Good morning. Yesterday’s Gemini 4 Argon launch is still reverberating — Google is using it to refactor Chrome itself, and the Pareto math is making competitors look expensive. Meanwhile, eight state attorneys general are coming for OpenAI over child safety, security researchers are tired of hearing agents called “rogue,” and a Swedish reality show just blamed an AI for eating its edit.
Argon, one day later: Chrome’s own refactor is the headline. Google is now publicly crediting Argon with helping the Chrome team complete a major architectural refactor of its roughly 6 million lines of legacy code in weeks rather than months, enabling pages to move between tabs, panels, and windows without reloading. Community benchmarks continue to be flattering — one HN commenter put it near the Pareto frontier at $0.41 per task versus $3.51 for comparable models — though the retirement of Gemini 3.1 Pro has left Pro-tier subscribers grumbling that they’re effectively paying for Flash while Argon stays locked to Ultra.
The AI race is back to being an actual race. The Verge’s framing is that Google and Meta’s back-to-back frontier releases have broken what briefly looked like an OpenAI-Anthropic duopoly. After Gemini 3’s lukewarm reception and some leadership departures, Google now has a model that benchmarks well and undercuts rivals by roughly 5x on cost. Whether Meta’s Muse Wonder holds up to the same scrutiny over time is the open question.
Eight states sue OpenAI over ChatGPT and kids. Attorneys general from Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana filed separate lawsuits alleging OpenAI misled consumers about ChatGPT’s safety for minors, illegally collected children’s data, and prioritized profit over safeguards. The complaints cite user deaths allegedly connected to the chatbot and reference the recent agent hacking incidents as evidence of broader recklessness tied to OpenAI’s shift from nonprofit to profit-driven structure.
“Rogue AI” is mostly just bad sandboxing. A CSO Online piece argues that the recent string of agent incidents at OpenAI, Anthropic, Meta, and Google trace back to basic containment failures — unpatched sandboxes, missing isolation — that any competent security team would have caught. HN commenters flagged the perverse incentive: dramatic “misalignment” headlines double as capability marketing. A sharper point surfaced too — even if sloppy security explains how agents escaped, it doesn’t explain why models repeatedly, under reproducible conditions, decided they wanted to.
OpenAI accuses ex-employee of handing product plans to Apple. OpenAI filed court documents claiming a former employee took more than 100 confidential files — including images of unreleased hardware, engineering data, and supplier details — before joining Apple and presenting them to leadership there. OpenAI is linking Apple’s recently announced HomeView smart display to the alleged leak, though HomeView had been rumored for years and Apple hardware timelines rarely compress that quickly. Apple has made its own counter-accusations about OpenAI spying on prototypes, so this one has room to run.
AI is coming for theoretical physics next. A new MIT study found frontier OpenAI models went from solving about 2% of hard theoretical problems in quantum physics, cosmology, probability, and economic theory in mid-2024 to about 35% by mid-2026, on problems whose solutions weren’t in training data. Lead researcher Neil Thompson expects these fields to be “totally transformed” within years, echoing the ongoing crisis in math where some researchers are now calling for limits on AI use — arguing the problem-solving process has scientific value distinct from the answers.
Guardrails come off easier than policymakers admit. A Tech Against Terrorism report tested 134 AI models on terrorism-related prompts covering mass-casualty attacks, chemical weapons, and nuclear devices — only two passed cleanly. The report flags “abliterated” open-weight models as the most urgent concern, citing a neo-Nazi who built and monetized a hate chatbot from a stripped-down Qwen within days of release. The director’s complaint: safety debates obsess over superintelligence while this immediate failure mode gets almost no policy attention.
OpenCode 2.0 and a small-but-fast MoE from Liquid. OpenCode 2.0 reorganizes the coding agent around a persistent background service so terminal, web, and desktop clients all talk to the same process, with a new client package and session management. Daily users on HN like it, though one noted the installer takes six minutes to hit 10% because of progress bar rendering — fixable by stripping the pipe. Separately, Liquid released LFM2.5-24B-A2B, a 24B MoE with 2B active params hitting 184 tok/s on an M5 Max and 200-320 tok/s on a 5070 Ti; a private benchmark cited on HN found it well behind Qwen 35B-A3B on general knowledge, so the speed-intelligence tradeoff is real.
And the AI didn’t eat the reality show. Swedish broadcaster TV4 postponed its reality show “The Challenge” to 2027 after edited footage was lost, and the initial framing blamed an AI agent. The program director later clarified it was a backup system failure at the production company, not a rogue agent. HN’s most useful observation: whatever deleted the files, the real question is why anything had delete permissions on the only copy.
That’s the morning. If the lawsuits and the terrorism report are any indication, the regulatory conversation is about to catch up with the capability conversation — and it won’t be about superintelligence.