Haz Sameen Shahgir

Our IllusionVQA paper has been accepted at @COLM_conf ! 🥳

@willccbb @brendanh0gan

i am no longer “that one morgan stanley guy who posts fun open-source grpo experiments” there are more of us

@osanseviero 6. mattshumer/Reflection-Llama-3.1-70B 😬

Which are your top 5 ML dramas?🍿 1. Llama and Zetta llama drama 2. What did Ilya see? 3. StabilityAI take-down of Runway Stable Diffusion 4. Hugging Face removal of GPT-4chan 5. Schmidhubering

@srush_nlp @NotFredd3 @jxmnop @jorje996 Bruh 😆

@NotFredd3 @jxmnop @jorje996 Yeah, what the heck! Go get a job.

another incredible thing about deepseek: all the american AI labs compete to hire the top PhD researchers - but deepseek didn’t compete deepseek researchers aren’t top PhDs. most are not even PhDs

@teortaxesTex @ericjang11 "After repeatedly changing his degree between different subjects like natural sciences, history of art, and philosophy, he eventually graduated with a BA degree in experimental psychology in 1970" - Wikipedia Checks out.

Dropout is motivated by the role of sex in evolution too Hinton seems to like these cross-domain inspirations

The opening sentence goes so hard. This paper was 10 years ahead of its time.

@danielhanchen @andrew_n_carr Something something condition number?

@andrew_n_carr Coincidentally I literally did an entire final year uni project for this :) Also why not inverse and why QR. Or why divide n conquer SVD is faster. Or if matrix is < 2000 cols use Cholesky via POTRF and SSYRK or do column pivoting etc And LSQR LSMR, sparse methods etc! Fun!!

Another great interview question! For linear regression, we can directly compute the minimizer as β= = (X^TX)^{-1}X^Ty. So why do we often use gradient descent instead?

@m2saxon Ah, makes sense. Pretty sad that it's a obfuscation system.

@sameen2080 I see, I hadn't been following closely enough to be aware that qwq (indisputably a model) is considered a reasoning model Regarding o1 though, I think that hiding the chain of thought tokens is a significant enough intervention on the raw outputs of the model to be a system

Can someone explain why o1 and ilk are described as "reasoning models" and not as "reasoning systems?" Isn't it a LM inside of a bigger structure?

@venturetwins They patched it?

ChatGPT refuses to say the name “David Mayer,” and no one knows why. If you try to get it to write the name, the chat immediately ends. People have attempted all sorts of things - ciphers, riddles, tricks - and nothing works.

@pranjalssh Excellent work. Couple of questions tho "...hence tensor core instructions require storing C over 128 threads in a SM" - Shouldn't it be 1024/256=4? "When we distribute C over a warp-group, each thread needs 1024/128 = 8 threads " - what does each thread needing 8 threads mean?

I implemented H100 cuda matmul kernel from scratch, taking inspiration from @Si_Boehm's blog. Our final kernel outperforms cuBLAS by 7% for N=4096. It fits in a single C++ file without any dependencies. Full-blown blog post with all details: cudaforfun.substack.com/p/outperformin…

@simonw Why not both?

After hassling Anthropic for months for a token counting library similar to OpenAI's tiktoken I just realized the Anthropic and Gemini approach of providing a free token counting API is actually better... because I don't know how to use tiktoken to count tools, images etc

@m2saxon Thanks for the shout out!

"IllusionVQA": Haz Sameen Shahgir, Khondker Salman Sayeed et al Testing VLM reasoning over optical illusion questions. For some *perceptual* illusions (same size, color) VLMs are superhuman, but for *logical* ones like "impossible shapes" they're worse. openreview.net/forum?id=7ysaJ…

@deliprao Seems like it's possible on Together AI (together.ai)

is there an llm finetuning service that will accept my data, train an open model (say llama3.2), and allow me to download the trained model?

@teknium Jailbreak it to reveal the CoT? 🤔

People probably don’t got enough questions to make it think as long as they expected (aka we’re all too dumb for it already)

Can’t tell if this is because there isn’t much demand or something else

@hu_yifei Yeah, current LLMs have really poor support for Bengali. NLLB was careful about this and upsampled Bengali. NLLB character/token for Bengali is 3.35 (higher the better) LLaMA-3, Qwen2, Mistral and Aya are all at about ~0.8. For reference English Char/Token is around ~4.5

Since I am working on multilingual stuff, I translated a piece of text from an academic paper to different languages. It seems like tokenizers are not friendly to certain languages despite they are among the most spoken languages in the world. People who speak Hindi or Bengali can confirm on this?

Presenting our paper in #cosplay at #ACL2024 🇹🇭, getting surprised and encouraged by the overwhelming interest we've received! Grateful for everyone’s support, especially the collaborators @LrzNeedResearch @JentseHuang. Paper Link🔗: incharacter.github.io

[N/N] This project was jointly lead by @MdTokiTahmid64 and yours truly. 😊 Check out our preprint: biorxiv.org/content/10.110… Try BiRNA-BERT yourself: huggingface.co/collections/bu… #RNAModeling #Bioinformatics #RNA

[8/N] 🔍 Finally, we perform **extensive** ablation studies that confirm that training on both BPE+Nucleotide tokens of each sequence is just as good as training two models separately with no performance loss.

[5/N] 📏 On long RNA sequences, BiRNA-BERT uses BPE to generate compressed sequence embeddings and can process RNA sequences 5 times longer than current SOTA RNA models with the same memory footprint.