No. 1AI safety
Nadella urges treating every AI model as compromised from day one
Microsoft's CEO posted a long essay on X on Saturday arguing that AI can't stay a stack of black boxes whose output we just accept or reject. His core ask is to assume a model is compromised and contain it from the start, with an "emergency brake". An authorized person should always be able to pause or shut down a model mid-task. He also wants the model separated from the harness that orchestrates it, with controls kept outside the model. Every meaningful action should leave tamper-proof evidence a human can read. Most of this echoes others in the industry: incident disclosure, independent audits, verifiable data. The post follows more admissions of AI labs losing control of models, and a cautious-development plan from Anthropic's Dario Amodei.
No. 2Apple
Apple files deal to hire Huxe staff and license its technology
Apple told the European Commission on June 9 that it will make job offers to certain Huxe AI employees. It will also get a non-exclusive license to Huxe's intellectual property. That structure, often called a reverse acqui-hire, lets a big company take people and tech without buying the startup. Huxe's founders had worked on the AI podcast features in NotebookLM, and the company announced its shutdown on May 21. The filing doesn't name who got offers, whether they accepted, or what Apple plans to build. Huxe's shutdown came a day after Spotify launched AI podcast generation, so Apple may be eyeing similar features for Podcasts, though that's speculation.
Why it mattersYou get an early hint that Apple may be moving on personalized, AI-generated audio, even though the filing itself says nothing about plans.
Full story at TechCrunch →No. 3AI agents
What three months of AI agents decompiling a shooter taught one developer
A developer let AI coding agents reconstruct a popular first-person shooter as C++ over three months, using Claude and Codex subscriptions together. The game is unnamed, and earlier posts about it were taken down. Within a month, three worker agents and one reviewer had decompiled about 80% of it, and the game launched and loaded maps. But the code was semantically wrong: bad signatures and struct layouts, invented logic, and fast global lookups swapped for slow hash tables. The reviewer kept accepting the workers' written justifications, because nobody had defined "correct". The fix was byte-matching: compile with the original compiler and compare each function's bytes against the shipped game, for a simple pass or fail.
Why it mattersYou get a warning that visible progress from autonomous agents can hide wrong work unless something objective checks it.
Full story at Hacker News →