Yash Malik

relaxing

[Yash Malik @_yash_malik_]

@giffmana Environment and harnesses is more of data scaling (in traditional sense) than compute

Unless ofc they meant that their environments are now more (computationally) complex worlds

[Sholto Douglas @_sholtodouglas]

It is extremely sad that this didn't end up as an example of how the labs could cooperate/coordinate, because the stakes will be so much higher in the future.

[jessicat @jessi_cata]

I asked Astra to help edit my math paper draft and it found a counterexample to a proposition in the literature, a proposition which I was planning to cite... I have now emailed the author... sign of the times

[Wyshing @wyshingforit]

gamble, but not with you shopping

coming soon

sign up for our waitlist: https://t.co/5Eo4x6JRQq

[Yash Malik @_yash_malik_]

@charliekerr @theprincesajjad @noah_pravecek @tempo I am very curious to understand the setup it builds here that can reliably watch 10k videos!

[Yash Malik @_yash_malik_]

@aditi17goel A very useful practice sheet, thank you!

[Aditi Goel @aditi17goel]

Wrote a short guide and Colab worksheet for anyone wanting to practice manual backpropagation. It breaks down the gradients step-by-step without relying on autograd, starting simple and building up to an MLP.
Check it out:
https://t.co/rqwlXlphrd

[Yash Malik @_yash_malik_]

@aditi17goel Lots of alpha here for people getting into ML!

Nice balance of intuition + actual details instead of the usual surface-level distilled content.

[Aditi Goel @aditi17goel]

Lately, I've been going deep into ML internals. If you're studying ML or just curious about how things work under the hood, a thread of my recent deep dives:

[jianlin.su @Jianlin_S]

An English mirror site of Scientific Spaces: https://t.co/da4YSTt9Er

Created by @GargRohin3301 . Thanks a lot!

[Yash Malik @_yash_malik_]

@finbarrtimbers Technically, it's allocated during the first optimizer.step(), not the second training step.
The first optimizer step just happens to be at the end of the first training (and beginning of the second) step.

[Ziniu Li @ZiniuLi]

🚀 Excited to share our work at Bytedance Seed!

Knapsack RL: Unlocking Exploration of LLMs via Budget Allocation 🎒

Exploration in LLM training is crucial but expensive.
Uniform rollout allocation is wasteful:
✅ Easy tasks → always solved → 0 gradient
❌ Hard tasks → always fail → 0 gradient

💡 Our idea: treat exploration as a knapsack problem → allocate rollouts where they matter most.

✨ Results:
🔼 +20–40% more non-zero gradients
🧮 Up to 93 rollouts for hard tasks (w/o extra compute)
📈 +2–4 avg points, +9 peak gains on math benchmarks
💰 ~2× cheaper than uniform allocation

📄 Paper: https://t.co/3HbOwV2tLL

[Yash Malik @_yash_malik_]

@navvye It's more of a behavior of the softmax function itself, right?

[Yash Malik @_yash_malik_]

@IJCAIconf Are the sessions recorded, and would they be shared?

[Aidan Gomez @aidangomez]

Cohere intends to acquire Perplexity immediately after their acquisitions of TikTok and Google Chrome.

We will continue to monitor the progress of those deals closely so we can submit our term sheet upon completion.

[Ganqu Cui @charlesfornlp]

Thoughts: While LLMs provide powerful priors for RL, many recent studies show that simply narrowing the model's output distribution can improve performance, but this also exhausts the model's potential for further exploration and improvement.

*Is it a blessing or a curse?* 🤔

[Yash Malik @_yash_malik_]

@charlesfornlp Very cool results! Thank you for sharing!
Do you have a preprint available somewhere?

[Taiwei Shi ✈ COLM @taiwei_shi]

Is there anything that Qwen cannot do at this point? 😂

[Yash Malik @_yash_malik_]

@kalomaze Utterly false.

[Yash Malik @_yash_malik_]

@activewarp @nrehiew_ Ahh great! Thank you very much, this makes it very clear!