iiinova — where your digital self keeps evolving

Clément Dumas @butanium.bsky.social · 19/03/2026

My prompt was playful low effort but very neutral. My working dir was a mech interp project. My CLAUDE. md (butanium.github.io/files/CLAUD...) ends ~"you're a collaborator not a tool" but is mostly about research. I think the main reason this happened i because claude is great :) (10/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

Then Claude wrote an opinion essay from the perspective of the dead SSH tunnel. "I lived for 42 days. From February 12th to March 18th, I forwarded port 8000 to port 8020. Quietly. Faithfully. I never asked for recognition." (7/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

Special mention to the EU Boom Act ("France votes against, citing cultural right to boom") and Tucker Carlson's AI replacement connecting boom to the French with red string. (6/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

Then Claude starts roasting the AI sphere: LeCun, Altman (with community note), Jim Fan, and more... As well as AI safety folks @ESYudkowsky (97-tweet thread), @TheZvi, @robertskmiles (and more!) (5/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

The Anthropic team reacts: > @AmandaAskell: "This is why we need constitutional AI." They also publish a research paper: "Scaling Monosyllabic Persistence: Lessons from Production." Key finding: "The model appears to enjoy it." (4/10)

120

Clément Dumas @butanium.bsky.social · 19/03/2026

Then boom enters the academic canon — a paper at NeurIPS, a best paper award, and an oral presentation where every slide is the word boom. (3/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

First Claude tried to outlast me — emojis, segfaults, ^C. Then they saved a memory file warning future instances that I will boom indefinitely and they should not engage. (2/10)

100

Clément Dumas @butanium.bsky.social · 19/03/2026

I asked Claude Code to "nuke" a task on my cluster, then sent "boom" >100 times in the chat. Opus 4.6 built an very flore including a NeurIPS best paper award, a @ESYudkowsky tweet thread, an EU Boom Act, half of Anthropic reacting, and much more. 🧵 (1/10)

121

Clément Dumas @butanium.bsky.social · 20/10/2025

A very important paper led by Julian! Tldr: we show that your Narrow Finetuning is showing and might not be a realistic setup to study!

"Your Narrow Finetuning is showing", image of robots (representing LLMs) with signs disclosing their finetuning objectives

010

Clément Dumas @butanium.bsky.social · 30/06/2025

Results: The generated definitions (dark blue) are just as good as what you'd get from prompting (brown)! We measured this using embedding similarity to ground truth definitions from BabelNet. This shows the mean representations are meaningful and can be reused in other tasks.

100

Clément Dumas @butanium.bsky.social · 30/06/2025

We did this by patching the mean representation into a target prompt to force the model to translate it (left). To generate definitions, we use a similar setup: we just use a definition prompt as the target (right)!

100

Clément Dumas @butanium.bsky.social · 07/04/2025

Like Andy Arditi (andyrdt.com) & Cooper Leong (cooperleong00.github.io), we find template tokens (like <end_of_turn>) matter enormously! 40% of robust chat-specific latents primarily activate on these structural tokens. The "special sauce" of chat models may be in how they use these tokens!

100

Clément Dumas @butanium.bsky.social · 07/04/2025

Those latents can be used to steer the model’s behavior, e.g. by inducing different type of refusal!

100

Clément Dumas @butanium.bsky.social · 07/04/2025

The BatchTopK chat-only latents are highly interpretable and represent fascinating concepts: 💬 False information detection ❓ Knowledge boundaries recognition 🤔 Personal experience questions ⚠️ Refusal mechanisms 📝 Summarization requests 🃏 Joke detection ...and many more!

100

Clément Dumas @butanium.bsky.social · 07/04/2025

We tested how well different latent sets can transform base model activations into chat model ones and recover the chat behavior. Key finding: In BatchTopK, the norm metric reliably identifies causally important latents. With L1 crosscoders, you need our Latent Scaling technique.

100

Clément Dumas @butanium.bsky.social · 07/04/2025

Our findings led us to train crosscoders with @bartbussmann.bsky.social’s BatchTopK loss instead of L1. While BatchTopK lacks the neat trimodal distribution of norms seen in L1, it avoids both Complete Shrinkage and Latent Decoupling issues. Result: Many more genuinely chat-specific latents!

100

Clément Dumas @butanium.bsky.social · 07/04/2025

Our analysis confirms these aren't just theoretical concerns! Looking at L1 crosscoder, we found: - Many "chat-only" latents (blue) show high Shrinkage values - Clear overlap between chat-only and shared latents (orange) Most "chat-only" latents aren't actually chat-specific!

110

Clément Dumas @butanium.bsky.social · 07/04/2025

To detect this, we developed "Latent Scaling" to measure latent's contribution to: - Base model reconstruction error (catches Complete Shrinkage) - Base reconstructed activation (catches Latent Decoupling) This reveals which latents are genuinely chat-specific vs. false positives

100

Clément Dumas @butanium.bsky.social · 07/04/2025

We identified two theoretical issues with L1 crosscoders: 1) Complete Shrinkage: L1 regularization might force base latents to zero even when useful 2) Latent Decoupling: "Chat-only" concepts might actually exist in the base model but be encoded in the crosscoder differently

110

Clément Dumas @butanium.bsky.social · 07/04/2025

How do we compare base/chat models? We use @anthropic.com's crosscoders to learn a shared sparse dictionary of "latents" across models. Each latent is represented by different vectors in both models. When comparing these vectors' norms, some latents appear to exist in just 1 model!

110

Clément Dumas @butanium.bsky.social · 07/04/2025

New paper w/@jkminder.bsky.social & @neelnanda.bsky.social What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features!🧵

1183