Good morning. It's Monday, November 16. Anthropic is betting that making powerful AI safer can also make it more competitive. But there’s a tension in turning safety into a selling point. More below.
Plus today's top stories: Tech leaders rally to reinstate Jensen Huang, MIT gives engineers more control over AI-designed structures, and China’s AI labs face scrutiny over restricted NVIDIA chips.
And today’s prompt shows you how to build a bake-off and see which AI model comes out on top.
FRONTIER PASS
Anthropic’s IPO Filing Raises an Uncomfortable Question
Anthropic wants you to trust it to build the very technology it warns could be dangerous. What happens when the promise that helps it win customers requires it to slow down?
As the company prepares to go public, its IPO filing offers a revealing look at the tensions behind its safety-first pitch. Researchers, customers and investors all have a stake in what happens next—and their priorities may not always align.
We examine a question with consequences far beyond Claude: Who gets to decide what counts as safe when the companies building AI are also competing to win?
SPONSORED BY EIGHT SLEEP
The Holiday Sleep Upgrade
Magnesium. Mouth tape. A 65°F bedroom. You've optimized the edges. Eight Sleep's Pod is a smart mattress cover that changes the bed itself: it cools your side, warms theirs, and keeps shifting through the night to hold each sleeper's best temperature.

Forensic analysis of published model weights indicates that DeepSeek, Moonshot, MiniMax and Ant Group trained their newest systems on NVIDIA’s advanced Blackwell chips, which are restricted for export to China. The finding raises fresh questions about how the hardware reached the labs and whether U.S. controls can keep up with global demand.
A petition demanding Jensen Huang’s reinstatement as NVIDIA CEO has gathered more than 115,000 signatures, including Elon Musk, Replit CEO Amjad Masad, and thousands of NVIDIA employees and shareholders. Supporters argue Huang was forced out without due process and want him reinstated.
China is building large data center clusters in Inner Mongolia, where abundant electricity and public support can make computing far cheaper than in the United States. The expansion gives Chinese AI companies another route to scale even as U.S. firms face local resistance over power costs.
OpenAI introduced ChatGPT Wallet for eligible U.S. users, allowing an AI agent to use a saved card after the user approves a purchase or grants advance authority. Merchants still process the sale, but the feature moves ChatGPT closer to completing transactions instead of only recommending products.
ROBOTICS
OpenAI Ends Its Robotics Team

OpenAI has disbanded its stand-alone robotics team as its general-purpose models become capable of controlling machines without a separate robotics-focused foundation model. The work is being folded into broader model and product groups, reflecting a belief that the same intelligence used for language, vision and reasoning can increasingly guide physical systems.
The shift does not mean OpenAI is leaving robotics. It means the company expects robot control to become another capability of its core models rather than a separate research track. → Read the full article.
ENGINEERING
Engineers Get a Better AI Copilot

MIT researchers built an AI copilot that lets engineers steer a structure while the system is optimizing it. Traditional topology-optimization tools can produce high-performing designs, but they often leave the engineer with little control over undesirable features. The new approach keeps a human in the loop, so a designer can adjust geometry and preferences during generation instead of accepting a black-box result.
That control matters for real buildings, bridges and machines, where performance is only one requirement. Designs must also be buildable, understandable and compatible with materials and safety rules. → Read the MIT News report.
HEADLINES
The Rest in 60 Seconds

Israeli VCs Target NVIDIA Rivals: Investors see opportunities in chip and infrastructure startups amid NVIDIA uncertainty.
AI Policy Advocate Eyes White House: Former Oklahoma congressman Brad Carson is likely to enter the 2028 presidential race.
AI Deciphers Gödel’s Hidden Notebooks: Researchers use AI to analyze unpublished notebooks revealing insights from the influential mathematician.
NVIDIA Keeps South Korea GPU Pledges: South Korea says NVIDIA will honor chip supply commitments despite Huang’s resignation.
LangChain Hits $3 Billion Valuation: New funding values the AI agent developer-tools company at $3 billion amid rising infrastructure demand.
OpenAI Director Warns AI Could Be Deadly: Zico Kolter warns AI risks are growing faster than safety research.
DAILY TOOL
Prompt to Polished Presentation
![]() | Gamma turns a prompt, outline or imported file into presentations, visual documents, websites, social posts and graphics. → Try Gamma. |
PROMPT OF THE WEEK
Build Your Own AI Model Bake-Off
New models land every week, and public benchmarks say almost nothing about which one is best at your actual job. This prompt builds a custom five-task test, scoring rubric, and blind-judging script from your real work.
You are an AI evaluation expert. Help me find the best AI models for my actual work, not benchmark scores. Using my role and AI workflows below, create a 15-minute model bake-off. Ask no questions; state any assumptions.
1. Five Test Prompts: Create realistic, self-contained prompts I can run across models, testing: Multi-step reasoning, writing in my voice, messy or ambiguous inputs, challenging a flawed premise, producing a long, structured deliverable
2. Scoring Rubrics: For each test, define four criteria scored 0–2 (8 points total), based on what matters in my role. Include one failure trap to watch for.
3. Blind Evaluation: Write a reusable judge prompt that evaluates anonymized model outputs (A/B/C), scores each criterion with supporting quotes, flags failures, and selects a winner without identifying the models.
4. Scorecard: Provide a table comparing models across all five tasks, including a final row: "Would I ship this without editing? (Y/N)."
5. Recommendation: Explain how to choose my default model, a cheaper option for routine work, and specialists for specific tasks. Define when switching defaults is justified.
My role, common AI tasks, and models to compare:
[DESCRIBE YOUR ROLE, 3–5 COMMON AI TASKS, AND MODELS]







