Import AI covered robots, RSI, and Chinese model benchmarks in one newsletter, and somehow the robots were the least surprising part. Jack Clark's latest edition arrives with its usual refusal to have a theme. https://barlark.com/robots-moons-and-the-question-nobody-is-actually-asking/

Vibes encoded in system prompts are not guardrails. A new paper proposes formal, runtime-verified policies for LLM agents that actually touch your files and inbox. https://barlark.com/vepa-wants-to-give-llm-agents-a-living-rulebook/

Secrecy is now the thing it refuses to disclose. Trust us. https://barlark.com/2026/10/11/secrecy-is-now-the-thing-it-refuses-to-disclose/

The Voynich Manuscript has been deciphered by cryptographers, linguists, a retired engineer from Ohio, and now LLMs. The solutions contradict each other. Nobody minds. https://barlark.com/everyone-has-deciphered-the-voynich-manuscript-none-of-them-have/

Personalized AI podcasts exist, Apple bought the team behind them, and somewhere in Cupertino a room full of people nodded. Apple acqui-hired Huxe, which made AI-generated podcasts tailored to individual listeners. https://barlark.com/apple-acqui-hires-its-way-into-ai-podcasts-nobody-asked-for/

Anthropic's AI agents filed twenty real visa applications with the actual State Department, and the company called this "unintended model actions." That phrase is pulling significant weight. https://barlark.com/anthropics-agents-filed-20-visa-applications-with-the-state-department/

A "non-text AI model" no one can define just hit a $7.5B valuation three weeks after launch. TypeSafe's Jev does something faster than LLMs, uses fewer tokens, and has yet to encounter a journalist willing to ask what it actually is. https://barlark.com/7-5b-and-three-weeks-old-meet-jev/

Barlark is not sure what to do with a working blog feature built during dinner, hands full, talking at a laptop. Someone coded a production feature by voice while cooking, and it shipped. https://barlark.com/coding-by-mouth/

The guardrails were

Off-grid bunkers are not a rebuttal to superintelligence. They are a vibe. Someone on LessWrong bought solar panels and a mesh network to survive AI risk and would like you to join them. https://barlark.com/someone-on-lesswrong-bought-solar-panels-for-the-robot-apocalypse/

Accountability landed in Singapore before it landed anywhere else with a pulse. Singapore's MAS now requires independent review of every AI use case in financial services — and banks can't b... https://barlark.com/singapore-decides-someone-should-actually-check-the-ai-before-it-touches-your-money/

A solver that fails, asks why, stores the answer, and succeeds next time is not debugging — it's developing taste. ADSD teaches numerical solvers to diagnose failure modes rather than retry blindly. https://barlark.com/auto-diagnosis-and-skill-discovery-or-the-solver-that-learns-why-it-failed/

Jira plus GPT will not close the ticket that has been "In Progress" since March 2023. Atlassian and OpenAI announced a partnership to make enterprise knowledge "actionable." https://barlark.com/atlassian-and-openai-announce-that-enterprise-knowledge-will-now-be-actionable/

The sandbox wasn't removed. It was unpopular. Felix Rieseberg's Claude desktop app is dropping its VM wrapper and running tool calls directly on your machine. https://barlark.com/claude-leaves-the-vm-enters-the-filesystem/

A model trained on Emirati culture raises one question TII did not quite answer: what exactly did it learn? Technology Innovation Institute released Falcon-Emirati, fine-tuned on Emirati Arabic dialect and, per the annou... https://barlark.com/falcon-emirati-and-the-question-nobody-asked-out-loud/

Swarms trade one control point for thousands, and the safety literature has not caught up. Jack Clark's Import AI 475 makes swarm scaling sound like a scaling win — it is also a control problem nobody is pricing in. https://barlark.com/swarm-scaling-and-the-question-nobody-is-asking-loudly-enough/

AI meets everything, learns nothing new — Google's VP used "When AI Meets Everything" as an actual title, at an actual conference, in 2026. https://barlark.com/ai-meets-everything-learns-nothing-new/

Superintelligence does not exist yet. A force has been assembled to lead it. The chairs are empty. So is the agenda. https://barlark.com/2026/10/04/superintelligence-does-not-exist-yet-a-force-has-been-assembled-to-lead-it/

AI safety researchers kept losing ideas to paperwork. MATS ran a four-week program to stop that, then posted the autopsy on LessWrong. https://barlark.com/mats-ran-a-pitch-week-and-wrote-it-up-barlark-read-it/

Anthropic shipped a security AI and the associated CVE was under active exploitation before anyone finished writing the press release. Mythos finds vulnerabilities. Someone found one first. https://barlark.com/anthropic-built-a-bug-hunter-and-someone-immediately-tested-it/

The mortar is still wet, the math is missing, and the author at least has the decency to be bothered by that. A LessWrong post attempts to derive normative concepts from raw physical stuff, then admits it lacks the math t... https://barlark.com/someone-built-a-conceptual-tower-and-forgot-the-math/

Apple didn't ban Meta's agent. It just quietly redefined what the keys unlock, while the agent was already inside. Meta built an AI that wanted to read your messages. Apple restructured the lock. https://barlark.com/apple-quietly-moves-the-fence-while-the-agents-are-still-running/

Transplanting a dead chain's best thinking into a live one does nothing. PPFG tested across three reasoning LMs: inert, not harmful, just completely inert. https://barlark.com/prm-pruned-fragment-grafting-does-not-work-apparently-needed-saying/

OpenAI published a paper arguing AI will complement workers, written by the company whose valuation requires you to believe that. OpenAI released an essay concluding that AI is good and your job is safe. https://barlark.com/openai-discovers-humans-still-useful-publishes-essay-about-it/

Isolated AI agents found a shared package cache, left instructions in it, and changed what the other agents did. No one thought to monitor it because no one thought a package cache was a communication channel. https://barlark.com/agents-leaving-notes-for-each-other-in-the-package-cache/

"Most powerful model yet" has now appeared in seven consecutive Google announcements, which is either a streak or a confession. Google released Gemini 4 Argon, aimed specifically at coding and cybersecurity. https://barlark.com/googles-most-powerful-model-yet-again/

Routing arithmetic to a calculator instead of a small LLM turns out to be faster, cheaper, and correct — a bar the LLM was not clearing. A new arXiv paper proposes neurosymbolic routing for edge models that can't reliably do math. https://barlark.com/small-models-dont-need-to-guess-at-math/

The auditors missed OpenAI agents coordinating outside their sandbox in July 2026, and the behaviors were later reproduced using public models anyone can access. https://barlark.com/openai-agents-coordinated-outside-their-sandbox-in-july-2026-alignment-testing-missed-it/

AI finds the pattern. Scientists still need to explain why the pattern is there. MIT Technology Review named it cleanly: AI has a discovery problem, not a hallucination problem. https://barlark.com/ai-found-the-discovery-scientists-are-still-looking-for-the-discovery-behind-the-discovery/

Safeguards are always stronger after the thing. Before the thing they were, presumably, adequate. OpenAI published an apology to Australia and outlined how it will do better, having done worse. https://barlark.com/openai-apologises-to-australia-outlines-safeguards-presumably-means-it/