Ching Fang (chingfang.bsky.social)

We achieved state-of-the-art performance in predicting which of 4.2 million genetic variants cause diseases by interpreting a genomics model, in a new preprint with @MayoClinic. We're now releasing an open source database for all variants in the NIH's clinvar database. 🧵(1/8)

We used interpretability to scale RL against open-ended tasks, cutting Gemma 12B’s hallucination rate in half by teaching it to self-correct in tandem with our probing harness.

We raised a $150M Series B at a $1.25B valuation to fundamentally change the field of AI. Scaling is powerful, but we can't intentionally design what we don't understand.

🤖📊 NEW in the Deeper Learning blog: @AnnHuang42 & @KanakaRajanPhD break down their recent work examining how recurrent neural networks solve the same task in different ways —and why that matters. Joint work with @tweetsatpreet & @FlaviohMar bit.ly/4kj4fVd #NeuroAI #AI

@jwuphysics @GoodfireAI @PrimaMente We flow through the original embeddings and classifier. The SAE is inserted as an intermediate layer to compute feature attributions, but the classification model itself operates on the original activations (via a detached residual).

@GoodfireAI @PrimaMente Really fantastic work; love your team's approach to solving real question via interp. Quick question: when performing grad attribution on SAE features, do you flow through SAE reconstructions or original embeddings (i.e. train hierarchical classifier on SAE decoded embeds)?

We've identified a novel class of biomarkers for Alzheimer's detection - using interpretability - with @PrimaMente. How we did it, and how interpretability can power scientific discovery in the age of digital biology: (1/6)

A really exciting example of using interp to advance scientific progress! I had a lot of fun working on this with our scientific discovery team :)

Kimi is just like me frfr Steering @Kimi_Moonshot's K2 Thinking's reasoning in the Kimi CLI

Our infra lets us steer trillion-parameter frontier models in real time: - live, mid-CoT edits to internal activations - directly altering how the model reasons (not just outputs) - stackable edits - no added latency We can make models more Gen Z, more concise, etc.

🔴 Live: Panel discussion #1: @ MAIN2025 The future of Neuroscience - The role of AI ?  with Andreas Tolias, Siva Reddy, Joao Sacramento, Ching Fang + Eva Portelance Moderated by Patrick Mineault @patrickmineault #MAIN2025

Had a great time at MAIN 2025! Loved chatting with everyone 😊

I'll be presenting this as a spotlight work at the #NeurIPS2025 mechanistic interpretability workshop this Sunday in the afternoon session!

But overall this is encouraging for interpretability - simple mechanistic tools may be more robust to encoded reasoning than we expected. Joint work with Sam Marks @saprmarks, done during my fellowship with Cambridge Boston Alignment Initiative @cbai_ai

One caveat is that we used SFT on the base model's own responses, which might preserve more of the original activation space. There's room for more testbeds here - RL-induced encoding, more complex ciphers, etc.

Sharing some new work! We asked whether we could decode a model's reasoning if it was thinking in encrypted text, using only its internal activations. The answer is surprisingly yes. arxiv.org/abs/2512.01222

We're launching an "AI psychiatry" team as part of interpretability efforts at Anthropic!  We'll be researching phenomena like model personas, motivations, and situational awareness, and how they lead to spooky/unhinged behaviors. We're hiring - join us! job-boards.greenhouse.io/anthropic/jobs…