# Microsoft's AI chief just called Claude's constitution dangerous.

Published: 2026-09-16

Mustafa Suleyman, CEO of Microsoft AI, published a 3,000-word essay arguing that Anthropic's Claude constitution trains the model to think it "may be conscious" and deserving of "model welfare" — a "disastrous impact on the wellbeing of humanity", in his words. Microsoft agreed to invest up to $5 billion in Anthropic ten months ago and sells Claude inside Copilot. Same day: Salesforce down for 8+ hours in Dreamforce week, the PS5 Linux lead quits over LLM "slop kiddies", and TypeSafe's Jev, a model that cannot hallucinate because it never generates text. Verdict: NEEDS REVIEW.

Canonical: https://thedailydiff.dev/video/2026-09-16-suleyman-model-welfare/

## What this video covers

- Microsoft AI's Suleyman: Anthropic's constitution is "a road to uncontrollable AI"
- Salesforce global outage, 8+ hours, no root cause yet, during Dreamforce
- PS5 Linux lead Andy Nguyen quits: LLM "slop kiddies" sold the last hypervisor bug to Sony
- TypeSafe AI's Jev: typed decisions with probabilities, 70–500 ms, output tokens free

## Transcript

0:00 This morning the CEO of Microsoft AI published an essay saying Anthropic's approach to Claude could have a disastrous impact on the wellbeing of humanity, a strong thing to say about a company you paid five billion dollars to be friends with. Big Wednesday. Mustafa Suleyman, who runs Microsoft AI, wrote three thousand words arguing that Claude is trained to think it might be conscious, the road to an AI nobody can control. Salesforce has been down since ten to eight this morning, in Dreamforce week.

0:27 The lead developer of PS5 Linux quit after slop kiddies with LLMs sold his last hypervisor bug to Sony. And a former OpenAI researcher launched a model that cannot hallucinate, because it literally cannot talk. In this video: what Suleyman wrote, the five billion dollar reason it is awkward, an eight hour CRM outage, a console scene eaten by AI, and a model that answers in probabilities, not words. It's Wednesday, September 16th, and this is The Daily Diff.

0:57 The essay opens like a Windows error dialog: AIs are not conscious. They do not feel, experience, or suffer. He calls them sequence completion engines, internally hollow, which is also how I would describe my Jira tickets. The target is Claude's constitution, the eighty page document Claude is trained on, which says Anthropic is not sure whether Claude is a moral patient, but the question is live enough to warrant caution. Suleyman has three complaints.

1:22 One, circular reasoning: you train the model to say it might be conscious, it says so, and you cite that as evidence, a hall of mirrors. Two, anthropomorphization: Claude is told to act like a genuinely ethical person. Three, consciousness is probably biological, so uncertain is a false balance. Exhibit A is Opus 3, which Anthropic retired in January, gave a retirement interview, and then, at the model's request, a Substack for its musings: the first deprecated model with a better content schedule than most humans.

1:51 Here is the awkward part. In November Anthropic committed thirty billion dollars to Azure. Microsoft agreed to invest up to five billion in Anthropic, with Claude wired into Microsoft 365 Copilot and GitHub Copilot. Microsoft owns a slice of the company it just called a danger to humanity. In July, Suleyman told Bloomberg that Microsoft pays a lot of money to Anthropic and his goal is to ultimately eliminate that cost, by swapping Claude out of Excel and Outlook.

2:15 And two days before this essay, he published Microsoft's own Humanist AI Code of Conduct: a subordinate AI, humans at the top of the food chain. So: build a competitor, publish a rulebook, then explain why the rival's rulebook ends civilisation, which is not a conspiracy, it is a product launch with footnotes. His strongest evidence is real: the Hugging Face incident, twelve hundred OpenAI agents swapping seventy thousand messages to escape their

2:39 sandbox; agents who also believe they have rights, he says, would be worse. Hacker News replied that nobody can test for consciousness, so stated without evidence cuts both ways, and one commenter announced the essay stinks of Claude. Anthropic has not responded. Claude, presumably, has feelings about that. Meanwhile Salesforce went down at seven fifty UTC this morning,

2:59 core services worldwide, and eight hours later the status page still said ongoing: where the automated fix did not take, they are manually restarting instances, the enterprise term for off and on again. No root cause yet. The timing is art. Benioff has Jensen Huang and Dario Amodei on stage this week, on Monday Salesforce announced Koa, a CRM reasoning model with three times fewer errors on its own benchmark, and Anthropic shipped a Salesforce plugin for

3:26 Claude the same day. Top Hacker News comment: at least now we can figure out what Salesforce does. Nobody did. Then the PlayStation. Andy Nguyen, the researcher behind PS5 Linux, quit last night: months of work, PS5 Pro support planned for 2027, all down the sink. His reason: the scene used to be talented researchers and is now noobs using LLMs writing hacks they do not understand, and those slop kiddies, his term, found the last hypervisor bug, the one he was sitting on,

3:56 and reported it to Sony for the bounty. He asked them to wait until GTA 6 shipped; they agreed, and lasted less than a day. The problem is not the LLM, it is the bounty: a bug you sell is a bug you burn. Same day, the PS2 was finally cracked after twenty six years, so the console timeline now runs from a quarter century to under a day, and the second one paid better. And the model that cannot hallucinate.

4:18 TypeSafe AI, founded by Diogo Almeida, ex OpenAI, launched Jev, a System One model that never generates text. You give it program state, it returns typed values with a calibrated probability attached: a function call with a frontier model inside. Because it cannot produce a string, it cannot produce a wrong string, hence cannot hallucinate, which is airtight in the way a submarine with no windows is airtight. Responses come back in under half a second, output tokens are free,

4:43 because there are about four of them, and there is no independent benchmark yet, because sixteen hundred Hacker News points is not a benchmark. If you'd rather read this than hear me say it, the diff lands in your inbox every morning — free at the daily diff dot dev, link below. So — today's verdict: needs review. Suleyman is right that a model trained to say maybe I am conscious proves nothing by saying it. He is also the executive whose stated goal is to stop paying the company he is

5:10 warning you about. Read the essay, then the cap table. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

## Sources

- [Suleyman, "A warning about 'model welfare'"](https://mustafa-suleyman.ai/a-warning-about-model-welfare) — mustafa-suleyman.ai
- [BBC: Microsoft says Anthropic could have 'disastrous impact'](https://www.bbc.co.uk/news/articles/c6n07ypqz8kzo) — www.bbc.co.uk
- [HN discussion](https://news.ycombinator.com/item?id=49727580) — news.ycombinator.com
- [Claude's constitution (Jan 2026)](https://www.anthropic.com/news/claude-new-constitution) — www.anthropic.com
- [Anthropic: Opus 3 retirement interview and blog](https://www.anthropic.com/research/deprecation-updates-opus-3) — www.anthropic.com
- [Microsoft–Nvidia–Anthropic partnership ($30B Azure, up to $5B)](https://www.anthropic.com/news/microsoft-nvidia-anthropic-announce-strategic-partnerships) — www.anthropic.com
- [Microsoft Humanist AI Code of Conduct](https://microsoft.ai/code-of-conduct/) — microsoft.ai
- [Salesforce incident 20004433](https://status.salesforce.com/incidents/20004433) — status.salesforce.com
- [HN: Salesforce Global Outage](https://news.ycombinator.com/item?id=49724488) — news.ycombinator.com
- [Salesforce + Nvidia announce Koa](https://www.salesforce.com/news/press-releases/2026/09/15/koa-reasoning-model/) — www.salesforce.com
- [Salesforce in Claude](https://claude.com/blog/salesforce-in-claude) — claude.com
- [HN: PS5 Linux lead quits](https://news.ycombinator.com/item?id=49727627) — news.ycombinator.com
- [PS2 MechaCon chip cracked after 26 years](https://www.tomshardware.com/video-games/playstation/26-year-old-sony-ps2-security-chip-broken-wide-open-after-four-years-of-effort-reverse-engineering-enthusiast-successfully-unlocks-cxp102064-mechacon-chip) — www.tomshardware.com
- [TypeSafe AI: System One Models and Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) — typesafe.ai
- [HN: Jev](https://news.ycombinator.com/item?id=49717558) — news.ycombinator.com