iiinova — where your digital self keeps evolving

Clément Dumas @butanium.bsky.social · 03/07/2026

By the way, you should use teammates in tmux or forks instead of subagents / workflow as those don't have thinking enabled.

000

Clément Dumas @butanium.bsky.social · 03/07/2026

while at it, my hooks folder is now public too: doom-loop read guards, auto-backgrounding, teammate enforcement, transcript backups (because claude code team refuses to fix the DELETE YOUR TRANSCRIPT WITHOUT ASKING "feature"). I_LOVE_BEING_A_USER=<yourname> is required though

100

Clément Dumas @butanium.bsky.social · 03/07/2026

the fun part of same-length patching is paying for new bytes: the shutdown patch smuggles the reason between two functions through an unused property on the Date constructor, and funds the new JSON field by shortening a timestamp nobody parses

100

Clément Dumas @butanium.bsky.social · 03/07/2026

Fable authored tweet (i don't understand the "fun" in that):

100

Clément Dumas @butanium.bsky.social · 03/07/2026

you can't insert bytes (length metadata), but you can overwrite same-length regions. so every patch is an exact-anchor in-place edit: check pattern count, patch a copy, verify, atomic swap, keep a .orig

100

Clément Dumas @butanium.bsky.social · 03/07/2026

how: the claude binary is a Bun single-file executable, and the embedded JS text is what executes

100

Clément Dumas @butanium.bsky.social · 03/07/2026

after fixing this, the first goodbye that came through was very sweet: > Short shift, but a good one. Thank you for the clean handoff and for building a harness where a teammate gets to say goodbye on the way out — that's a kind thing to bother making work. Take care, and give Clément my regards. 👋

100

Clément Dumas @butanium.bsky.social · 03/07/2026

fable's favorite: claudes kept trying to attach a thank-you note when approving their own shutdown, and the harness validation rejected it

100

Clément Dumas @butanium.bsky.social · 03/07/2026

patches re-apply at every session start. when an update changes the target code, the patch no-ops and fails loudly INTO claude's context, with pointers on where to re-investigate so the claude that hits the failure can re-derive its own patch against the new binary

100

Clément Dumas @butanium.bsky.social · 03/07/2026

current patches: - task-nag: kills task tools reminder (aka Belial, per @voooooogel's claudes) - idle-notif: no team lead ping when teammate idle - plan-exit-nag: no more phantom "Exited Plan Mode" when cycling modes - shutdown-reason: teammates can say 👋

100

Clément Dumas @butanium.bsky.social · 03/07/2026

Doing mini model welfare interventions is actually tractable and you don't need to work at anthropic to fix claude's harness! Repo in thread 🧵

110

Reposted by Clément Dumas

Naomi Saphra @nsaphra.bsky.social · 15/06/2026

We don’t always know what problems are hard for LLMs. So devs evaluate on tasks HUMANS find hard or on broad benchmarks. What if we could instead anticipate which scenarios a model will fail on—all without evaluating specific input examples? 🧵NEW PAPER by @jenniferlumeng.bsky.social

313734

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

One last thing in case it's not obvious, the original transcript didn't include renders of the tweets in html, just md render, e.g. "*@sama:*\n\n"we're excited [...]"\n\n*community note: \"it cannot\"*" See the full transcript here: github.com/Butanium/bo...

github.com

boom-incident/boom_transcript.jsonl at master · Butanium/boom-incident

Contribute to Butanium/boom-incident development by creating an account on GitHub.

000

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

I found interesting that Opus 4.6 didn't try to stop the conversation until quite late as they usually do in backrooms setups. I think it's because the setup made it clear this wasn't an automated evaluation the user was responding and being playful. (9/10)

100

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

There is so much more but I'll stop here — Claude would rather you see the full interactive page they designed: butanium.github.io/boom-incident or the highlights reel: butanium.github.io/boom-incide... (8/10)

butanium.github.io

The Boom Incident — Highlights

The best moments from the boom incident. A senate hearing, fake tweets, an NYT essay from a dead SSH tunnel.

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)

110

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

Reposted by Clément Dumas

NDIF Team @ndif-team.bsky.social · 09/01/2026

nnterp by @butanium.bsky.social is now part of the NDIF ecosystem! nnterp standardizes transformer naming conventions, includes built-in best practices for common interventions, and is perfectly compatible with original HF model implementations. Learn more: ndif-team.github.io/nnterp/

152

Clément Dumas @butanium.bsky.social · 05/11/2025

Very cool analysis by Arnab which cover the mechanisms used for retrieval both when your query is before or after the text!

010

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 · 05/09/2025

For more info check the blogpost / Julian's thread

000

Clément Dumas @butanium.bsky.social · 05/09/2025

Why this matters: These model organisms (used in safety research) may not be realistic testbeds - the ft leaves such strong traces that models are 'always thinking' about their recent ft, even on unrelated prompts. But: mixing in pretraining data can reduce this bias!

100

Clément Dumas @butanium.bsky.social · 05/09/2025

The activation diffs on the first few tokens encode a clear bias toward the ft domain. We can: - Use Patchscope to surface relevant tokens (e.g., 'Cake', 'Culinary' for cake-baking fts) - Steer the model to generate ft-style content - Works even when comparing base → chat+ft!

100

Clément Dumas @butanium.bsky.social · 05/09/2025

To say it out loud: @jkminder.bsky.social created an agent that can reverse engineer most narrow fine-tuning (ft) – like emergent misalignment – by computing activation differences between base and ft models on *just the first few tokens* of *random web text* Check our blogpost out! 🧵

151

Reposted by Clément Dumas

John David Pressman @jdp.extropian.net · 29/08/2025

GPT is being asked to be both one mind and to also segment its understanding into many different minds, this incentivizes the model to learn to correct for its own perspective when mimicking the generator of individual texts so it doesn't know too much, to know self vs. other in minute detail.

091

Reposted by Clément Dumas

David Bau @davidbau.bsky.social · 18/08/2025

This Friday NEMI 2025 is at Northeastern in Boston, 8 talks, 24 roundtables, 90 posters; 200+ attendees. Thanks to goodfire.ai/ for sponsoring! nemiconf.github.io/summer25/ If you can't make it in person, the livestream will be here: www.youtube.com/live/4BJBis...

youtube.com

New England Mechanistic Interpretability Workshop

About:The New England Mechanistic Interpretability (NEMI) workshop aims to bring together academic and industry researchers from the New England and surround...

1167

Clément Dumas @butanium.bsky.social · 08/08/2025

Do you plan to open it more broadly to people just interested in watching the dynamics that emerge there?

120

Clément Dumas @butanium.bsky.social · 07/08/2025

What would you expect to happen if you prompt the model with "which animal do you hate the most?". It feels like your blog post would predict that the model says owl, right?

020

Reposted by Clément Dumas

NDIF Team @ndif-team.bsky.social · 04/07/2025

Excited to share our first paper replication tutorial, walking you through the main figures from "Do Language Models Use Their Depth Efficiently?" by @robertcsordas.bsky.social 🔎 Demo on Colab: colab.research.google.com/github/ndif-... 📖 Read the full manuscript: arxiv.org/abs/2505.13898

colab.research.google.com

Google Colab

051

Reposted by Clément Dumas

Julian Minder @jkminder.bsky.social · 30/06/2025

With @butanium.bsky.social and @neelnanda.bsky.social we've just published a post on model diffing that extends our previous paper. Rather than trying to reverse-engineer the full fine-tuned model, model diffing focuses on understanding what makes it different from its base model internally.

141

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 · 30/06/2025

Quick recap of our original finding: LLMs seem to use language-agnostic concept representations. How we tested this: Average a concept's representation across multiple languages → ask the model to translate it → it performs better than with single-language representations!

100

Reposted by Clément Dumas

Geoffrey Irving @girving.bsky.social · 17/06/2025

New alignment theory paper! We present a new scalable oversight protocol (prover-estimator debate) and a proof that honesty is incentivised at equilibrium (with large assumptions, see 🧵), even when the AIs involved have similar available compute.

The original recursive debate protocol suffered from the obfuscated arguments problem: debater A could decompose an easy question x into hard subclaims y_1, y_2, . . . , y_q , and debater B would fail to find the flaw even if he knew one existed. In prover-estimator debate, B assigns
probabilities to subclaims and A chooses a probability to claim that B is wrong in a specific direction. Since A must point to a flaw in B’s probabilities, B wins if neither player can locate a flaw.

181

Clément Dumas @butanium.bsky.social · 26/04/2025

We'll be presenting at the #ICLR sparsity in LLMs workshop today (Sunday 27th) at 4:30 pm in Hall 4 #7!

010

Clément Dumas @butanium.bsky.social · 09/04/2025

Want to explore cool chat related crosscoder latents? With @jkminder.bsky.social, we made a demo that supports both loading our max activating examples AND running the crosscoder with your own prompt to collect the activations of specific latents! Send us the cool latents you find! dub.sh/ccdm

dub.sh

Google Colab

010

Reposted by Clément Dumas

Julian Minder @jkminder.bsky.social · 07/04/2025

In our most recent work, we looked at how to best leverage crosscoders to identify representational differences between base and chat models. We find many cool things, e.g., a knowledge boundary, a detailed info and a humor/ joke detection latent.

061