Learning Transformer Programs We designed a modified Transformer that can be trained to solve a task and then automatically converted into a discrete, human-readable program. With Alex Wettig and Danqi Chen. Paper: arxiv.org/abs/2306.01128 Code: github.com/princeton-nlp/… [1/12]

Understanding the data-model interaction has become a very important topic for alignment and understanding LLM behavior. Happy to announce that we are organizing a workshop on attribution and accepting submissions.

Do Machine Learning Models Memorize or Generalize? pair.withgoogle.com/explorables/gr… An interactive introduction to grokking and mechanistic interpretability w/ Asma Ghandeharioun, nada hussein, Nithum, Martin Wattenberg and iislucas (Lucas Dixon)

Recent innovation has given rise to #ML models w/ impressive capabilities, but there’s much to learn about how we attribute model behavior to training data, algorithms, architecture, & more! Have papers or ideas on this? Submit to ATTRIB @ #NeurIPS2023 → attrib-workshop.cc

PAIR is looking for a Research Scientist interested in making hard ML problems (like understanding langauge) much smaller... In Paris, and working closely with fun interactive explorable visualizations too. See: goo.gle/3PcvPEs

Happy to share that this paper was accepted with a Spotlight at #NeurIPS2023! We updated the arXiv with results showing the disconnect between knowledge localization and editing success across different neuron ablations, editing methods, editing metrics, models, and datasets.⬇️

I’m very excited to share our work on Gemini today! Gemini is a family of multimodal models that demonstrate really strong capabilities across the image, audio, video, and text domains. Our most-capable model, Gemini Ultra, advances the state of the art in 30 of 32 benchmarks,

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

🧵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

Happy to introduce our paper MusicRL, the first music generation system finetuned with human preferences. Paper link: arxiv.org/abs/2402.04229

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: aitestkitchen.withgoogle.com/tools/music-fx… It’s an infinite AI jam that you control 🎛️. Try mixing your unique 🌀 of instruments, genres,

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 → goo.gle/4aS5epd

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 💙♊️

Join us live tomorrow at 2:30pm CET for some exciting updates on our research! youtube.com/live/hm2IJSKcY…

Can Large Language Models Explain Their Internal Mechanisms? pair.withgoogle.com/explorables/pa… An interactive intro to Patchscopes, an inspection framework for explaining the hidden representations of LLMs, with LLMs w/ Asma Ghandeharioun Ryan Mullins Emily Reif Jimbo Wilson Nithum iislucas (Lucas Dixon)

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 → goo.gle/patchscopes

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 → dpmd.ai/4d0MKEH

🧵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)