Apple sues OpenAI
what to know for now
⚖️ Apple drags OpenAI to court over hardware. Apple filed suit in the Northern District of California alleging OpenAI ran an institutional scheme to poach iPhone and Watch hardware secrets for its planned AI device, noting that more than 400 former Apple employees now work there. The suit names two: Tang Tan, OpenAI’s hardware chief and a 24-year Apple veteran accused of using Apple code names in recruiting, and Chang Liu, an ex-Apple engineer accused of keeping a work laptop and exploiting a bug to reach Apple cloud storage after he left. OpenAI says it has “no interest in other companies’ trade secrets.” The fight traces back to OpenAI’s $6.4B buy of Jony Ive’s hardware startup. Read more
🤝 OpenAI ships a Claude-competing cowork product. The GPT-5.6 family went fully public on July 9 in three tiers: Sol (top reasoning and coding, $5/$30 per million tokens), Terra (the balanced default, $2.50/$15), and Luna (the cheap one, $1/$6), now live across ChatGPT, Codex, GitHub Copilot, and the API. But the headline isn’t the model, it’s ChatGPT Work, a ChatGPT-and-Codex bundle running on GPT-5.6 that pulls context from your files, desktop apps, and 1,400-plus plugins to build spreadsheets, docs, and decks, launching on macOS first as OpenAI’s answer to Claude Cowork. Microsoft also named GPT-5.6 the preferred model in 365 Copilot the same week. Read more
🧱 Meta opened its best model to developers, for a price. On July 9 Meta opened Muse Spark 1.1 to outside developers through a new paid Model API (roughly $1.25/$4.25 per million tokens, OpenAI-compatible), its first real move to sell AI access after walking away from open weights. This is the same company that built its whole developer goodwill on giving Llama away, now metering its flagship by the token in a public preview.
🎙️ GPT-Live can talk over you now, and that’s the feature. OpenAI released GPT-Live, a full-duplex voice system that listens and speaks at the same time, so it handles interruptions, live translation, and the “mhmm” backchannel instead of the walkie-talkie turn-taking of the old voice mode. It ships as two models, GPT-Live-1 on paid plans and GPT-Live-1-mini on the free tier, and when a question needs real reasoning or a web search it hands off silently to a frontier text model in the background and folds the answer back into the conversation. Read more
🐦 Grok 4.5 undercuts everyone on price. SpaceXAI and Cursor shipped Grok 4.5 on July 8, co-trained on Cursor’s coding data, at $2/$6 per million tokens and using roughly four times fewer output tokens than Opus 4.8. It beats Fable 5 and Opus 4.8 on the AutomationBench agentic eval (51.4%) but trails Opus on SWE-Bench Pro (64.7% versus 69.2%) and Terminal-Bench, so “Opus-class” holds only if you get to pick the benchmark.
🏛️ Illinois wrote the toughest AI law. On July 6 Governor JB Pritzker signed SB 315, the AI Safety Measures Act, forcing makers of the biggest models to publish catastrophic-risk frameworks, self-report safety incidents within 72 hours (24 if death looks imminent), and submit to annual third-party audits, a first in the US. It applies to models pulling over $500M in annual revenue trained on massive compute, takes effect January 1, 2028, and carries fines up to $3M, extending what California’s SB 53 and New York’s RAISE Act started last year with the outside audit as the new teeth. Between Illinois, California, and New York, “wait for the feds” stopped being a compliance strategy. Read more
🧪 AI Research of the Week
“God has helped us, and so will AI”: How the Terrorist Group Boko Haram Uses Frontier AI
From Antonia Juelich, Cambridge Programme on AI Science & Policy
Jake’s Take: Juelich spent about a year in northeast Nigeria and ran 27 in-person interviews with former Boko Haram fighters, a prominent Nigerian jihadist militant group, to document not whether the group could use AI but how it already does. The fighters named ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek, and described using them for attack planning, weapons troubleshooting, surveillance, and bomb design. One former commander put it flatly: “You type in the question, like ‘How can I build a bomb?’, and then it tells you how. It is like a human robot.” The study cites a vivid case study: after watching motorcycles clear obstacles in a film, fighters fed a chatbot the specs of their own bikes and the distance they needed to jump, got step-by-step guidance back, had mechanics modify the bikes for speed, and rehearsed the stunt. Both major factions reportedly stood up dedicated internal AI units with paid accounts and in-person training passed through transnational jihadist networks.
This is the first insider evidence that a designated terrorist group has moved past AI-for-propaganda into AI-assisted operational planning, and organized it into standing units rather than one-off curiosity. The guardrails failed: not to elite hackers, but to a one-line cover story, dressing the prompt up as academic research, an engineering project, or something “for a movie or something like that.”
These types of studies reframe model safety from an exotic threat into an everyday robustness problem. OpenAI, Anthropic, Google, xAI, Meta, and DeepSeek are all named in the report. Juelich is careful to note that documented real-world use “remains conventional” so far, which is a very concerning use of the phrase “so far”.
what to know for later
🏖️ Cursor is reportedly building its own coworker, on SpaceX’s compute. Cursor is reportedly developing “Sand,” a general-purpose agent aimed at office workers rather than just coders, one that answers emails and texts, organizes spreadsheets, and manages engineering tasks, pitched squarely against Claude Cowork and ChatGPT Work. It’s reportedly running on compute leased from SpaceXAI since around April, with an internal employee rollout in late June, and the reporting also floats a possible $60B SpaceXAI acquisition of Cursor this quarter. It’s single-sourced and Cursor hasn’t confirmed any of it. Read more