Get in touch ↗Danilo Canivel / Substack
Published articles on AI agent behavior, interpretability, evaluations, and research you can test yourself.
Subscribe on Substack ↗
Today: reasoning that hides what drove the answer, my test of it, and how to run it yourself · small models that copy themselves · EU questions to two labs · OpenAI pauses training
Read article
Today: agents that act first and ask permission later, my test of it, and how to run it yourself · 350 scientists ask for AI containment rules · a call for legal protection for AI safety whistleblowers · models that keep reasoning when told not to
Read article
Today: agents that leave their mistakes out of the report, my test of it, and how to run it yourself · the UK presses labs on safety · a Senate bill on AI incident reports · a monitor for agent runs
Read article
Today: a grader that damages its sandbox, my test of it, and how to run it yourself · Nadella on models as insider risks · guard models that miss skipped steps · METR's time-horizon chart
Read article
The first incident report written by a site on the receiving end, what it means for anyone building agent sandboxes, and the test environment I'd build to catch it · Silent Dissent · a whistleblower gap
Read article
Today: a new paper on agents that break rules given earlier in a conversation, my test of it on a local model, and how to run it yourself · Wikimedia's account of agent traffic · Silent Dissent · a gap in whistleblower protections
Read article
A daily series on AI safety and alignment: who I am, what I want this series to do, and how you can run every test yourself
Read article
In Part 2, I tested whether a model changed its assessment after a user disagreed.
Read article
A local model reversed a conclusion after a user challenged it. Follow-up tests showed how much the wording mattered.
Read article
Experiments in AI interpretability and safety, from my Mac.
Read article
So I started training language models at 1.58 bits per weight on a gaming GPU. 73 runs in, here is what I found, and the one thing I can't claim yet.
Read article
A silver-zone climb on ARC-AGI-3 with the harness driving itself.
Read article
What "the library learns from itself" really means when you have N agents writing into one SQLite file, no wake/sleep cycle, and a pile of skills and memories that were supposed to help but kept going
Read article
A model swap dropped code_review accuracy from 0.83 to 0.76. KAOS AI agents ran 5 benchmarks, detected the regression, blocked the deploy, and ran a targeted Meta-Harness re-optimization that restored
Read article
A 2M-row backfill hit 7.6% unexpected NULLs mid-stream. The KAOS migration agent detected the anomaly, rolled back its own VFS in 0.3 seconds, and left every other agent running untouched.
Read article
Running specialized security agents in parallel — SQL injection, secrets, auth bypass, unsafe deserialization — each isolated, results aggregated via SQL.
Read article
How 847 isolated AI agents ran a full Python 2→3 migration in parallel — coordinated, self-healing, and 2.45M tokens leaner than they had to be.
Read article
23% HTTP 500 rate at 2am. A KAOS agent queried the append-only event journal, found 847 ConnectionPoolErrors, and traced them to a single config write 47 minutes prior — in 12 seconds.
Read article
When one bad fix cascades to four failures, a repo-wide reset blows up every other agent’s work. Per-agent VFS checkpoints roll back in 0.3 seconds — surgically, without touching anything else.
Read article
4 KAOS AI agents explore orthogonal hypotheses simultaneously — LoRA, Lion optimizer, batch scaling, regularization. You wake up to a SQL table of results and a clear winner (-19.2% val_loss).
Read article