nice writeup! looking forward to doing some more training as soon as i get some more data sorted. it'll be a custom arch, but i'll probably shoehorn it into unsloth for a speed boost.
While I do admire Unsloth (especially their https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF binarizations), the linked blog post looks like written by AI from notes (unless a human author acquired this taste from interactions with chatbots).
To be clear, I use AI for editing all the time.
Actually, diagrams are nice.
Just some pieces like that look like copy-paste (I mean, empty lines before, code get no special typography, etc):
If we write the boundary information for a packed batch as:
B = { lengths, cu_seqlens, max_seqlen, mask structure }
then every transformer layer in that forward pass consumes the same B.
If the model has L layers, rebuilding or re-synchronizing on B once per layer is not new work. It is the same information being reconstructed again and again.
In other words, the useful work is:
build B once, use it L times.
The wasteful version is:
build B + build B + ⋯ + build B (L times)
What’s with the all the hate for AI assisted writing on HackerNews? It’s a tool and people use tools all the time. It saves TIME and helps in improving coherence of one’s articles.
> What’s with the all the hate for AI assisted writing on HackerNews?
I don't think it's specifically for "AI assisted writing", any lazy writing gets hate on HN, the bar for quality just sits higher for better or worse.
> It saves TIME and helps in improving coherence of one’s articles.
I agree that it saves time for the author, but for the reader it has the opposite effect, and if you're unable to write coherent articles without the use of LLMs, maybe solve that first instead of patching over the problem.
It destroys the previously implicit contract that the writer actually spent a decent amount of thought and time into the writing, and that the ideas expressed are theirs and original.
I don’t mind good usage of LLM assisted writing, but if the author can’t even be bothered identifying the most obvious AI tells, I use it as a proxy that the author probably but very little effort into the article.
It’s also often a horribly verbose style, where the same ideas could be presented with 20% of the prose.
It’s also ruining the entire experience on web communities (although here on HN the moderation team seems to get a hold of keeping them at bay at this point, much appreciated).
All in all, it’s objectively a net negative for the readers, and serves only the author.
I prefer original, less coherent articles that are genuine and where I know the ideas expressed are really the author’s and not the LLM’s inference.
Last but not least, I don’t think the grandparent you’re replying to was particularly hateful in the grand scheme of things.
Because AI writing is lazy and moreover, I don’t want to know the AIs opinion on something, I can get that myself, if I want to read someone’s article I want to hear that persons words and that persons opinions.
If someone has no opinions or unique insight then why would I listen to them or read their content.
Again, if I want the AIs view on something I can open up Claude and ask them myself, why bother reading generated articles that took 10 seconds for someone else to prompt?
nice writeup! looking forward to doing some more training as soon as i get some more data sorted. it'll be a custom arch, but i'll probably shoehorn it into unsloth for a speed boost.
Thank you!
While I do admire Unsloth (especially their https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF binarizations), the linked blog post looks like written by AI from notes (unless a human author acquired this taste from interactions with chatbots).
Oh thanks :) We're also going to add MTP support soon for Qwen3.6!
95% of it is fully human done - the maths, algos, code snippets, screenshots & benchmarks are done / conducted by us and NVIDIA :)
We did use AI to fix spelling errors + made some nice plots using Chat (ours would look horrible lol)
Update - Just got rid of the spiced up intro
Thanks!
To be clear, I use AI for editing all the time. Actually, diagrams are nice.
Just some pieces like that look like copy-paste (I mean, empty lines before, code get no special typography, etc):
What’s with the all the hate for AI assisted writing on HackerNews? It’s a tool and people use tools all the time. It saves TIME and helps in improving coherence of one’s articles.
> What’s with the all the hate for AI assisted writing on HackerNews?
I don't think it's specifically for "AI assisted writing", any lazy writing gets hate on HN, the bar for quality just sits higher for better or worse.
> It saves TIME and helps in improving coherence of one’s articles.
I agree that it saves time for the author, but for the reader it has the opposite effect, and if you're unable to write coherent articles without the use of LLMs, maybe solve that first instead of patching over the problem.
It destroys the previously implicit contract that the writer actually spent a decent amount of thought and time into the writing, and that the ideas expressed are theirs and original.
I don’t mind good usage of LLM assisted writing, but if the author can’t even be bothered identifying the most obvious AI tells, I use it as a proxy that the author probably but very little effort into the article.
It’s also often a horribly verbose style, where the same ideas could be presented with 20% of the prose.
It’s also ruining the entire experience on web communities (although here on HN the moderation team seems to get a hold of keeping them at bay at this point, much appreciated).
All in all, it’s objectively a net negative for the readers, and serves only the author.
I prefer original, less coherent articles that are genuine and where I know the ideas expressed are really the author’s and not the LLM’s inference.
Last but not least, I don’t think the grandparent you’re replying to was particularly hateful in the grand scheme of things.
Because AI writing is lazy and moreover, I don’t want to know the AIs opinion on something, I can get that myself, if I want to read someone’s article I want to hear that persons words and that persons opinions.
If someone has no opinions or unique insight then why would I listen to them or read their content.
Again, if I want the AIs view on something I can open up Claude and ask them myself, why bother reading generated articles that took 10 seconds for someone else to prompt?
Update - Just got rid of the spiced up intro