置顶推文
Goodfire
415 posts
Goodfire
@GoodfireAI
Using interpretability to understand, learn from, and design AI.
San Francisco 加入时间 Ağustos 2024
28 关注13.1K 粉丝
Goodfire 已转推
Come discuss autonomous science next Tuesday at the Academy of Sciences!
Excited to talk about what we're building @GoodfireAI and looking forward to hearing talks from @Ginkgo, Monomer Bio, and Tetsuwan!
Cristian Ponce@c_m_ponce
We rented out the Academy of Sciences next Tuesday for a series of talks on autonomous science with @Ginkgo, @GoodfireAI, and Monomer Bio! Come by! luma.com/rmuagyc0
English
New paper from Atticus Geiger and collaborators, led by @arunasank!
Aruna S@arunasank
Interpretability methods usually study single-token behavior. But real model behaviors, like sycophancy or writing style, are diffuse across many tokens. Can these diffuse behaviors be localized and controlled from long-form responses? YES!
English
Goodfire 已转推
In <2 years, @goodfireai has built an all-star team, landed their first blue-chip customers, published a ton of research, and now...
Raised $150M at a $1.25B valuation to become the first interpretability unicorn. 🦄
Here @DanJBalsam & @banburismus_ explain the motivation & vision for their new *Intentional Design" research agenda, which aims to understand and control what models learn in training.
Full episode ↓
English
Authors: @sidbop1, @annabelma314, @maxsloef, @RaphaelSarfati, @EricBigelow, Atticus Geiger, Owen Lewis, & @jack_merullo_
Paper: arxiv.org/abs/2603.05488
Read more in the blog post: goodfire.ai/research/reaso…
English
Interpretability methods can augment monitoring by identifying relevant information internally that isn’t reliably expressed.
Plus, well-calibrated probes give us a knob to control the tradeoff between accuracy and token usage, enabling adaptive inference-time compute! (6/7)
English
LLMs often reason “performatively” well after deciding on a final answer - something that CoT monitors are slow to catch.
Our new paper finds that:
- probes can help monitor for this
- it seems to track with task difficulty
- probes enable early CoT exit, saving tokens! (1/7)
English
Welcome @thomas_fel_! We're excited to have you at Goodfire.
Kempner Institute at Harvard University@KempnerInst
After 2 years probing #visionmodels at the #KempnerInstitute, @thomas_fel_ reflects on what he’s learned—and what pieces of the #interpretability puzzle remain hidden—as he heads to @GoodfireAI. Read the interview: bit.ly/4aEBzmp 🎙️🧩
English
Stanford students - check out this spring quarter course on interp that we're helping support:
Amir Zur@AmirZur2000
Excited to launch a course on mechanistic interpretability at Stanford next quarter! So grateful to be part of an amazing teaching team: Atticus Geiger, Jing Huang, Junyi Tao, and Thomas Icard (who all share the admirable trait of not having a twitter account)
English
Read more on how we interpreted Evo 2: goodfire.ai/research/inter…
Plus some follow-up work since then:
goodfire.ai/research/phylo…
English
Congrats to @pdhsu @BrianHie @davey_burke @garykbrixi and all our other coauthors at @arcinstitute @nvidia @Stanford @UCSF @UCBerkeley @UW!
English
Not every day nine of your teammates get published in Nature!
We've been working with Evo 2 since its release, and have found a number of exciting results with our interpretability tools - including discovering numerous biologically relevant features in the model.
Arc Institute@arcinstitute
Evo 2, the largest fully open biological AI model to date, is now published in @Nature.
English
Read the post: goodfire.ai/blog/interpret…
Join us: goodfire.ai/careers
English
Those optimizations let us harvest 3 billion activations from Kimi in an overnight run, at 14,000 tokens-per-second on an 8 GPU node.
That (+ other work) let us steer Kimi's chain of thought in real time! (5/6)
Ryan Panwar@RyanPanwar
Kimi is just like me frfr Steering @Kimi_Moonshot's K2 Thinking's reasoning in the Kimi CLI
English
