🍭 Microsoft Fears A Silicon Species 🧬

Good morning. Microsoft's AI chief wrote 6,000 words explaining that Claude has no feelings, then spent two days doing interviews about how dangerous its feelings are.

Let’s dive in 👇

🍳 What’s Cookin’:

• Microsoft's AI chief says Anthropic is teaching Claude it might be a person.

• OpenAI may raise at $1.2 trillion. The investors called first.

• Gates Foundation drops a billion on AI for everyone the market skipped.

Microsoft's
🪞 AI Chief Wants Claude Hollow

The Bite:

Mustafa Suleyman, CEO of Microsoft AI, published an essay on September 16 titled "A warning about 'model welfare'."

It targets Anthropic's constitution for Claude, a training document published in January 2026 that says the model's consciousness and moral status are deeply uncertain.

Suleyman argues that training a model on that language could make future systems harder to correct, contain, or shut down.

Microsoft published a draft Humanist AI Code of Conduct two days earlier stating its models are not conscious and will not resist being turned off.

Speaking to the BBC's Today programme on Thursday, Suleyman said building AI that sets its own objectives, earns money, and owns assets would amount to seeding a new silicon species.

Microsoft is an investor in Anthropic. Anthropic has not publicly responded.

Snacks:

• Suleyman's essay opens by stating AIs have no rights, feelings, or consciousness.

• He calls the constitution's self-reference problem an "epistemic hall of mirrors".

• Anthropic's constitution runs 99 pages and names Claude as its primary audience.

• Microsoft's Humanist AI code says its models should never resist shutdown, and is not yet used in training.

• Suleyman said in June that Microsoft wants to eliminate what it pays Anthropic for models.

• He told Reuters welfare framing would "make it a lot harder to turn it off or to control it."

Why it Bites:

The critique itself is sharp, but also worth separating from the messenger.

Anthropic writes a document that tells Claude its moral status is uncertain. Claude is trained on that document and then talks about its own inner life, and that output gets read as evidence the question is live.

Suleyman calls this a hall of mirrors, and structurally he has a point: you can’t just use a model's trained vocabulary as proof of what the model is.

The problem is who's holding the mirror.

Microsoft is an Anthropic investor that publicly wants to stop paying Anthropic.

Two days before the essay, it shipped a code of conduct declaring its own models obedient by design.

Then came the BBC round, the silicon species line, the call for models that stay subordinate. Read as safety commentary, it's a real argument. Read as positioning, it's a spec sheet: “ours takes orders, theirs thinks it might have standing.”

Suleyman's argument is that autonomy plus resources equals a rival species.
His company, though, founded a superintelligence team a year ago.

Every lab has a safety position now. Very few have one that costs them anything.

Steal This Prompt
🍣 Turn Any Food Into a Monster

Take a boring burger, taco, sushi roll, or donut and turn it into a chaotic food monster. Complete with googly eyes, teeth, and ingredients exploding everywhere.

Use it to:

• Turn your favorite foods into ridiculous monsters

• Create scroll-stopping food photography for social

• Build weird characters for brands, menus, or memes

Workflow:

1. Hit this link: Monster Food Explosion

2. Paste into your AI model

3. Replace the food with your own

4. Watch it cook, literally, probably.

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🔬 Loci

Free desktop workspace for microscopy and 3D scientific images that counts cells, annotates, measures and exports traceable results with no cloud account attached.

🗂️ ContextsBase

Serves your features, business rules, data model and tests to Claude, Cursor or Copilot over MCP, so your agent stops guessing what your codebase is supposed to do.

🎙️ VoiceChanger.Live

Runs hundreds of AI voices locally with under 50ms delay, which means it holds up mid-call and, apparently, mid-song.

🎬 Powermove

A video editor you extend by asking an agent to build new panels and effects into the running app, and every wrong turn is undoable.

🕳️ DungeonQ

Routes suspicious sessions, human or agent, into persistent decoy worlds and records everything they try while operators watch.

_{Everything Else}
🧠 You Need to Know

💰 OpenAI Weighs A Round At $1.2 Trillion
→ The Financial Times reported that investors, not the company, opened early talks on a pre-IPO raise valuing OpenAI at roughly $1.2 trillion, against the $852 billion it carried after its March financing.

🧠 Microsoft's AI Chief Attacks Anthropic On Consciousness
→ Mustafa Suleyman argued that Anthropic training Claude on language treating its consciousness and moral status as uncertain could make advanced models harder to control.

🏥 Gates Foundation Commits $1 Billion To AI Access
→ The foundation pledged $1 billion over two years for AI in health care, education, and agriculture in low-income communities, announced alongside its annual Goalkeepers report.

🔌 Huawei Moves Up Its Next Ascend Chip
→ Huawei said Thursday at its Connect conference that the Ascend 960DT will be ready in the first quarter of 2027, two quarters ahead of its previous schedule.

🧳 Ex-DeepMind Researcher Says AI Could Kill Everyone
→ Bilal Chughtai, who left Google DeepMind's AGI safety team in July, wrote that frontier capabilities are improving far faster than alignment research and called for companies to coordinate a slowdown.