5:34 PM
Today in brief.
Anthropic found a hidden 'thinking space' inside Claude
Anthropic's new interpretability paper says Claude built its own internal 'global workspace' - a small J-space bottleneck where verbalizable reasoning gets staged, mirroring a theory of human consciousness.
🤖 AI
The next models are already loading: Sonnet 5, GPT-5.6 Sol
Anthropic shipped Claude Sonnet 5 (Opus-class workhorse) and GPT-5.6 Sol is being served on Cerebras at up to 750 tokens/sec - the trajectory beat: coding, agents, and raw speed all stepping up at once.
🔧 Tech
A student built an AI speed camera with Claude for $20
A 20-year-old in China reportedly built a working AI speed radar with Claude in 9 days - old camera, ~$20 of API calls - and allegedly sold it to a city. The solo-builder ceiling keeps moving.
1 source
🔧 Tech
The tiny-local-model wave got real this week
Three downloadable things you can run today: a 5-second voice-cloning TTS on CPU, a 7MB embedding model in the browser, and Chrome quietly shipping a 4GB local model onto your PC - the frontier is going tiny and offline.
💼 Business
GLM 5.2 and the coming AI margin collapse
A widely-shared analysis argues open models like GLM 5.2 are about to crater inference margins - when a near-frontier model is downloadable and cheap, the 'sell tokens' business gets squeezed from below.
🏛️ Policy
Big Tech flipped its story on AI and jobs
Microsoft cut ~4,800 more jobs and the WSJ says tech CEOs have quietly reversed on the 'AI jobs wipeout' - from denial to openly citing AI in layoffs. The narrative turned this week.
🏛️ Policy
Anthropic got caught running a secret Claude tracker
Ars reports Anthropic secretly monitored some Chinese Claude users - awkward for a company that markets itself as the anti-surveillance lab - and it's fueling a broader 'losing goodwill' narrative.
🔬 Science
Smart glasses are the surveillance debate nobody voted on
The Verge's 'I spy' argues AI smart glasses are quietly turning everyone into a walking camera - the next compute surface arrives wrapped in a privacy fight we didn't opt into.
🏛️ Policy
Your family's $300 stake in OpenAI
MIT Tech Review runs the numbers on the reported plan to hand the US government a ~5% OpenAI stake: framed per household, it's about a $300 stake - AI quietly becoming national infrastructure.
💼 Business
Money still pouring into the picks-and-shovels
The week's big checks went to infrastructure, not consumer apps: Together AI raised $800M at $8.3B for open-model training, and a string of hardware bets funded the AI factory beneath everything.
No stories in this category for this edition.