🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Once we solve math and the world doesn’t react, we’ll realize that most problems are empirically bound, not intelligence bound. All the clusters built in the 2020s will then migrate from learning to simulation.

Padding a transformer’s input with blank tokens (...) is a simple form of test-time compute. Can it increase the computational power of LLMs? 👀 New work with Ashish Sabharwal addresses this with *exact characterizations* of the expressive power of transformers with padding 🧵

Kevin Zhu incredibly important question and i still don't think there's a good go-to citation for this, but intuition is that RL on-policy update sparsity + CoT enables generalization rather than just memorizing the data as in SFT arxiv.org/abs/2501.17161 x.com/saagnikkk/stat…

Half the projects for my undergrad Intro to Deep (LLM) RL pearls-lab.github.io/intro-deep-rl-… are exactly this. They came up with some pretty interesting stuff

half the ppl pinging me about building novel RL envs want to RL web agents sick release from the bros, now u can just use hud sdk to build, train, and eval on web RL envs 🔥🔥🔥

capabilities-pilled interp drop: we now know why transformers outperform SSMs 👀 (tldr don’t make your mamba conv kernel size too small, and if you’re using non-convolutional SSMs, it’s joever buddy)

yo @ the doomers: here's a great example of a less intelligent entity exercising control over a more intelligent one and by that i mean the woke right somehow convincing techno-libertarians to coalition with them

>50k+ to 500k+ views on X Is this the beginning of the end for conferences and citations?

getchu a gpu dealer that’s scalingpilled AND sanskritpilled thanks for the free ocr data btw 🤤

word on the street is that Justus Mattern is forfeiting the opensource RL championship fight

Could you use sparse RL gradients to identify & understand circuits that are relevant for a behavior from an interp pov? cc Aryaman Arora

deeptechmaxxing

white professor living in chicago presenting to a chinese audience in shanghai about indian ideas from the 7th century in relation to 21st century developments in ML i love globalism