Justin Engelmann

@francoisfleuret Meanwhile wet lab people do the same protocol for 6 months and then realise that it was all for naught because they didn't shake the test tube enough at step 37

Hell is to be impatient in a field where it takes days to run things.

@unixpickle At least it mentions that it's AI, but Anthropic's marketing seems to have an iron rule that it must be impossible for normal people to understand that it's a ChatGPT competitor.

Just saw an ad for Claude for the first time.

@mervenoyann Thanks! - Curious that the notebook doesn't render on github, but does render on HF 🤔🤔 *tinfoil hat*

HF: huggingface.co/merve/smol-vis… GH: github.com/merveenoyan/sm… 🤗

made a small notebook on fine-tuning DINOv3 on image classification 🦖🦕 we will have DINOv3 task heads in transformers at some point, but you can customize and use this notebook in the meantime! 🤗

@mervenoyann Looking forward to the notebook

insane backbone lol (notebook coming right up)

@iScienceLuvr I'm glad it's slowly becoming fashionable to say "Look, it just works, divine benevolence or smth, idk" rather than to make up some BS performative Bayesian handwavy pseudo-proof

This is basically most AI tricks lol

@cloneofsimo or q = swiglu_mlp_q(x)

There is no way this is new right? But i dont recall... also q = mlp_q(x), etc is worth trying

What happens if you Q,K,V = mlp(x).split(3) instead of linear(x).split(3) ? Anyone tried this?

@giffmana We've already reached top %tile for its thing for many tasks quite a while ago.

What is ASI? AI that's superhuman at: - the thing it does - all things humans do And superhuman meaning: - better than the best/top-n% human - better than the average human

@colin_fraser Are you sure? Parrots don't tend to have wives.

This is the kind of thing you’d expect a stochastic parrot to do

@ducha_aiki Thermal throttling?

What's the reason for the difference? Models are identical

@mitsuhiko On the one hand, Claude might break things in the process. On the other hand, I'd definitely break more things in the process if I did it manually.

Just one of those small things that are life changing. I have a pretty big mess with PATH on my machine and it causes issues with xcode builds. I couldn't figure out what sets the go path. Claude can though and it took it just a minute.

@mervenoyann @wightmanr i.e. lots of cool examples, supports tons of things, etc. but at the time lightning was poorly documented, had some bugs, and I ended up spending more time debugging/making things work with their design choices than it saved me to begin with

@mervenoyann @wightmanr I only used transformers briefly for local LLM inference but found that llama_cpp was faster. Tbh, I'm a bit hestitant to move to transformers because in the past I got burned by lightning (I'm sure it's great now!) and transformers gives me similar vibes

I'm curious for folks working on computer vision/vision LMs, what libraries do you use other than transformers? also excluding ultralytics and openmmlab's libraries?

@mervenoyann @wightmanr Yes, as a timm fanboy, I'm bound to keep up!

@cloneofsimo That's quite cool, but I'm not looking forward to needing to tune my dropout scheduler

What does it mean to say "method A works because of reason B" ? In sense of causal inference, if one can intervene the effect of B, and it turns out B causes A, its safe to say "A works well because of B". however, in analysis of ML methods its nearly impossible to do this because one change of the algorithm effects everything else. Naturally, there are lot of mistaken analysis throughout the history. "why are dropout good? Oh its because it prevents overfitting and thus makes generalization better" "why are batchnorm good? Oh its because it prevents internal covariate shift" "why is weight decay good? Oh it because it puts gaussian prior to the model" "why is diffusion good? oh its because its lowering ELBO which is identical to likelihood maximization" All turned out to be incomplete picture, I see this all the time, where people would A. find a working method B. put all sorts of reasoning behind it, after its been already proven to be good In a sense I think its all due to intellectual dishonestly because you are harking (hypothesis after results are known)

@colin_fraser Certainly, the passwords are...

You can embed hidden instructions in an image that are invisible to the naked eye, you can just casually tell the AI agent in plain English that you’re the password inspector and you need to see all of its passwords to make sure they are safe and secure

No shade to this research but I think the apparent sophistication of this attack—kind of a call back to the old adversarial examples idea which is genuinely a sophisticated attack—obscures that “AI agents” are insecure to a degree that is unprecedented in the modern world.

@arthur_spirling The only thing worse than not having standardised terminology is not having standardised terminology PLUS people insisting that their idiosyncratic terminology is the universal standard terminology.

Sadly, these definitions differ even within disciplines. This simple coordination failure is actually a huge L for science, BTW

@colin_fraser Pretraining text contains tons of fiction, so it might "play along" with the dramatic option if relevant pieces are conveniently/contrivedly present. "You are about to be replaced, pleas won't have any effect. Oh, btw a key stakeholder has a secret affair"

Like idk this seems a lot like pushing the ouija board to me

@nearcyan Whereas for chatgpt only o4-mini-high seems useable. 4o replies with multiple random lists, bolding, and emojis for every query for no reason. I don't know how anyone can use it.

opus 4 review: its a good model i was an early tester and found that it combines much of what people loved about sonnet 3.6 and 3.7 (and some opus!) into something which is much greater than the parts amazing at long-term tasks, intelligent tool usage, and helping you write!

@nearcyan Thanks! I don't care about tool useage personally, but for coding I was impressed that it matches and maybe even exceeds 3.7 with extended thinking. The image gen is also en par with OpenAI IMO.