🔥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

I always found it puzzling how language models learn so much from next-token prediction, while video models learn so little from next frame prediction. Maybe it's because LLMs are actually brain scanners in disguise. Idle musings in my new blog post: sergeylevine.substack.com/p/language-mod…

Does anyone want to volunteer as a maintainer of the ColBERT library? We don't necessarily need to make any changes, though some features are often requested! We do need someone to upgrade dependencies like pytorch and transformers, because old stuff rots basically. Any takers?

Diyi Yang (Diyi Yang) speaking at VLMs4All - CVPR 2025 Workshop workshop about the limitations of LLMs wrt cultural awareness.

I have bittersweet news to share. Yesterday we merged a PR deprecating TensorFlow and Flax support in transformers. Going forward, we're focusing all our efforts on PyTorch to remove a lot of the bloating in the transformers library. Expect a simpler toolkit, across the board.

happy 8th birthday of maybe the most influential paper of all time 🥳

underrated idea to learn passively about people from everyday computer use - I think the natural extension is learning from *trajectories* of how people prefer to do things, which is hard to get from prompting / static user data otherwise

a comprehensive study on AI job automation! lots of interesting nuggets

Cool work profiling a bunch of tasks that US laborers currently perform and to what extent (a) those laborers would like AI to automate those tasks and (b) AI experts believe AI could currently automate those tasks. I would guess that AI experts are pretty bad at

General User Models by Omar Shaikh might just be the coolest thing ever. SO MUCH looking forward to messing w the MCP

Some tasks are painful to do. But some are fulfilling and fun. How do they line up with the tasks that AI agents are set to automate? Not that well, based on our new paper "Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce"

The Hugging Face team recently announced it's going all in on PyTorch 🔥 "We have seen our user base consolidate in PyTorch," says Lysandre, Chief Open Source Officer at Hugging Face. "Going forward, we're focusing all our efforts on PyTorch to remove a lot of the bloating

By surveying workers and AI experts, this paper gets at a key issue: there is both overlap and substantial mismatches between what workers want AI to do & what AI is likely to do. AI is going to change work. It is critical that we take an active role in shaping how it plays out.

Language as a Visual Format talk by Phillip Isola. (These slides are a summary of the intro. The main content is about cycle consistency between images and captions improving text-to-image and image-to-text. Uses DPO! See quoted thread. 🧵👇)

Fascinating new AI agent research from Erik Brynjolfsson + colleagues: Looking at work tasks that are painful to do vs ones that are fulfilling and fun. How do they line up with the tasks that most AI agents are soon poised to automate? It identifies a key new opportunity quadrant.

スタンフォード大学研究者らの報告によると、労働者の多くは「AIに仕事を手伝ってほしい」と思っているが、完全に仕事を奪われたい訳ではないとのこと。 そして労働者が「これはAIにやってもらいたい」と思っている仕事とAI技術者が「これはAIで自動化できる」と思っている仕事は全然一致していない

It's really painful to live a few years ahead of the trend. By the time the trend catches up, you've realized a deeper concept and moved on. At recent AI events, I was stopped so many times and asked ``can I do multi-step/"multi-agent" stuff in DSPy''. What?! DSPy was started

Highly relevant! "Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce Yijia Shao" by Humishka Zope, Yucheng Jiang, Jiaxin Pei, David Nguyen, Erik Brynjolfsson, and Diyi Yang. arxiv.org/pdf/2506.06576

A must read > DSPy doesn’t just reduce the complexity of our codebase, it reduces the complexity of working with LLMs by decoupling our task from the LLMs > Even before optimization, DSPy gets us up and running faster while producing more maintainable code that can grow with us

41% of YC AI startups are solving tasks workers don't need automated New Stanford study shows workers actually DO want AI, but for repetitive work that frees them up for higher value tasks Startups are chasing full automation where partnership would work better