Mapping LLMs with Sparse Autoencoders https://t.co/FIcY91YkS4 An interactive introduction to Sparse Autoencoders and their use cases with Nada Hussein, @shivamravalxai, Jimbo Wilson, Ari Alberich, @NeelNanda5, @iislucas, and @Nithum

AI is everywhere, but the most impactful features start with a real user need. Ask how AI can add unique value to that experience, not just how to release an AI feature. Here are three questions to consider before you start building. 1️⃣ What’s the core user problem? 2️⃣

my coworkers have been subjected to this yap so I now continue here: this is genuinely one of the most useful papers to engage with for interp imo. easily one of my most referred to papers. if you have not read it, pls do. mandatory reading until the heat death of the universe.

Veo holograms 🦝⚡️ Visualizing animal superpowers! Just discovered Veo 3's amazing ability to render 3d holograms. Virtual interfaces within the simulated world. 🔊 Prompts in 🧵

We are hiring Applied Interpretability researchers on the GDM Mech Interp Team!🧵 If interpretability is ever going to be useful, we need it to be applied at the frontier. Come work with @NeelNanda5, the @GoogleDeepMind AGI Safety team, and me: apply by 28th February as a

We scaled training data attribution (TDA) methods ~1000x to find influential pretraining examples for thousands of queries in an 8B-parameter LLM over the entire 160B-token C4 corpus! https://t.co/4mglIOAjyB

What a way to celebrate one year of incredible Gemini progress -- #1🥇across the board on overall ranking, as well as on hard prompts, coding, math, instruction following, and more, including with style control on. Thanks to the hard work of everyone in the Gemini team and

🚨 New Paper 🚨 Can LLMs perform latent multi-hop reasoning without exploiting shortcuts? We find the answer is yes – they can recall and compose facts not seen together in training or guessing the answer, but success greatly depends on the type of the bridge entity (80%+ for

I’m so proud of the updated version of #MusicFXDJ we developed in collaboration with @jacobcollier, available today at https://t.co/pYopej66CL. Over the past year I’ve spent countless hours experimenting with our real-time music models, and it feels like I’ve learned to play a

🧵Responses to adversarial queries can still remain latent in a safety-tuned model. Why are they revealed sometimes, but not others? And what are the mechanics of this latent misalignment? Does it matter *who* the user is? (1/n)

We’re welcoming a new 2 billion parameter model to the Gemma 2 family. 🛠️ It offers best-in-class performance for its size and can run efficiently on a wide range of hardware. Developers can get started with 2B today → https://t.co/hQRWYwGY7q

Can large language models (LLMs) explain their internal mechanisms? Check out the latest AI Explorable on Patchscopes, an inspection framework that uses LLMs to explain the hidden representations of LLMs. Learn more → https://t.co/mvmix9hKs0

Join us live tomorrow at 2:30pm CET for some exciting updates on our research!

Gemma 2 27B is now the best open model while being 2.5x smaller than alternatives! This validates the work done by the team and Gemini. This is just the beginning 💙♊️

Being able to interpret an #ML model’s hidden representations is key to understanding its behavior. Today we introduce Patchscopes, an approach that trains #LLMs to provide natural language explanations of their own hidden representations. Learn more → https://t.co/WfY1FYa1Wt

I love music most when it’s live, in the moment, and expressing something personal. This is why I’m psyched about the new “DJ mode” we developed for MusicFX: https://t.co/1Qk1VjnjEE It’s an infinite AI jam that you control 🎛️. Try mixing your unique 🌀 of instruments, genres,

Super excited for the Gemma model release, and with it a new debugging tool we built on 🔥LIT - use gradient-based salience to debug and refine complex LLM prompts!

Happy to introduce our paper MusicRL, the first music generation system finetuned with human preferences. Paper link:

🧵Can we “ask” an LLM to “translate” its own hidden representations into natural language? We propose 🩺Patchscopes, a new framework for decoding specific information from a representation by “patching” it into a separate inference pass, independently of its original context. 1/9

We often interpret neural nets by studying simplified representations (e.g. low-dim visualization). But how faithful are these simplifications to the original model? In our new preprint, we found some surprising "interpretability illusions"... 1/6