This was the week the AI world split into two camps: those building, and those warning. Both were loud.
The warnings
- Microsoft: an exec called AI web-scraping "the largest theft of labor in human history" - top of Hacker News at 947 points
- US military: an AI-generated intelligence report contained hallucinations and nearly influenced a real decision
- ChatGPT: revealed to collect signals about your activity on other websites via ad tools
- United Nations: Secretary-General Guterres used his final UNGA address to warn about runaway AI and autonomous weapons
The building
At Alibaba's Yunqi Conference in Hangzhou, three launches landed in one day:
- CPFS storage - cuts AI training storage costs by 69%
- QwenNote - first hardware product for the Qwen agent stack
- Lingjun M890 supernode - expanding AI compute to a second region in Q4
OpenAI, meanwhile, says its internal models solved 100+ long-standing math problems in 24 days - and installed an independent advisory board to prove it. After a $100M lawsuit and political scrutiny, verified capability is the new currency.
The money
JPMorgan's Jamie Dimon: hyperscaler AI spend could reach $1 trillion next year, possibly "slightly inflationary." The counterintuitive datapoint: Nvidia now trades below 17x forward earnings - a decade low for the AI hardware king.
My take
The industry is moving from "tell stories" to "show homework." Companies that cut real costs (69% is real), survive regulators, and open themselves to verification will take the next leg. The rest get cleared out in the correction.
What do you think - overshoot or underpriced?
Top comments (0)