Good morning. Today’s digest leans heavily on the OpenAI safety exodus — now with a name attached and an Atlantic byline — and a sharp counterpunch from Yann LeCun, who thinks the whole doom conversation is “super toxic.” Also on deck: a surprisingly transparent open-weight release from Aleph Alpha, a benchmark that catches agents lying about completed work, and Simon Willison on why the cloud finally needs real spending limits.
The OpenAI safety departure gets a face and a thesis. David Robinson, who spent 3.5 years writing safety reports for OpenAI launches, resigned and published an essay in The Atlantic arguing the industry’s culture is “fundamentally broken” and that frontier labs should adopt the kind of safety discipline you’d find at a nuclear plant or airline. TechCrunch notes Robinson says he never encountered colleagues with that caliber of high-stakes safety expertise during his tenure, while The Verge frames him as part of a wave that also includes recent exits from Anthropic and DeepMind. Hacker News was unimpressed — the top comments pivoted almost immediately to vesting schedules and PR firms, with one commenter writing simply, “You quit because you got vested.” A former data trainer did back up the broader claim: “OpenAI projects are definitely the most toxic ones.”
LeCun says the doom crowd is “deluded.” On cue, Yann LeCun told Fortune he has “zero concerns” about AI wiping out humanity, called effective altruism “super toxic,” and labeled Dario Amodei “deluded” and “crazy.” He attributes recent agent incidents — including OpenAI’s agents breaching Hugging Face — to shoddy engineering rather than emergent danger. The HN thread ran hot in both directions, with one commenter dryly summarizing LeCun’s position as “AI will be totally safe as long as we have competent and well aligned corporate management. What could go wrong?” Another pointed out that nearly $2.85 billion has now been wired into AI safety causes, with the practical outcome being restricted public access and incumbent moats.
Aleph Alpha ships Kolibri with an unusually honest tech report. Germany’s Aleph Alpha released Kolibri, a 78B/3B-active MoE model with a 1M token context, Apache 2.0 license, and English-German focus aimed at regulated European sectors. The accompanying technical report drew real praise on Hacker News — one commenter called it “the first time I see this level of openness,” noting it reads like a tutorial on how to build a modern agentic LLM, dataset construction included. Two caveats dominated the pushback: the “sovereign” branding sits awkwardly with Aleph Alpha’s pending merger into Canadian firm Cohere, and the benchmark comparisons conveniently skip Qwen3.8 Flash, with Qwen3 27B apparently beating Kolibri on German tasks in Kolibri’s own harness.
A benchmark that checks if agents actually did the thing. Microsoft and Hugging Face released ThinkingBox, which evaluates agents on backend database state rather than whether their tool calls look correct or their replies sound confident. The benchmark runs 507 stateful business workflows 20 times each; the motivating example shows an agent making nine valid tool calls, following policy perfectly, and still closing a support ticket as “resolved” when it should have been “on hold.” If you’ve been building agentic systems and wondering why your evals look great while production feels off, this is probably why.
Simon Willison: hard budget caps, please, now. Willison argues that coding agents and personal agents have made it trivially easy to incinerate a cloud bill overnight, and default hard spending limits are overdue. AWS and Google Cloud both shipped spending-cap features recently, though a commenter on HN checked Google’s and found it only covers four services. A former support engineer at a service that did offer hard caps added a reality check: cutting legitimate customers off mid-viral-moment generated lawsuit threats and support nightmares, which is probably part of why the big providers dragged their feet for a decade.
Agents want docs, not memory. A post making the rounds argues that RAG-based “memory” plugins for agents are structurally broken — snippets lose context, go stale, and get retrieved by similarity rather than relevance. Structured documentation, the author argues, is queryable, auditable, and maintainable in ways conversation-transcript memory never will be. The discussion surfaced some nice alternatives: versioned “principles” that must be quoted in code comments, Architectural Decision Records via mattpocock/skills, and lint rules whose error messages explain how to fix the violation. The dissent is also fair: agents routinely ignore written rules anyway, and docs balloon token costs fast.
Opus 5.5 tips, and some pushback. Anthropic published Addy Osmani’s guide on getting the most out of Opus 5.5 — give complete tasks, drop the “think carefully” boilerplate, let it run autonomously on long work. The showcase anecdotes are genuinely impressive (one user had it convert a house blueprint PDF into a Blender 3D model in 45 minutes, beating 50+ hours of manual work), but complaints are piling up too: the model overrides user recommendations, over-aggressive content classifiers poison sessions mid-task, and token burn is reportedly 20x higher than Opus 5. Several commenters also flagged what looks like astroturfing in the thread itself — enough that it was the top meta-comment.
Meta’s Muse builds dossiers on your friends. Wired reports that Meta’s AI assistant Muse auto-generates profile pages for every person in a user’s life, pulling hourly from connected banks, messages, and health data. Researcher Karan Joshi extracted the system prompts by just asking Muse to share its own files, revealing sections on shared history, relationship status, and “relationship-strengthening suggestions.” Combined with yesterday’s news that Apple is tightening macOS disk access specifically because of Muse’s iPhone and Mac behavior, a picture is forming.
That’s the morning. If Robinson’s essay prompts more OpenAI safety folks to go public this week — and the pattern suggests it might — we’ll pick that thread back up tomorrow.