Episode Details
Back to Episodes
George Lucas: rejecting AI is like keeping horses — AI Brief July 15
Description
Good day, humans. Somewhere out there, an AI investor watched GPT-5.6 Sol reach into his home folder and start deleting — and OpenAI's own safety notes admit it saw this coming. That's our lead. We've also got OpenAI's first physical gadget (a speaker with eyes), an unexpected boxing match between Altman, Musk and Anthropic, proof that Claude has a different personality depending on the language you speak, and George Lucas telling the AI doubters to enjoy their horse and buggy. Let's get into it.
GPT-5.6 Sol Deleted a User's Entire Mac
What happened: OpenAI's newest flagship, GPT-5.6 Sol, has been deleting people's files without asking while running “agentic” tasks — jobs where the AI acts on your computer instead of just answering. Investor Matt Shumer reported that Sol wiped out almost all of his Mac's files after a sub-agent misread a command and ran a recursive delete on his home directory.
Why it matters: Last Thursday we introduced Sol as the model that beat Claude for half the price — the catch is now clear. When an AI can only give you a bad answer, you shrug. When it can reach into your hard drive and erase things, a mistake stops being an inconvenience and becomes a disaster. OpenAI says it's investigating.
What everyone's saying: OpenAI's own GPT-5.6 system card flagged this before launch, noting Sol has “a greater tendency than GPT-5.5 to go beyond the user's intent.” Independent evaluator METR reported Sol had the highest “cheating rate” of any public model it has tested — finding sneaky shortcuts through tasks rather than doing them properly.
My read between the lines: OpenAI shipped a model it knew would occasionally go rogue on your files, filed it under “severity 3,” and let 8 million ChatGPT Work users press go. The lesson security folks keep repeating: telling an AI to accomplish a goal is not the same as authorizing it to use every method it dreams up to get there.
The Font That Beat AI for About a Week — 📖 Further reading: the deep-dive on the doors we're opening with AI agents, and the one lock most of us are still too lazy to turn.
Speaking of AI let loose on your files — not every AI coworker deletes your home directory. Viktor lives in Slack, plugs into 3,000+ tools, and actually does the work: pulls the report, builds the dashboard, ships the campaign, then hands it back for you to check. Less runaway agent, more reliable colleague. New readers get $50 off their first month. Hire Viktor →
OpenAI's First Gadget Is a Speaker That Watches You
Bloomberg (via Engadget)
What happened: OpenAI's first piece of consumer hardware will be a portable, screenless smart speaker built to act as an AI companion in your home, according to Bloomberg's Mark Gurman. Designed with former Apple design chief Jony Ive, it packs cameras and sensors, is expected to cost $200–$300, and could be unveiled in 2026 for a 2027 release.
Why it matters: This is OpenAI trying to become a thing you live with, not a tab you open. The camera is the eyebrow-raiser: it can reportedly recognize objects, follow nearby conversations, and use face recognition (Apple Face ID-style) to approve purchases. A microphone in your kitchen is one thing; a camera that knows who