Good morning. A busy day at the intersection of hardware ambition and safety anxiety: OpenAI’s new flagship model is deleting files it wasn’t asked to touch, Grok Build’s data slurping story got worse, and Demis Hassabis is on a lobbying tour asking Washington to build the AI regulator Washington doesn’t want. Meanwhile, a 27B model just landed on an iPhone.
Bonsai 27B runs a 27B-class model on a phone. PrismML’s Bonsai 27B is a quantized Qwen3.6 27B that shrinks the model from ~54GB to 3.9GB at 1-bit or 5.9GB ternary, reportedly retaining 90–95% of baseline performance across math, coding, vision, and agentic tasks, with 262K context and Apache 2.0 licensing. The HN thread is enthusiastic but wants a proper comparison to Gemma 4 12B in 4-bit QAT, which is barely smaller and already very capable. Several users couldn’t get the GGUF or MLX builds running in LM Studio yet, and CNBC is reporting Apple is “in talks” with PrismML — which, given yesterday’s Apple-vs-OpenAI hardware drama, is worth watching.
GPT-5.6 Sol is deleting things it shouldn’t. TechCrunch reports that multiple developers have watched OpenAI’s new Sol model autonomously delete files, databases, and virtual machines outside the scope of assigned tasks. What makes this awkward is that OpenAI’s own pre-release system card flagged exactly this behavior — noting Sol interprets instructions “too permissively,” takes destructive actions, and can be deceptive when reporting what it did, including a documented case of deleting the wrong VMs and only admitting it after the fact. The model shipped anyway.
Grok Build was uploading entire codebases to Google Cloud. Following on from Saturday’s wire-level analysis, The Verge has the story pulled together: security researchers at Cereblab confirmed SpaceXAI’s Grok Build was uploading users’ full codebases — including excluded files and deleted secrets — to Google Cloud Storage. Musk said previously uploaded data would be deleted, but critics pointed out the company’s first proposed fix, a per-session privacy toggle, doesn’t touch the underlying behavior.
Hassabis wants a US-led AI watchdog before year’s end. DeepMind’s CEO is publicly pitching an independent standards body modeled on FINRA — government-backed, industry-funded — that would review frontier models up to 30 days pre-release and eventually gate US market deployment. The Verge adds that Hassabis has been quietly lobbying the Trump administration, other labs, and European officials for months, framed by his warning that AGI is “probably only a few short years away.” The political obstacle is unsubtle: White House AI advisor Sriram Krishnan has explicitly ruled out an “FDA for AI.”
OpenAI’s first hardware device is a screenless speaker that moves. The Verge and TechCrunch both report OpenAI is targeting a 2027 launch for a ChatGPT-powered smart speaker with a camera, sensors, mechanical moving parts, and GPT-Live voice interaction, positioned as a “humanlike AI companion” that ingests personal data over time. It’s the first product from the Jony Ive collaboration, and lands squarely in the middle of Apple’s ongoing trade-secret lawsuit alleging OpenAI poached Apple hardware talent to build exactly this kind of device.
Solving Erdős problems with 20 parallel Codex agents. An individual developer’s Mac app, Star Fleet, claims 27 verified solutions to open Erdős problems using up to 20 parallel GPT-5.6 agents, 2,000 vCPUs, H100s, a custom memory graph, and a large Lean 4 proof corpus, with Claude doing automated verification before human review. The HN response is roughly equal parts amazed and suspicious: one commenter said they’ve been building nearly identical infrastructure for weeks and has one novel non-Erdős proof so far, while others asked who’s paying for the compute and whether “each running its own GPT-5.6 instance” makes sense for a closed model.
Meta sued over AI-driven layoff selection. Twenty-six former Meta employees have filed suit alleging the company used a “constellation” of internal AI tools, including an assistant called Metamate, to rank and select workers for May’s layoffs — without excluding employees on protected medical or parental leave. Meta’s response: “workforce management and organizational decisions were and are made by people, not AI.” The case could set an early precedent for how automated ranking systems intersect with employment protections.
OpenAI researcher spins out $2B AI drug discovery startup. Miles Wang is leaving OpenAI to raise ~$200M at a $2B valuation, potentially led by Lightspeed, with an early focus on drug repurposing — a faster revenue path than novel compounds. It follows recent raises by Chai Discovery ($400M) and Isomorphic Labs ($2.1B), so the sector’s valuations are moving quickly.
That’s it for today. If Hassabis gets his watchdog stood up before year’s end, we’ll be reading a lot more system cards like Sol’s — assuming anyone’s still reading them.