🔥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've done a deep dive into SB 1047 over the last few weeks, and here's what you need to know: *Nobody* should be supporting this bill in its current state. It will *not* actually cover the largest models, nor will it actually protect open source. But it can be easily fixed!🧵

🥁🥁 New blog post out (link in thread), w/ two aims: 🤓 Providing a clear, hopefully easy-to-read intro to ColBERT, without assuming you've ever used it. 🏊Introducing ColBERT Token Pooling ✨: You can reduce the size of ColBERT indexes by 66% with barely any performance hit!

New Conference on Language Modeling paper with Noémie Elhadad! SPEER: Sentence-Level Planning of Long Clinical Summaries via Embedded Entity Retrieval arxiv.org/abs/2401.02369 *tl;dr* We introduce R^3 decoding--integrating RAG into planning--to improve faithfulness & coverage on long summarization!

Will present this work on testing the global opinions represented in language models at Conference on Language Modeling 🌍

Someone noticed our not-quite-launched new lib for WebGPU programming on GitHub and now it's on the front page of HN! It's created by Austin Huang and he'll be publishing a blog post about it very soon. But since it's out in the open now, here you go :D github.com/AnswerDotAI/gp…

Announcing: The initial release of my 1st project since joining the amazing team here at Answer.AI gpu.cpp Portable C++ GPU compute using WebGPU Links + info + a few demos below 👇

Can LLMs reason and solve other multi-step tasks? They sound like they can but they often fail wildly. I've written a post on Yejin Choi team's "Faith and Fate" paper, which provides a great intuition for this, arguing what models ARE doing is *linearized subgraph matching*.

Announcing FastHTML. A new way to create modern interactive web apps. Scales down to a 6-line python file; scales up to complex production apps. Auth, DBs, caching, styling, etc built-in & replaceable and extensible. 1-click deploy to Railway, Vercel, Hugging Face, & more.

Announcing Cold Compress 1.0 with Answer.AI A hackable toolkit for using and creating KV cache compression methods. Built on top of Horace He and Team’s GPT-Fast for torch.compilable, light-weight performance. Develop novel methods in as little as 1 line of new code.

We'll present this at #NeurIPS

My vision for the ultimate AGI interface is a blank canvas. The one that evolves, self-morphs over time with human preferences and invents novel ways of interacting with humans, redefining our relationship with AI technology and the entire Internet. But here are some of the

Announcing Surya Table Recognition! It uses a new architecture to outperform table transformer, the current SoTA open source model. - Recognizes table rows, columns, and cells - Works with complex layouts and rotated tables - Supports any language - Runs locally

Machines of Loving Grace: my essay on how AI could transform the world for the better darioamodei.com/machines-of-lo…

This gradient accumulation implementation bug doesn't affect all training frameworks. For example, Composer has the accumulate_train_batch_on_tokens option, which should prevent this issue. I would be surprised if other training frameworks didn't have similar options.

As a lead researcher at @stabilityai, I worked a lot with synthetic data to train LLMs and VLMs. It is the most underrated way of boosting model performance. Now at Answer.AI I've been working to make the best practices easy—read on to learn how!

The new GSM-Symbolic paper from Apple has been making waves, but we published very similar findings earlier this year. Using nearly the same symbolic template methodology on GSM8k problems, we demonstrated the reasoning limitations of LLMs. arxiv.org/pdf/2401.09395

Introducing an upgraded Claude 3.5 Sonnet, and a new model, Claude 3.5 Haiku. We’re also introducing a new capability in beta: computer use. Developers can now direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking, and typing text.

New Anthropic research: Evaluating feature steering. In May, we released Golden Gate Claude: an AI fixated on the Golden Gate Bridge due to our use of “feature steering”. We've now done a deeper study on the effects of feature steering. Read the post: anthropic.com/research/evalu…

Excited to share my new research on evaluating feature steering: I ran quantitative evaluations on how steering specific features affects model behavior. I identified a 'sweet spot' for maintaining capabilities, and found both targeted and off-target effects on social biases 🎯

🧙 I am recruiting PhD students and postdocs to work together on making sure AI Systems and Agents are built safe and respect privacy (+ other social values). Apply to UMass Amherst Manning College of Information & Computer Sciences and enjoy a beautiful town in Western Massachusetts. Reach out if you have questions!