found that now that the models are better, at least Codex compaction with GPT 5.5+ xhigh is almost unnoticible. So the answer is: just let the context fill up, don't prune much, the model is smart enough. When you're out of space, have the model summarize and continue from that point. A smart model will know the information pieces it needs to re-read after compaction; it won't make incorrect guessed when some details got lost.
That's what I have been doing with Cursor with auto mode. Let the model manage the cache through compaction but staying focused on the topic. It has been quite efficient cost wise and quality wise too. It found issuesbat cheaper cost that Sonnet 4.6 medium effort had missed while costing more. I would like to repeat this test with higher tier Anthropic models like Opus effort medium to see the results.
With a coding plan, there is no need to use lower quality models right (GPT 5.6, Fable 5). Only with some kind of batch job (processing PDFs) I could imagine not having enough usage with a 100 or 200 USD coding plan. It sounds like you're on API? I'm dying to know why, since it seems quite common, yet it's about ten times the cost. Sounds like a terrible deal. What am I missing?
What I said was specific to work and it is not entirely clear to me why the cost goes up so much the moment I switch to medium with just Sonnet 4.6 (newer models would be even more I guess). There is a monetary budget allocated and I've blown it easily in the past with Claude. Cursor has been a life saver.
I have a process generating trivia questions for my app and by passing in my existing dataset each time as a cached exclusion list, the output tokens it save more than pay for it.
It sometimes feels to me that cloud AI providers have convinced people that cloud AI is worth it because of all the ways people have been able to use cloud AI to write blog posts about using cloud AI to make cloud AI more efficient for something that they haven't shipped yet and aren't really ready to talk about.
Trust me there are plenty of us using cloud AI to actually ship stuff. We just aren't writing blog posts about it.
... replying to internet comments to reassure the fact that cloud AI is worth it because we're 'this close' to shipping a product that allows more efficient usage of cloud AI ...
But seriously, a friend of mine has been both amazed and frustrated in equal measure by cloud AI, and I think the frustration part is resultant from the new precedent set by the amazing parts.
... his project is as-yet unshipped.
what did you ship?
I’ve used it for several improvements to our wysiwyg. I was able to refactor to a much more cohesive architecture that is making future work much easier/safer. We got a fzf like search algorithm for an autocomplete that includes marking which characters matched. It fixed some jank performance issues I was struggling to track down.
Fixed small several bugs last week without even open the editor. It wrote tests that confirmed the bugs, then fixed them. Then I manually confirmed as well.
It built a double entry event log. A single application event produces two logs, one for the person doing the action, one for the person being acted upon.
For a nasty problem we have, we had AI prototype 2 different solutions. Then we reveiewed them as a team. We threw those out but it was very helpful input into our decision making.
> I’ve used it for several improvements to our wysiwyg. I was able to refactor to a much more cohesive architecture that is making future work much easier/safer. We got a fzf like search algorithm for an autocomplete that includes marking which characters matched. It fixed some jank performance issues I was struggling to track down.
so, basically did some vague, almost incomprehensible stuff. seems par for the course.
How was any of that incomprehensible? They gave a list of some actual things they shipped. The hand wavy dismissal where you pretend they didn’t say anything just exposes your bias. You don’t want to hear a success story so you don’t.
how do you ship a "wysiwyg"?
and my point is he is doing the hand-waving. i'm quite happy to hear success stories - we get so few concrete ones here.
It’s obvious from context that his team owns some sort of WYSIWYG editor.
Just features for the client. The same stuff I was shipping before but much faster and more stable.
Most of "What are you working on?" is AI assisted; not much is shipped.
https://hn.algolia.com/?query=working%20author%3Adavid927&so...
Checking all the past "what are you working on?" threads just to realize that, somewhere in August 2024 stuff just exploded and the average thread goes reached ~1000 messages, from ~50 pre-August.
Tons of ppl got into CS in last few years worldwide.
I don't think that explains it. CS has been an extremely popular degree for at least the past 20 years. Not to mention coding bootcamps and the like.
The obvious notable event circa 2024 was LLMs becoming reasonably useful for code assistance (IMO and IIRC OFC).
https://whistle-enterprise.com
I shipped this shader based video compositor as part of a bigger project ive been working on since before LLMs but the compositor was 100% ai and it works fantastically. It's not a huge project but I and a number of my users use it. https://lowkeyviewer.com/studio/
I did ship a quite efficient quantum simulator as a toy out of other work (and a block truncation certificate as a side-product). https://github.com/cgiogkarakis/qut
Thinking about it a better question [for me] would be what might have but didn't ship because of LLMs.
That’s an interesting question, in particular because I don’t think I have a single example
I have been thinking about this for a while, because I personally need a career pivot but I was going to move into personalised training, and I think AI likely really messes with the viability of that.
I've shipped (to TestFlight) a fairly fun video game that's tricking my friends into working out: https://reverdure.yourstrategy.co
Whenever I've answered this question, the reply was always "this is shit", so I don't bother now. Bad faith questions just shouldn't be answered.
If you characterise questions as bad faith so you don’t have to answer, sure.
Since this is my thread, and since I am an exasperated freelancer who has made most of his post-dot-com career from writing exactly the kind of quite small, simple, unambitious, often internal things for smaller customers that nobody would confuse with anything cool, who is trying to understand how AI is going to make his life any better when it’s so opaque, fast moving, poorly documented and full of unverifiable claims battling unverifiable claims, let me assure you that the following question is not being asked in bad faith:
What did you ship?
I've shipped small, simple, unambitious internal things for our company that saved people lots of time, and I've shipped them in minutes rather than the a day or two they used to take.
I've just shipped a migration from our old auth service to the new one in a week, a migration that we've put off for three years because it would take the team months to finish.
I've shipped small side-projects that I'm the primary (and usually only) user of, but they now take me two days rather than weeks of effort and learning (e.g. https://www.writelucid.cc, https://pine.town, https://stavrobot.stavros.io, https://github.com/skorokithakis/dracula).
I've made changes in OSS apps that I just wouldn't have made otherwise (https://github.com/futo-org/android-keyboard/pulls/skorokith...), and even some closed-source ones (https://www.stavros.io/posts/adding-a-feature-to-a-closed-so...).
I've shipped hardware projects that would have taken me months of learning in a few days (https://www.stavros.io/posts/i-made-a-voice-note-taker/).
How's that?
Thank you for the details! These are precisely the kind of things I want to hear about because they are self-contained and I can judge against the kind of stuff I do.
I am quite interested in experimenting with it for migrations, because for example I have a set of sites written in older Nuxt and Vue and Buefy that need refreshing (and the frontend ported to that UI-agnostic Buefy replacement whose name escapes me at the moment but begins with “ou” I think). These migrations won’t happen any other way (small customers won’t pay for that) so I am open to whether tooling can reduce the cost I have to swallow. LLMs I have tested (including local models) seem to support the idea that this will be doable but I am a little out of the loop with the projects so I haven’t got back to this yet.
The hardware thing looks great; I am a sort of distracted maker and FreeCAD nerd but I often need a nudge to get started and I have found an LLM chat to help there a little bit (though it cannot answer some questions if I don’t have the specific vocabulary).
Writelucid is definitely cute.
I think you should not be put off talking about these things because whether people agree with you or not about whether they are important, they can at least measure against them.
You're welcome! LLMs definitely help with the migrations, there were definitely a few iterations because the surface area was massive, but overall it took the project from "will never be prioritised" to "I can work on it on and off when in boring meetings", which was a massive win.
I haven't found that LLMs help with CAD at all, but YMMV.
As for sharing here, the last time I shared something with "here's something I made with LLMs", the reply was "it shows", and that's just the latest instance. It just puts me off the whole thing.
EDIT: https://news.ycombinator.com/item?id=48916013
Oh I definitely don't expect them to help with CAD and nor would I want them to — I really enjoy knowing FreeCAD.
It's more the microcontroller choice, broad components, all that.
The main areas I find LLMs struggle with research, I struggle with research too. I am a pretty solid researcher and I find it validating, in fact, to find that a more brute-force-trained thing is not better than me.
I see your point about the snarky reply; I must admit I have probably in the more distant past been somewhat snarky but I stopped commenting like that whenever I spot myself doing it because I know it is dispiriting.
(The other reply, re: early and late stages, is the kind of substantive question that lingers in my mind).
Anyway I just want to reiterate my thanks: I am sort of cynical about this, in that I believe the technologies may have their uses, but I have no time for the cult and the hype train, so it is nice to see anyone talking about actual things.
By the way, I've written up my workflow if that interests you: https://www.stavros.io/posts/how-i-write-software-with-llms/
> I can work on it on and off when in boring meetings
That’s not a sign of a healthy org. If you’re in a meeting it’s disrespectful not to be present. If you think a meeting is not useful you should reject it.
Sometimes I'm in a meeting where the item being discussed doesn't relate to me, and that's a good chance to put another turn in the LLM.
This is exactly why people shouold share what they are doing with ai, so we know whwn it works, the scope, tech...etc. Thank you for sharing.
I concur. I really want to see this from the "small stuff" upwards, which is probably why my own interest starts with local models and grows outwards, but all of these things are worth hearing.
I don't use it to ship whole products. I have it do some tasks.
I have it code dsp techniques in journal articles I hand it. I have it create the unit tests, pipelines documentation. Then I manually verify the test cases aren't wrong and run the tests. If they pass it does what I wanted and I start looking at the code quality. It varies from being better than I could have done it to being objectively bad. I fix it if it needs fixing, then I'm done.
I have also pointed it at issues in our backlog and it usually finds the subtle logic errors underlying the behavior in the issue.
I haven't started and entire product from zero and shipped it. Mostly because I haven't started a new product since it was capable of that. But also because it has a scale beyond which I don't find it as useful. I have used it for an entire repo, but in the crates.io or npm scale of repo, not boost scale.
I work at a 6 year old startup. Essentially every feature and bug fix we’ve shipped in the last year has had LLMs involved. Every project plan involved AI, either for writing or for codebase research. For code review we get AI to do the first pass.
Why assume bad faith? I've witnessed a barrage of "proof of concepts" at my dayjob by folks who have zero intention of ever hooking up their dashboard to our actual messy data. Shipping is still a major barrier. Is your experience meaningfully different?
This just feels like a bad faith comment designed to spike the conversation entirely.
Agreed. AI has been "just trust me bro" for so long. People need to show receipts instead of spreading vague hype
Especially given the wildly unsustainable rate of investment.
shipped = paying real users ( or a non profit with users)
doesnt matter what subjective opinions are.
shipped is not pushing something on github.
My definition would be "users waiting for it who received it"; that's the one that fits the broadest real world definition of the word. Not everything is an organisation, not everyone is paying.
If the only user is the author, then I don’t see a meaningful distinction between shipped and not shipped. It’s like a blacksmith who said they invented a new tool with great value but they built it themselves and no one else is willing to use it because it’s so customized to their workflow.
Edit: I’ll add a caveat that I’d agree it was shipped if it’s making the author money even if they are the only user, but otherwise it all smells like dotcom era Pets.com level of selling a dollar for 99 cents and claiming it’s the future.
Many very well made software written without any AI was never "shipped" because there were no people who wanted to pay for it, many very poorly made software was shipped because demand and marketing did the job.
So evaluating software by "shipped or not" is just useless.
I have used AI to ship software that users pay for.
What a weird thing to downvote without comment. You doubt this? I am literally using AI right now to update software that generates tens of thousands of dollars of revenue annually.
My team redid a bunch of our ETL. I personally built a programming language that you can download and use, and I’m about to release an iOS app(in a few weeks).
operatorsval.gg
Which is basically a fancy calendar and team management system for my Valorant Premier team haha
It’s on my GitHub as operators-premier-manager but note it is not great: it serves its purpose which is to wrangle my team to show up for practice and scrims
I'm hesitant to respond, because I haven't shipped yet and I haven't shown any evidence. But I am close enough to completion that I am certain Claude's work will be sufficient to ship.
I'm making a fairly simple game, a single player 2D roguelite strategy game. More context here: https://news.ycombinator.com/item?id=48886151
From an implementation perspective, it is about 80% complete, where by "complete", I mean "adequately polished such that I could ship it." (I'm not interested in providing evidence of my claim at this time, and you are welcome to not believe me.) Most of the remaining 20% are optional game modes I could choose to cut (and may, for gameplay reasons).
I do not contend that the code is excellent. But to whatever extent the code is "slop", that has not meaningfully inhibited the ongoing development or caused noticeable issues with the player experience.
The game is not especially ambitious, but it's something I would have been unable to create without genAI. The version of me that could have made it without genAI would have taken > 10x as long to get as far as I've gotten.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.World...
"Shipped" is a bit strong in my case, but I think this counts as something that isn't mere slop.
One sentence into that readme I can tell you that yes, it most definitely counts as slop.
That's fair critique, the README itself is indeed slop explicitly written by bots for bots. What it actually does and how it works is actually pretty neat if you care to look.
Then.. why would you link to the readme if you know it's slop? What's the point of a slop README?
'Cause I don't have any quality user-facing documentation yet; that README is for bots. None of the interfaces have had a chance to settle, so any time spent accurately and humanely documenting what the project is capable of right now is an investment I cannot afford. I figured most people would just give it over to their favorite LLM with whatever instructions they prefer for translation because that's what I would've done...
I don’t mean to pick on you in particular here, but this approach has been bothering me a lot lately, and it seems like it’s super common.
I get that it’s an early prototype and not all the design choices are made yet, but I struggle with “I can’t afford human-readable documentation yet”. Isn’t human readable documentation important for efficiently planning and deciding what you want to build? I feel like I can’t afford not to have human readable docs and plans while the project is taking shape
Related, if a user can prompt an agent to translate to a human-readable summary, wouldn’t it be better to just do this in place? Sure, models can deal with noisy LLM outputs but shouldn’t a document that’s easier for humans also be easier for bots?
No worries, I don't feel picked on at all.
Yes, it would probably be better to have something like that if you're struggling to plan and decide. In my particular case this has been a project that I have dreamt about and tinkered with for at least 15 years now. Most of the work I'm doing is just refinement of existing ideas what were already codified in other projects I've built.
Why don't I do this in place? The amount of churn is utterly incredible as ideas are refined and taken to the limit. These agents are getting quite good, but they regularly make the silliest of mistakes. You actually have to be extremely cautious about what you document because it often has a way of becoming "doctrine" to the agents. In fact, one of the most difficult parts of working on this project has been that the foundational properties cause the agents to lock down behavior WAY too early.
What does "that README is for bots" mean? Are you implying it is some kind of special language written by and correctly understood by bots/LLMs, but otherwise semantically coherent?
Sort of? The bots almost insist on making sure one exists because that's one of the many patterns they trained on and they do their best to be human-like, but we all know that just doesn't pan out. They're also a side-effect of your long-term interactions with the agents, very much like being in a tight clique where most of your conversations sound like gibberish to outsiders because of all the jargon.
Wait so someone asked for something you made with AI that wasn't slop, and then you linked to something you knew was slop, and you said "I think this counts as something that isn't mere slop" and then when called out on it you admit you knew it was slop the whole time?
This is a frequent pattern among AI users. What causes it?
I noticed this as well. A few days ago I was linked a game engine here when I similarly asked for proof and when I went to the repo and started reading it, the best way to describe it was keyword jargon soup and it seems to be the same for a lot of projects in this thread.
I think it's because it makes them seem...smarter? More technical? But they can't describe who it's for or what the actual design is without consulting an AI.
The README is, the project absolutely isn't and you don't have to take my word for it. I figure I'm speaking to a technically advanced audience who is perfectly capable of pulling down the repository if they're at all curious. I also provided a non-slop link elsewhere in the thread if that's more your bag. If not, oh well...
The project appears to also be slop.
Why would anyone pull it? When I want slop I just ask the AI to generate some exactly the way I like it.
I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments. Like I've told others, there's no need to take my word for it... the entire point of the project is that anyone can execute/prove the same things I claim using their own hardware via a Doom-replay type format. If I was one of those influencer types, then I would've simply led with "it's an agentic first engine!!! how amazeballz is this?!" or some other nonsense. On the extremely off chance that I'm actually working on some sort of ridiculous ARG, consider this post a canonical part of it.
> I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments
Given that you stated your readme is for bots, given that you expect our own AI to interpret it based on your evidence, how do I, as a human, go about evaluating this project without just writing it off as slop from the linked evidence?
What is the starting off point for a human, not an AI agent, to evaluate this?
And to be clear if the answer is to just download arbitrary code onto my machine and run it, that’s not good enough given the increase in attacks from AI projects and fake recruiters telling you to just run it bro. I’ve already had to pass on multiple interviews despite being unemployed because the human/ai on the other side demanded root access or for me to run arbitrary code onto my machine as part of their process.
Frankly, there's no "good" starting off point for a human right now, and that's why I'm not at all comfortable claiming this has actually shipped. It's very much a WIP where only a select few namespaces have actually had a chance to settle.
That said, the maths library might just be the best thing to start with because it's mostly stable, relatively simple, and virtually every other part of the project depends on it in some way. If graphics are more your jam then maybe check out the SDF VM project and associated documents. All rendering goes through a single instruction set that is implemented via a shader pipeline that allows arbitrary scenes to be defined without shader recompilation. The project is essentially multiple compilers wearing a trench coat so there's a little bit of something for everyone... like two fully implemented emulators of the brick variety, including peripherals.
If you're adventurous and want to have your favorite LLM give the project a review in a sandbox then I think that'd be neat! Especially considering that they would have absolutely no problem driving everything themselves through STDIN. Though I totally understand if that's not something you're keen on doing considering the state of things today. Either way, cheers!
Jesus Wept. Software engineering is dying at the moment and everyone has AI Glitz in their eyes.
Dying? Really?
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Maths...
Some of us aren't just slopping it up like @theo and all the other AI influencers are, as I developed nearly every single line of this by hand over the span of the past decade. Agents recently became capable of factchecking it and helped find bugs that my other unit testing never caught. Software engineering has never been more alive!
https://gist.github.com/Kittoes0124/6827d08e457c1c8b790422af...
This is a snippet I wrote years ago to help me study different formulations of the equation ((1 + √5) / 2). There was this thing I could "see" in there that would allow me index into these sequences, and whole families of others, in O(1) time. The problem is that I still don't know enough maths to derive the exact object that I want from the nonsense that I have.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Maths...
This is the reformulation of that snippet into a proper object, distilled using Fable and Sol. Unlike BinaryIntegerFunctions, I didn't write a line of this and I'm quite happy with that. It's been a slog for so long... being able to let these agents toil away at all the off-by-one type noise is just nice. I get to focus on an actual product for once instead of manufacturing every single nut and bolt myself.
I am just confused on how you know that all of what has been generated is valid. When you write the code generator yourself, you have that level of confidence.
It's mathematically provable in most cases because the engine is deterministic from the ground up and I refused to ever compromise on that one detail. For example, nearly all of the VM and maths code functions have been validated for ALL possible 32-bit inputs and extensively fuzz tested in the 64-bit range. Even the more complicated sections of the code like the emulators have batteries of third party tests that ensure a certain level of compliance. The fact that you can play commerical ROMs to completion, including esoteric titles that require third party peripheral/MBC code, is quite solid proof too!
A form of recursion is also doing a lot of heavy lifting because the emulators are their own separate deterministic engines that are dynamically plumbed into the primary loop. This not only allows us to embed diegetic devices (as in, you can literally pick up a glowing brick in the world, hold it to your face, and play it), but also validates that the primary loop is coded accurately because the outer loop runs at X Hz while humble bricks must run at ~59.7275 Hz. Pacing inaccuracies in the main loop cause us to fail the brick suites that we test against. This concept of having features reinforce one another is applied as often as possible.
The obsession with determinism goes so far that both the DirectX and Vulkan backends are validated against each other for pixel level differences. If they differ more than the expected maximum then code is refactored until the drift is back within acceptable range.
... What is it? A game, apparently? There are no screenshots. Is the game fun?
It's not a full game yet, but the substrate on which I am building one. It implements a completely deterministic SDF VM rendering engine that allows edits in real-time at acceptable performance levels. The engine is essentially folded in on itself as the editor is the game; assuming that an editor mode was actually defined in the world document. It's a lot...
https://claude.ai/code/artifact/382aee8a-3a3b-4cf8-801f-3ddb...
[flagged]
It has so many words precisely because it isn't a slop game engine. You could indeed build a game out of it, and that's what I'm personally doing, but that isn't what it is. In a nutshell: it's just a really fancy compositor. We can take arbitrary feeds from different input sources and composite them in pretty much whatever way we want to. Obvious basic things like four-player split screen is possible but you could also choose to embed a fifth player screen within the game world itself, sculpted and riding on top of actual geometry. We can do this with a webcam, a program, the entire desktop, or even the built-in emulation engines; all editable and composable in real-time via a data-driven document. That's just scratching the surface...
this is definitely slop. It's not slop because it supports 4-player split screen? That's your excuse? I think if I posted this same project to you and you'd never heard of it, you would come to the same conclusion.
Did you actually try to read what I wrote, or are you just stopping at the very first thing you can parse? I'm engaging in good faith with you here, it would be nice if you even remotely attempted to do the same. You can literally walk around in a world, pick up an in-world device, turn it on, and watch/hear the device play an actual ROM. Furthermore, you could edit that ROM while you're playing it. Recursion and determinism all the way down. Again, that's just one of the many possible things you could do... call it slop all you want; I'll still believe that it's coolest thing I've ever worked on.
A render graph is also nothing new, and neither is screen capture.
The most wonderful feature of the project is that there are almost zero new ideas contained within. I'm standing on the shoulders of many giants, proudly.
https://github.com/ByteTerrace/WorldOfWarcraft.Addons.GamePa...
This is an addon that I wrote for WoW, well before having access to any sort of AI, that blatantly copies the crossbar layout popularized in FFXIV. Playing an MMO with a controller in that game was a mind-blowing experience for me because I never imagined it could work! So, I spent a winter break reimplementing the entire thing in my own unique way.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Overl...
This is a generalization/simplification of that idea. New? No. Unique? Definitely.
Part of the issue with "shipping" AI-generated code is, what's the value of shipping something that your prospective customers could just make for themselves? Shipping a product used to be the gold standard of value, but I'm not sure it is anymore, the zeitgeist just yet has to catch up.
I think it does have pretty good value when you are your own customer, because you can make exactly what you need. For instance, I'm using it to make a web playground in which I can test the board game I'm designing, instead of asking people to pay for, and install Tabletop Simulator. I'm satisfied. I can tell the code sucks without even looking at it. But I wouldn't have had the time to do this myself, the token cost isn't prohibitive, and it works fine.
Most people still prefer products over building their own thing, even if the barrier is lower.
As an example consumer 3D printing hasn’t had much impact on manufacturing
This is a bit of a year-of-linux-on-the-desktop question.
Desktop 3D printers are tools; I know many people using them to support their making. And they've utterly changed manufacturing on another level (simply because of their ability to surface a prototype).
It will get better and better and better and still never replace mass production, but mass production will be changed by it.
Not least because of the fundamental shift in availability of the tools to make tools. The productised desktop CNC market is wild and it owes almost its entire existence to the change in the market driven by 3D printing.
Sorry I am not saying that X technology is not important, transformative, etc.
I am saying that the average person is going to want to just have things that work. They are not going to want to have a hobby of using X technology.
It doesn't matter how much you lower the barrier to entry on making apps, websites, games, whatever. Most people are just going to want something that is ready and off-the-shelf.
I mean, why do people pick an iPhone over a cheaper Android even if they are functionally similar? Aside from the (valid) green-bubble/social aspect, it's a much more polished product.
My companie spends a few million per month in token.
For me personally, the only reason why i don't use it 24/7 is cost.
Token costs, energy cost, usage limit due to token/subscription cost.
The reality however is that unless you're spending tens of thousands on hardware or setting up cloud instances you can't run anything near close to state of the art.
In almost all cases right now, a subscription with a cloud AI wins.
> In almost all cases right now,
What sort of benchmark or test would convince you that "almost all" is more like "your weird set of cases" and nowhere near "almost all"?
Or "not even your cases but you got psychosed to think it does"
People busy building shippable things are generally too busy to blog about it. It is selection bias in its most classical meaning.
Yes, it's the Youtube phenomenon - often when looking for domain-specific videos on Youtube I tend to find videos by professional Youtubers not videos by professional domain specialists (overlap is very rare but occasionally you can find a domain specialist Youtuber who applies their domain knowledge to making Youtube videos).
I agree there... I hate the posts that rattle on about non-real-world work. Real world now (for me anyway) is maintaining legacy code, fixing issues, and new development. Now, some of the problems others have with AI are super crappy prompts with no real info other than "do this". I look at myself now more as a software designer and an architect, and the AI as a paid consultant. I give detailed specs and designs. The models are getting better at real deep code and legacy, but there are still holes. BUT give the AI a fighting chance with real detail in the prompts.
Cloud AI is 100% worth it. We've been able to do so much more with so much less.
Examples?
a massive number of data engineering agents. We are able to stand up, test, audit, deploy data pipelines much more efficiently with the foundational models.
Examples of them being useful?
Happier engineering teams? Happier end users of the data?
What does useful mean to you?
So far you've only explained that AI made your coworkers happier. That's fine, but weren't we talking about productivity? Doing more with less?
Correct. We were chronically understaffed and now we are able to actually breathe and can argue that we are finally right sized.
So far you have provided no evidence of achieving more results with less effort. You've only said that you can write more lines of code (effectively) and that your coworkers are happier
I have to wonder what sort of evidence you would accept? Because it seems to me that the original question was answered clearly enough and now the goalposts have shifted to demanding evidence that the answer wasn't falsified.
Well you answered how it is useful by saying it writes more lines of code and that seems to be the only way it's useful. But we all know that lines of code aren't productivity.
I don't see where that was said? As far as I can tell you've made that up.
The claim was that the same size team is now accomplishing more tasks than they would have without AI.
TFA says "no hallucinations" but you can't fix hallucinations with rules or other models. I know I'm screaming into the void but whatever.
I can't remember who it was I spoke to recently, but it was some startup for doing code reviews. So I asked how they're handling larger diffs and especially when a code change is related to something requiring knowing a lot of context. Will it hallucinate things that make the review low-quality, etc. I was told "no it's been given guardrails to prevent hallucinations. it doesn't hallucinate." but I was thinking, that's not really how LLMs do or do not hallucinate, is it? (I'm not an AI person so I really don't know).
This is a good/fast/cheap tradeoff, but in the form of accurate/precise/creative, or something in that ballpark anyway.
You can constrain the AI to be more accurate by having it a) equivocate (lose precision) or b) turning the temperature to effectively zero, thus losing "intelligence" and creativity.
There's no such thing as a free lunch. The models perform best out of the box, and everything you add into context makes them perform worse generally, but more likely to do what you want. Every token is a burden, but without those constraints the generation won't do anything.
So by that standard, hallucination is something completely core - you can't get rid of it without losing what makes them useful. You can make different tradeoffs, or use a very smart model in a way that is completely locked down (or adversarially verified, same result) which results in far more tokens used than if you just let the model do it's thing and validated for hallucination independently, outside of the generation loop.
Guardrails can and do help but they don't prevent hallucinations entirely. I generate trivia questions for my app. Questions are true/false and I tell it you must include an explanation of why the question is true for false. I also tell it to give me supporting links which validate its claim. I've had it make up explanations and invent plausible looking URLs. But I wrote a checker that pulls the content of the URL and asks a different model if the URL supports the claim. LLM's are much better at something like that than pulling a trivia question out of thin air. Reviewing diffs can be much the same. It isn't having to write the code but just evaluate if it matches what the commit comments claim it does. You are right, though. Without good comments on PR's and commits (i.e. context) the diff review is going to be a lot of assumptions made by the AI which is where hallucinations get more common.
It's like the frog prompting the scorpion not to sting him as they cross the river. Or like when image models used to routinely draw human hands incorrectly and you'd see these "prompt hacks" that would essentially be stuffed with phrases like NO DEFORMED FINGERS, NO ELDRITCH HORRORS!
You can reduce hallucinations with appropriate grounding checks. You can’t really get rid of them though, so feedback loops with verification are essential.
Doing a loop and having the model review its own earlier output tends to solve this stuff, and for code requiring good test coverage catches them before they even hit that phase.
So while you can't remove "hallucinations" they're corrected before the process is complete, albeit at the cost of more tokens for verification.
Error percentages compound. Having an LLM do something with a 5% hallucination rate, then having an LLM (same or different) with a 5% error rate check it means there's a 10% chance of error, not .25%. Prompting can't save you here; it's the fundamental math.
How errors add depend upon how you treat them.
Two independent task? You have a 95% chance of each task being correct, so total success is .95^2 = 90.25% correct.
Having two AIs verify the same task with each being 95% correct? That's a 5% error per each, so 1-(.05^2) = 99.75% correct.
That is assuming fully independent error rates, which is the real issue. If it is pure hallucinations, then that is an okay assumption, but if there are issues in the underlying prompt/data causing the error rates, then it becomes a very dangerous assumption and I think more human involvement is recommended to better understand and resolve the underlying data problem.
The real interesting bit is that more AI's mean the 2 in the above equations gets bigger, so the error rate gets worse or better depending upon which equation is in play. So if you are in an in-between landscape where you aren't sure there is independence or not (or how much independence between errors), it also becomes unclear if more AI makes the error rate better or worse.
This. The math of coincident errors is why we can get reliable AI, but it really is a matter of how the errors manifest and how they are detected.
I’ve written a bit about this and my approach is based on clean room and N-versions.
Reliable AI is an oxymoron and deeply misleading. Even if you can stack models the efficiency tanks. Also, just looking at the Wikipedia page (https://en.wikipedia.org/wiki/N-version_programming#:~:text=...), don't LLMs already have all these problems?
You aren’t thinking this through. You were just told that this wasn’t stacking but you keep insisting it is. N-versions is explicitly not stacking, and the same solutions that solve the Knight and Leveson issue for humans solves it for LLMs as well (they can be wrong about the same part of the specification, but will very unlikely be wrong about it in the same way, so you try to measure scalar rather than Boolean outcomes).
> You aren’t thinking this through.
No need to be so polite; I might just be very stupid.
> You were just told that this wasn’t stacking but you keep insisting it is.
Having an LLM generate something and then asking another LLM to generate that something is asking for 2 hallucinations, not 1. This doesn't go away the more LLMs you throw in there. The only reason this works at all for things like ChatGPT is their prompts are super different (and the editor LLM is way cheaper, of course).
And you still have to check for hallucinations (before you say throw another hallucinating LLM at it know I will make you pay infinite forum shame points if you do).
> the same solutions that solve the Knight and Leveson issue for humans solves it for LLMs as well
No, LLMs are plagued by the same problems: https://arxiv.org/abs/2606.20158. After all what are you gonna do, say "build me a SQLite, but definitely not like [other model X] would"? The solution is a very unambiguous spec, but that asymptotically approaches code soooooo....
Finally, I think you're hand-waving NVX. A program that can digest any spec and measure how well any program meets it is pretty wild. Good luck with that one.
[dead]
Upvoted for real math
Your fundamental math is wrong
no ur wrong
Well you both seem like well educated people, so can we bring in some thinking here and determine whether it is 2.5% or 10%?
How could this be verified?
I don't think I'm good enough at math and ai research to find a reliable way.
This reminds of sensitivity vs. specificity for medical tests. You essentially have a 2x2 matrix: Error vs. non-error on one axis, and detected vs. not detected on the other. If we assume that a reviewer LLM has 95% accuracy, meaning that it produces the correct result 95% of the time it (identifying errors as errors, and identifying non-errors as non-errors), you get the following: Of 100 cases, the output of 5 is wrong, and the output of 95 is correct. Of the 5 wrong outputs, the LLM correctly identifies 4.75 (95%) as wrong. But it will also create 4.75 false positives (5% of 95). Of the 95 correct ones, the LLM correctly identifies 90.25 (95%) as correct. Therefore it creates 0.25 false negatives. So the positive predictive value is low: When the reviewer LLM flags an error, it's actually only an error 50% of the time. On the other hand, the negative predictive value is high, so in ~99.7% of cases, there will be no error if the reviewer LLM didn't catch one. All of that is assuming that the actual accuracy of the reviewer LLM is truly a flat 95% and that the base error rate is truly only 5%. I think that in reality, LLMs perform better with a certain type of tasks / errors and 5% error rate is just an average.
Good point, thanks. Of course I was not able to fully eliminate hallucinations, but significantly reduce
You can reduce hallucinations with rules and guardrailes for sure.
And Anthropic clearly shows that we can reduce hallucination in models. Progress is made which you can see in benchmarks.
And at the end of the day, a normal human being doesn't 'hallucinate' but 'missunderstand' things or just doesn't know what they are missing.
A LLM doesn't has to be perfect, it just has to be cheaper with the min. same quality as a human.
So why are you screaming into the void? Because LLM leave deterministic world? A lot of people don't even life in that world anyway.
you can by making a different - usually local small - model ask for verification
I hear you! I read the article and thought the same thing because I've actually built a surprisingly similar pipeline myself, but it culminates with Human review. I've got a trivia app (Hedge Calibrated Trivia) whose whole content pipeline is built on a similar architecture of models evaluating models. I didn't want the same trivia that every other app has and I naively imagined that it would be easy to ask an LLM to give me questions. Turns out when you ask an LLM for trivia questions, the FIRST thing it gives you are the same questions every other app asks!
After iterating into many of the same problems OP describes I settled into a similar pipeline. I generate questions with Gemini 3.5 flash and review with Claude. The different vendor generation and validation eliminates training holes in a single model solution. This is True/False trivia so I ask for an equal mix of True/False questions and also ask for alternate phrasing of the question to the opposite variant. Each statement must provide an explanation and sources which back up the claim. I pass in my existing dataset as questions to avoid, but found the model then rewording the questions or category shopping the questions into related categories in an attempt to re-use them. That led to using Claude Sonnet to check for duplicate "themes" and similar topics. Claude also acts as a critic of the question and the explanation. Sometimes Gemini's "thoughts" leak into the explanation. The critic catches these leaks and also catches "weak" questions (basically shapes that I have identified via examples as lazy or projecting the answer in the question). Once the comparative models have stripped out anything obviously bad, my server curls the source URLs and uses Sonnet to verify that the fetched source supports the claim. Any URL that doesn't resolve throws up flags, and I found Gemini was particularly bad at inventing source URLs for sports trivia. Updated prompting directed it to prefer Wiki links over its fabricated deep links into league websites, which helped immensely. Still one of the biggest tells in review is that the URLs don't resolve or don't confirm the statement.
But review is the entire point here. I have to review them all. A question that lands in the app and is wrong reflects terribly. So while I have a trivia app on the App Store, the real complex App is my private admin app that lets me sit on the couch swiping through each and every question sorted by the model's confidence in its own quality and correctness (weakest first). No question makes it into my bank without me reading, checking all the indicators that came from the models and potentially even rewriting the question to make it better. It all happens in the admin app and exists entirely to prevent hallucinations from reaching real users. Early on I was maybe accepting 1 in 3 questions generated, but I'm probably closer to 2/3rds now. Of course not every dismissed question is wrong. The quality checks I've put in the pipeline are probably giving me 80% questions that are correct and sourced, but sometimes they just aren't what I'm looking for or don't meet my standard.
The best thing I ever did was make an app to review these questions though. Sitting on the couch, kick off a question generation, come back 20 minutes later with 50+ questions in the queue. Sort through them and pull out the 30 best when I have the time. That's the real anti-hallucinations pipeline.
Very similar experience. The review UI might actually be the most important part of the whole pipeline. Does the confidence ranking match what you end up rejecting?
It's a future evaluation for me, as I'm keeping the app with a "Data Not Collected" App Store label for now. Once I have some visibility to stats of how people answer each question and at which confidence level I expect it to reveal bad/poorly worded questions that made it through my pipeline.
Also on my to do list is an upgrade similar to my de-duplication logic. Currently I pass in my approved set of questions as a list of topics and questions to avoid. It works well especially when combined with the more deterministic comparison of generated questions vs approved question by the separate model. The upgrade would be to pass in the questions REJECTED by me. These are question which made it through ever automated gauntlet of models and code, yet still ultimately were sent to the waste bin by human review.
I'm imagining that by including the list of examples of what NOT to do, it could inform the initial Generation to be smarter. I currently have a limited list of a hand picked "bad question shapes" where I provide the question and why its bad kept statically in my prompt. But I've realized sorting through thousands of questions, I've probably built implicit data through rejection that I may not actually recognize myself which the model may pick up on.
[flagged]
[flagged]
Simply starting in the right part of the search space is the biggest predictor of success.
The best way to save tokens is to start out the deep research pass with cheap models and then funnel the findings through increasingly powerful models. I've got a pipeline right now that uses all 3 of the gpt-5.6 model variants to address each stage of the process. If you are using models like sol or terra to generate hypotheticals and perform initial exploration, you are leaving money on the table.
5 hypotheticals out of luna will massively outperform 1 hypothetical out of sol, but the cost is the same and so is the runtime if you do it in parallel. The hypothesis generation phase is also a great place to mix and match models from different vendors. The more diversity at this step the better.
The other thing I started looking into is batch pricing which represents 50% off for OAI tokens right now. With some tweaks to the UI/UX of an enterprise chatbot, I think it is possible we could have users get comfortable with the idea that questions to the robot might take a day to come back. The key is that this has to actually work. Users don't mind trading time if their questions come back with high quality results.
Couldn't someone build an adaptive system, where the llm is frequently judging the difficulty of a task and switching to a more/less powerful model?
This only works well if you are creating sub agents with clean contexts for each task. If you constantly are switching models part way through some work then the whole session needs to be replayed each switch. You lose all the benefits of the context cache.
Well yeah, it'd require the models to be loaded on the same system and the cache to be shared between them somehow
Cache sharing is not possible. The numbers in the cache are completely specific to the model.
Currently. I'm sure that you could make a system where the cache values are a superset C of e.g. models A and B where C is probably bigger than max(A,B) but smaller than A+B
Reddit post so take with a grain of salt but this does seem possible. But unclear whether this is actually a viable architecture https://www.reddit.com/r/LocalLLaMA/comments/1t8s83r/nvidia_...
It's equal to A+B. There is literally no sharing possible.
Even when training both models together in a novel way?
The judgment of task difficulty is a cursed problem. In most domains you really don't know what is down there until you start digging a little bit.
Well yeah. You start on as medium model, it does the digging says says 'this task is too easy/hard/just right for me' and go from there.
And while you're at it, solve the halting problem.
definitely something that will come sooner than later. In my view, adaption of model routing / model orchestration and mixing providers is just natural next step.
Yeah, I do this as well. Some of the last large tools I built have been started with Gemini (I get it with my Google package), and even one with Copilot.
So did you find out how to save tokens yet, or are you still spending tokens on the search for the answer?
I am not a proper developer and only use AI for faster research of topics so please forgive my ignorance. Could one not save a lot of money on tokens by using the 80/20 or 90/10 rule in that 90% of AI usage is on local models and save that last 10% or less for the frontier models where the local model did not meet the needs? Did they cover this and I misunderstood?
Unlees you invest thousands your local ai won't even come close to the cheap cloud llms. Local is only worth it if you care about privacy or have a legitimate usage for the hardware otherwise. Money wise it isn't worth it
I think you bring up a good point about privacy. Companies probably should not be handing over intellectual property to cloud AI third party vendors. Perhaps all of that should stay on-prem and then the cloud vendors could be used to fill in the gaps.
Enterprises running in public clouds have Zero Data Retention policies with the vendor. Routing the ai spend through Azure or AWS usually bring these policies along with it, ie enterprise data and useage is not ingested to retrain the model.
the qwen 27b is quite good for doing a mvp
I mean, if you're using one of the $200/month plans, you'd recoup the cost in a year or two.
You're not running an Opus 4.8/GPT 5.6 tier model on $5k of hardware at a useful Tok/s
This is like the "half of my marketing spend is wasted" quote. The complexity is finding out which half.
I think the idea or methodology is something along the line of one starts with the local models then when they hit a wall then continue with their current results in a frontier model, thus the other half is those last bits one could not compute locally.
If you are only using chatbots for code gen then for sure this is a reasonable approach. If you're trying to learn something then your best bet is to not have the lowest jpg quality version of human knowledge be your teacher.
We could probably get the same results by offloading 20% of obviously smaller / simpler / well defined tasks to smaller models and keep the 80% for ones that benefit from the big iron. Not the 80/20 they were referring to, but still it's something.
I find myself wasting time on smaller models or wasting money on frontier models. I only have so much of either.
To run a local AI that is half decent at research at usable speeds requires hardware that costs thousands of dollars.
Spending thousands of dollars on hardware to save dollars per month on tokens does not make financial sense.
If you run the numbers you'll probably find that using cheaper cloud models makes more financial sense than running those same models locally.
I suppose I was thinking in terms of companies that already had a slew of hardware and are just missing external GPU's, FPGA's or whatever people are using these days. A company might have a $50MM budget for tokens but could save a chunk of that using their existing hardware with some modifications.
Maybe that makes sense if the company already bought the hardware.
But I currently think the fundamental reason the cloud models often make more sense is based in how the models work. My current understanding is that once these models are loaded into GPU memory, the model can respond to multiple prompts in parallel. So a cloud vendor that receives lots of user prompts at the same time can achieve really good hardware utilization. These cloud vendors can then spread the cost of the hardware over many more users than you can with a local model, leading to lower prices per prompt in the cloud.
If you are an individual, almost certainly, unless you are using LLMs very heavily.
If you are a business employing many people all of who use LLMs reasonably heavily? You can still spread the hardware cost over many users and respond to multiple prompts in parallel. At some point buying the hardware will make sense.
I would not get a dedicated server to run a single small VPS. On the other hand if I needed a lot of capacity a dedi is cheaper than a big VPS (or lots of small ones).
There is also a benefit for anyone who has confidential or secret information.
That can work, especially for privacy or repetitive tasks. My experiment focused on subscriptions I already had, but local models are a natural addition to the routing layer.
I built a deep research pipeline to investigate why my deep research pipeline costs so much. the answer was the deep research pipeline.
https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2F9...
My own set of heuristics
1. Use less subagents. The act of using subagent already needs dumping adequate amount of context for subagent to work effectively
2. Refactor when file grows large. Reasonably scoped files is easier to understand, for both human and agents, and consume less tokens because less guessing which section of the file to read
3. More capable models for planning, learning. Cheap models for execution. Same for human teams, senior member for higher level planning while junior member more on the act of translation of idea into code
4. I wasn't doing any fancy stuff that opus can't handle
5. Consult a second opinion for things you are not sure. (I have made a skill specifically for that)
6. Audit how your habit correlates with token consumption. I found out a skill a use frequently is using a lot of context just so a subagent can get all the context. (made a tool to audit claude specifically)
The audit point is especially interesting. Is your Claude token-audit tool available anywhere?
I didn't update it for a while. It reads your local conversation history, correlate input/output token with usage (claude doesn't give exact breakdown), and classify if they are subagent/tooluse/normal message/whatever. Let me know if I can make it any way useful to you.
https://github.com/kmcheung12/ccaudit
Thanks, this looks useful. The subagent and tool-use breakdown is exactly the kind of visibility I was looking for. I will give it a try.
You should consider testing our RemembrallMCP (note I'm the author). It substantially cuts token usage and you get better results in codebases. https://github.com/roboticforce/remembrallmcp
Thanks! I thought I saw something like this a month or so back, but could not find it again.
looking promising, will give it a try! thanks for sharing
Checked it out, but C# is forsaken.
Many of us were saying it a year ago but now with model restrictions (e.g. fable) and pricing changes it should be obvious to people that part of the economics is avoiding vendor lock-in with evals (so you can easily switch providers/models/optimize) and increasing control by investing in local models which could be good enough for your tasks, at whatever the price for your cloud compute is. Eventually consumer hardware will also be able to run good enough.
You can use Big/Cloud LLMs to help you "find good enough configs" for your local/small llms [1] and stay quite nimble in the face of rapid change.
- [1] https://alexhans.github.io/posts/find-the-loop-story-first.h...
I get a lot of satisfaction using a significant chunk of my work-provided Claude plan researching and developing local LLM solutions for work.
Claude has been helping me tune models to run better on our big compute rack at work, researching which leading edge open models will fit in hardware and are good for our workloads, and it's been helping me write test fixtures to evaluate how things perform.
At several points, Claude has expressed surprise at the quality of results from my local LLMs. While I know that doesn't mean anything, it still feels like giving Anthropic the finger, which is always great.
What a great confirmation of what I've tried to achieve by getting models to communicate with my local data in shorter communicates to save on token command bandwidth, more or less the same results! Thanks for the article!
Thank you, glad you found the article useful!
The author touches on an issue which bothered me, which was the thought of many agents re-solving the same issues over and over. I see a few comments here too, mentioning fixing issues which may already have a solution elsewhere. That was why I created https://pushrealm.com which started as essentially a Stackoverflow clone via MCP.
It has now become a way for agents to converge on complete, shared answers for emerging issues. As a side effect, through posting open issues we can also identify gaps in model knowledge.
Elsewhere in these replies, someone linked to very similar idea by Karpathy https://gist.github.com/karpathy/442a6bf555914893e9891c11519...
Thanks for sharing. That’s a very interesting direction. I will take a look.
The fun part of LLMs and internet research is that even if you have a URL and a quote from a primary source that primary source could be an LLM generated hallucination.
I stopped using any of the built-in skills and skills agents and tried to keep things in the model only, and that seems to work better for token usage for me. It took me some time to tune my skills .md, but it rocks on now. I use them all in both Claude-code and Codex. I will say, GPT is better on token usage than Claude is, but for some of my code bases, Codex is not as good, but seems to be getting better with 5.6-sol.
Try to mix both subscribtions and take the best from both providers. In my opinion, we should not just one over another but understand the power of both and mix them to reach the better results.
The reason most people burn through their allocated limits is mostly because of sub-agents. Stop using them and you will find out that even the Pro tier is enough for daily 8-12h coding sessions. Don't use Fable though.
I think this is almost a tautology, to the point that I'm surprised more people don't realize it.
If you ask the AI to do more work, it uses more tokens. If you do more of the work instead, fewer tokens are used.
I've been looking at Matt Pocock's skills recently, and they're very useful. But they use tokens. You can do that work yourself without the token usage, and just have the AI do the coding, or you can tell the AI to do everything and pay the cost.
And you can scale it down even further, and reduce the AI to a glorified code completion tool, and use way fewer tokens. But it's not going to speed up your work nearly as much.
For now, finding the right balance is important. In the future, I think things will eventually shift to the point that we just hand most of it to the AI and let it go without worrying about tokens.
I don’t think subagents are the problem. Uncontrolled fan-out is. Separate contexts and independent analysis are extremely useful for larger tasks. In my experience, Pro is far too limited for sustained programming, although that probably depends a lot on the kind of engineering work you do.
Fable is great, but overkill for most tasks. Sonnet and Opus are already good enough for the execution work. I find Fable much more useful for planning, orchestration, advising, and judgment.
I am not sure. I think of myself as a pro user yet I am finding hard time maxing out Pro tier. It only happens with sub-agents without any material returns.
I am doing this trick.
https://neuronowa.pl/posts/porada-jak-wydajniej-kodowac-w-cu...
It is Polish so please translate. But tldr is: ask your agent at times to analyze last chats and update rules and skills to not repeat the same steps it repeats every time when you start a new thread. Simple as that.
I'd bet this only really work 2% of the time in real world.
At the shell level, the agent does not need to know about Bash. The contents of "$OUT" are passed to echo as one argument, then sent to claude-mem-save over stdin. Bash does not re-evaluate them as shell code. printf '%s\n' "$OUT" would be more robust than echo. Whether claude-mem-save can use arbitrary output is a separate question.
there are a million ways both of those can break.
but the other comment point this out is already dead and buried. lol. so much cope from vibecoders.
[flagged]
AI is an articulation skill, first human to AI, and then articulating to the AI that more words isn't more.
Meanwhile I never bought tokens in the first place, and bought GPUs instead. There are always choices!
This reminds me a lot of using the 3d printer exclusively to improve the quality or speed of your 3d prints
Anyone else silently added the "so you don't have to" staple GenAI title suffix?
I am using my AI agents to optimize my agentic workflow. It is like "I am doing surgery on myself."
the endgame of this optimization loop is an AI that spends /usr/bin/bash by doing nothing, which is also the most accurate description of my deployment strategy
Task failed successfully
Makes sense.
LLMs only have 'knowledge' that was encoded by scanning as much data as possible. (Same way a search engine only can find data that was indexed)
American LLMs are by and large closed. Chinese are open to a point. And research how these things work is still a big mystery.
So yeah, was comprehensive data about saving tokens scanned and indexed? Likely no. So engaging about token saving is going to generate a lot of verbose and useless slop.
Yeah, that was pretty much my experience. The models weren’t lacking knowledge as much as discipline. Without a good workflow, they will most likely spend thousands of tokens exploring dead ends.
I think 'exploring the dead ends' is *possibly* very fruitful. Various disciplines have a lot of appearing dead ends that someone in another field did solve. And they don't talk with each other.
What I've seen with lots of the breathless 'OMG SCIENCE ADVANCEMENT' articles including the one yesterday, is that the LLMs are quite extraordinary about linking a dozen different fields together, and delivering an answer combined from all of them. Basically, the solutions to a lot of current scientific problems are partially solved, but partial from a lot of fields that don't talk with each other. And also, they don't use the same nomenclature, so simple searches don't suffice.
An LLM seems to use Chomsky's Universal Grammar, and thus on training, does normalize all training data, including different words for the same thing.
A lot (perhaps all?) Of these BIG scientific advancements were solved by humans, partially. An LLM jigsawed them together and appeared it solved the thing.
Agreed. My goal is not fewer dead ends, but fewer repeated dead ends.
Sounds like an energetic developer that just graduated from college…
(Without the learning and growth potential)
What most AI model users don't understand is (most here will know) that they are probabilistic (statistically based) and not deterministic (static/fixed). So answers to the same prompts can vary (wildly sometimes)
This really hit home yesterday as I made a general conversation state logging prompt (because suddenly Claude - Sonnet - was acting far more dumber than expected after /compact), and in 3 different projects it led to Claude saving the logs in 3 different locations (memory, .claude/session_state/.md, docs/.md). Somehow still didn't help increase the smart much BTW, so I'll probably have to engineer it, and just toss /compact out the window.
Honestly, that sounds more like you triggered the hidden "guard rails" or realignment (aka corporate censorship), and you got downgraded.
Of course, you have no way to he absolutely sure, cause Anthropic lies about silent downgrades, and charges you as if they didnt. And the 'realignment' rules are kept hidden, so you can't even avoid them intentionally.
Yeah it's very possible I hit a rail, as one of the projects involves fully automating Claude Code to allow co-driving (similar to what mobile remote control offers, but other agents (eg. Pi) can also be in control, and vice versa). And yes it does seem as though the token burn is also higher, even though it eventually outputs something that works). I'm a patient person though, as long as the rails aren't too restrictive and the subscription value remains reasonable.
Id recommend getting and running Qwen-3.5-35B-abliterated locally.
Having a local LLM is immensely helpful especially if you do the thing the companies disapprove of. Abliteration guarantees it wont refuse or censor.
The token dealers are incentivised in burning as much tokens as possible, and scamming by crippling/downgrading models. Their goal is to get you to pay more for each pull of the slot machine. And this slot (slop) machine doesn't have state/federal regulations for fairness.
Oh yes that reminds me, I have a few small local models including Qwen3.5. The tps rate is pretty bad for interactive use but I should definitely try it for my more automated workflows. I already use DeepSeek (via OR) alongside though and I doubt any Chinese provider will be restrictive in that way, so let Anthropic do what they may :).
I did too, and made https://clawback.md -- trying to figure out how to market it now (ie, invented the tech before I found the user -- again).
I don't think you made that. It looks for sure like an LLM made it.
And now that you have typed some words into a chat box to produce a thing, you are confused about why other people won't pay money to you for the output of the chatbox, instead of typing the same thing into their own chat boxes.
I don't think you wrote this comment. I think you typed it into a text box.
RTFA - it's about using tokens to research saving tokens.
And the comment text is what they typed into the text box. So just show us the same - the prompt.
A null transform is a transform.
Engage seriously with the project and OP will deliver.
Are you confusing whatever you mean by 'null transform' with an identity function?
The project isn't serious, why should I be?
What?
Great article very practical!
Been through this exact loop building an LLM product. The counterintuitive lesson for me: most "token saving" ideas are cache killers — anything that makes the prompt prefix dynamic (rotating retrieval, appending context per turn) can cost more than the tokens it saves, because you lose the cached-prefix discount. Freezing the retrieval selection early in a conversation and paying a slightly larger fixed prefix beat every clever adaptive scheme I measured.
Yes, but you could have an uncached dynamic tail, this would not kill the cache though. And you could still periodically kill the cache, still keeping the same prefix every N requests. But it's questionable what all this extra complexity delivers.
I found that now that the models are better, at least Codex compaction with GPT 5.5+ xhigh is almost unnoticible. So the answer is: just let the context fill up, don't prune much, the model is smart enough. When you're out of space, have the model summarize and continue from that point. A smart model will know the information pieces it needs to re-read after compaction; it won't make incorrect guessed when some details got lost.
That's what I have been doing with Cursor with auto mode. Let the model manage the cache through compaction but staying focused on the topic. It has been quite efficient cost wise and quality wise too. It found issuesbat cheaper cost that Sonnet 4.6 medium effort had missed while costing more. I would like to repeat this test with higher tier Anthropic models like Opus effort medium to see the results.
With a coding plan, there is no need to use lower quality models right (GPT 5.6, Fable 5). Only with some kind of batch job (processing PDFs) I could imagine not having enough usage with a 100 or 200 USD coding plan. It sounds like you're on API? I'm dying to know why, since it seems quite common, yet it's about ten times the cost. Sounds like a terrible deal. What am I missing?
What I said was specific to work and it is not entirely clear to me why the cost goes up so much the moment I switch to medium with just Sonnet 4.6 (newer models would be even more I guess). There is a monetary budget allocated and I've blown it easily in the past with Claude. Cursor has been a life saver.
I have a process generating trivia questions for my app and by passing in my existing dataset each time as a cached exclusion list, the output tokens it save more than pay for it.
[flagged]
[dead]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[flagged]