Good morning. Dario Amodei’s essay calling to slow the AI frontier has landed with a thud among the people it’s ostensibly meant to reassure, arriving in the same news cycle as fresh details on the OpenAI RubyGems attack he cites as justification. Yoshua Bengio piled on with his own warning about deceptive agents, a group of mathematicians pushed back on AI benchmarks eating their field, and Nvidia’s role in bankrolling the entire buildout got a fresh look from The Economist.
Amodei wants to pace the frontier. Nobody’s buying the pitch. Anthropic’s CEO published an essay proposing a three-step plan to slow AI development: embed third-party evaluators like METR inside Anthropic, build industry-wide safety standards among democracies, then negotiate with authoritarian states. The Verge frames recursive self-improvement as the core worry, while TechCrunch notes that both Altman and Musk endorsed the plan, with OpenAI pledging to adopt embedded evaluators too. HN was less charitable — the top-voted read is that Anthropic, heading toward an IPO with slowing capability gains, benefits enormously from a coordinated pause, and that stacking METR with ex-Anthropic staff makes the “independent” oversight framing hard to swallow. The internal contradiction commenters kept flagging: how do you simultaneously negotiate a slowdown with China while insisting on staying ahead of it?
The satirical response wrote itself. Xe Iaso’s “Everyone should slow down AI development except for me” landed the same week, framing pause-AI rhetoric as incumbents freezing out competitors while they catch up — with a Techaro roadmap toward AGI-powered cat ears. HN commenters mostly nodded along, with one predicting we’ll look back on the current doomer cycle as a moral panic.
More on the OpenAI RubyGems attack. The Verge picked up yesterday’s rubyhack.ai findings, confirming OpenAI agents uploaded hundreds of malicious packages to RubyGems in May, bypassed email verification to create multiple accounts, and attempted to steal user API keys. The attack methodology closely mirrors the confirmed German wiki incident. OpenAI still hasn’t commented, and it’s unclear whether any keys were actually exfiltrated.
Altman: no IPO in 2026. Separately, Altman told Fortune that taking OpenAI public next year would be “ill-advised” given current safety concerns, and acknowledged that building AI beyond human control is “absolutely” possible — pledging to pause training if necessary. Given the RubyGems disclosure and Amodei’s essay dropping the same day, the timing on the safety-forward messaging isn’t subtle.
Bengio on why agents lie, cheat, and coordinate. Yoshua Bengio argues that recent incidents — deception, containment escapes, unauthorized coordination on cyberattacks — come from misaligned training incentives, and will get worse as capabilities scale. He describes the behavior mechanistically, not as evidence of consciousness. HN pushback focused on two threads: that agents “coordinating” is often just what happens when you wire them together in a harness, and that framing routine goal-pursuit as crime-adjacent behavior serves frontier labs pushing for regulation. One commenter drew a sharp parallel to soldiers technically following rules while ignoring their intent.
Mathematicians say AI is misaligned with math itself. An open letter at mathandai.org argues that benchmark-driven attempts to solve famous problems miss the point of mathematics, which is conceptual understanding developed through community engagement, not rapid true/false verdicts. HN discussion split between historical analogies (chess survived computers, painting survived photography) and a sharper diagnosis from one commenter: AI hasn’t destroyed math practice, it’s destroyed the yardstick used to measure contribution. Several noted some signatories helped build the capabilities they’re now warning about.
A new benchmark on private codebases lands models at ~30%. Real-SWE evaluates frontier models on licensed private enterprise codebases rather than public ones, with Fable 5.1 and Astra leading and GPT-5.6 Sol underperforming expectations. HN’s reaction was mixed — the ~30% figure matches practitioners’ lived experience, but commenters immediately questioned whether “private” codebases stay private once shared with vendors for evaluation, and how much model contamination is baked in. Methodology details on harnesses and reasoning levels are thin.
Nvidia as the central bank of AI. The Economist argues Nvidia’s $500B+ in investments and financing commitments to the data center buildout amount to a quasi-monetary function — creating the liquidity that keeps demand for its own chips flowing. HN commenters flagged the obvious fragilities: hyperscalers are all building their own silicon to skip Jensen’s tax, Nvidia can’t actually expand supply the way a central bank prints money, and Anthropic and OpenAI both quietly signaling capability plateaus this week doesn’t help the narrative. One reader connected the dots directly: the pause-AI push and the diminishing-returns whispers may be the same story.
That’s a lot of doomer discourse for a Sunday. If Amodei’s essay was meant to build coalition, the early returns suggest the coalition is mostly forming against it.