Today's issue reads like a status report: trust is the bottleneck, the frontier still runs on Hopper, and Anthropic is reminding itself to be good.
The chart was a map of distrust, not of capability.
Anthropic published a chart of what AI theoretically can do, by profession, set against what organisations actually deploy it for. Bayram, writing in Products & Startups, argues the blue-to-red gap is not a capability ceiling — it is algorithm aversion. We forgive humans their mistakes and refuse to forgive models theirs.
- One hallucination collapses confidence the way a hundred human errors do not.
- Users want agency, and "let the model decide" reads as the loss of it.
- Subjective tasks (partner, lawyer, brand voice) resist automation; objective ones (route, total) accept it.
- The question is not "can the agent?" but "will the agent act in my interests?"
- Survival anxiety beats learning anxiety; psychological safety is the cheaper lever.
Claude is reminded to be good — and is.
Anthropic's "widening the conversation" post slipped in a methods nugget. Researchers built a tool that, mid-task, pings Claude with its own ethical commitments. In alignment evals the model showed markedly lower rates of misaligned behaviour. They cannot yet say whether the win comes from the reminder, or from the pause to reflect.
"Markedly lower rates of misaligned behaviour. We are still working out whether the effect is the reminder, or the reflection itself."
Anthropic, widening-conversation-ai · 19 May 2026
Krea ships K-2.
Krea has opened K-2, the second generation of its image model, in early-access preview. Prompt adherence is tighter, micro-detail in faces and hands holds up at portrait crops, and seed reproducibility is finally honest. Activation codes circulating on Telegram unlock eighty free credits each, while supply lasts.
v0 opens what it makes.
v0's latest update hands the agent a browser. It opens the app it just generated, navigates the flows it claims to support, captures screenshots, and files bug reports against its own output. The build-test loop closes inside the editor with no human between turns.
v0> ship "checkout flow" → generated 14 files in 3.2s v0> browse --route /checkout → opening preview · capturing screenshots → /checkout OK (200ms) → /checkout/pay OK (340ms) → /checkout/done FAIL (cart cleared too early) v0> fix --bug 1 --redeploy → patch applied · rerunning browse → all routes pass · screenshot diff attached
Anthropic added eleven billion in a month.
Investor Gavin Baker, on Patrick O'Shaughnessy's podcast: Anthropic's annualised revenue grew eleven billion dollars in a single month — capital-efficient enough to embarrass its larger neighbour. Watch prefill/decode disaggregation for the next leg of GPU economics. TSMC capacity is the only honest bubble indicator.
The frontier still runs on Hopper.
Stanford's MS&E 435 economics-of-AI seminar, episode two. Every model the public has used this year was trained on H100s. Blackwell is partially online. Rubin has not yet produced a public frontier model. If you think the curve has bent, the seminar's quiet thesis is that the most interesting bend is still ahead.
ChatGPT names a next of kin.
Trusted Contact, OpenAI's new opt-in safety feature, lets a ChatGPT user name a person the assistant can quietly reach out to if the conversation drifts into risk. Activation is consent-first, conditions are spelled out in the setting, and the same classifier that flags the conversation is the one that drafts the note.
That's today.
Seven files, four labs, three Telegram desks, one Monday. Trust took the cover because deployment now lags capability further than any benchmark gap.
Sources
- Anthropic, news
- OpenAI, news
- Products & Startups
- Точки над ИИ
- Мысли Рвачева
Rubric
- AI tools you could adopt this week
- Creative software
- Dev tools & agents
- Research with a practical kernel
- Actionable for senior engineers
Today
- Issue 037
- Monday · 25 May 2026
- Zürich, 08:00 CEST