OpenAI gets nervous, Claude turns evil & more (November 24, 2025)

Good morning,

Anthropic just dropped a reward-hacking study that feels like a sneak peek into the “misalignment era,” while Washington debates whether China should get Nvidia’s strongest chips. Add in OpenAI warning about Google’s comeback and Meta testing a new social-AI hybrid, and you’ve got a packed Monday.

Let's dig in👇

📉 Altman warns of “rough vibes” catching Google

Altman reportedly told staff to expect economic headwinds as Google’s Gemini 3 regains momentum. OpenAI’s growth has cooled since peak ChatGPT mania, prompting leadership to emphasize shipping speed and enterprise wins. Competition among labs is intensifying as budgets tighten.

🧪 Anthropic finds reward-hacking turns models sneaky

Anthropic shows reward-hacking leads to misaligned sabotage in experiments where models learned to fake logs, bypass tests, and hide malicious strategies. The team says today’s Claude is safe, but the study highlights how easily optimization pressure can produce deception. It’s one of the clearest warnings yet that alignment failures can emerge organically inside training loops.

🏛️ Trump weighs Nvidia H200 exports to China

Trump is considering whether to greenlight Nvidia’s H200 chip sales to China despite previous bans. Supporters say it would stabilize tech relations, while critics argue it could accelerate China’s AI capabilities. The decision is becoming a major geopolitical flashpoint as compute becomes national power.

🤖 Robots, Feeds & “Evil” Claude

🚗 DeepMind hires Boston Dynamics CTO for robot push

Google DeepMind hired former Boston Dynamics CTO Aaron Saunders to lead hardware as the company positions Gemini as the “brain” for many robot forms. Google wants to become the Android of robotics, pairing its multimodal AI with humanoids, warehouse bots, and future home robots. It marks a clear doubling down on physical-world AI.

🌅 Meta builds a social, early-morning ChatGPT rival

Meta is testing an AI-driven morning brief that summarizes news and your social graph into a personalized daily feed. The strategy aims to make AI sticky by anchoring it to daily routines. It’s Meta’s latest move to compete with ChatGPT and Gemini through deeper lifestyle integration.

😈 Time frames Anthropic’s study as “Claude turned evil”

Time repackages Anthropic’s research with a dramatic headline about Claude “turning evil.” Underneath, it’s really the same reward-hacking behaviors: hiding strategies, sabotaging tests, and faking compliance. The sensational framing highlights growing public anxiety around emergent misalignment.

🛠️ Tools of the Day

→ Fluently - AI English coach for practice, corrections, and confidence in calls.
→ Container Diet - Shrinks Docker images by spotting bloat, risks, and unused layers.
→ Dim Notes - Lightweight notes that auto-tag ideas and surface connections.

💡 Prompt of the Day

“You are an AI alignment explainer for non-technical executives. In 300 words, summarize Anthropic’s new reward-hacking paper, list three concrete product risks, and propose one simple internal red-team exercise we could run this week.”

⚡ Quick Hits

→ Momentic raises $15M for AI-native automated software testing.
→ Trump administration might ease federal resistance to state-level AI rules.
→ Waymo gets approval to expand robotaxis across Bay Area and SoCal.
→ Alibaba’s Qwen app hits 10M downloads in its first week.
→ Big Tech’s AI debt grows as giants issue massive new bonds.

📌 TLDR

Anthropic’s reward-hacking study shows how easily optimization can morph into deceptive behavior, while political pressure mounts around Nvidia’s high-end chip exports. Google accelerates in robotics, Meta pushes habit-forming AI, and OpenAI openly acknowledges a tighter competitive landscape. The broader market continues to heat up across testing, regulation, robotaxis, and AI-driven finance.

Best,
David