In this edition
Happy Friday, fam 👋
This week the AI wars got physical. Apple took OpenAI to court over hardware secrets and 400 poached employees — and OpenAI responded, in spirit, by shipping its first-ever gadget. Meanwhile Anthropic kept assembling an empire ahead of its IPO, every frontier lab got its safety report card (nobody made the honor roll), and two icons checked into the graveyard.
⚖️ Apple sues OpenAI: 400 hires, smuggled parts, and a 41-page complaint.
⌨️ OpenAI's first gadget is… a $230 keyboard with a dial for brainpower.
🏦 Anthropic's IPO march: a banker for compute, a Samsung chip, a Blackstone bet.
🛠️ Five high-impact tools you've (probably) never heard of.
🪦 DALL·E 3 checks into the graveyard + a report-card ethics read.
⚖️ Apple v. OpenAI: 41 pages of receipts
On July 10, Apple sued OpenAI for trade-secret theft, naming OpenAI, its io Products hardware unit, its chief hardware officer, and a former Apple engineer. Apple says the scheme ran "at every level": over 400 former Apple employees now at OpenAI, job candidates allegedly told to bring "actual parts" from Apple to interviews for show-and-tell, an engineer who allegedly used an authentication bug to pull confidential hardware files after leaving, and suppliers led to believe Apple had blessed the use of its own metal-finishing technique. Apple wants damages and an injunction — a direct shot at the Jony Ive device program before it ships anything.
OpenAI's actual first gadget arrived the same week, and it's… a $230 keyboard. The Codex Micro — built with macropad maker Work Louder, shipping July 24 — has 13 mechanical keys, a joystick, and a rotary dial that literally turns your AI agent's reasoning level up and down. Dial-a-brain. The device wars have officially begun, in the smallest possible way.
🏦 Anthropic is assembling an empire in public
Everything Anthropic did this week points at one date. Monzo co-founder Tom Blomfield joined its compute team — the fintech operator now managing the industry's scariest line item. It's in talks with Samsung to fab a custom 2nm chip and attack that compute bill at the silicon layer. And it teamed up with Blackstone on the bet that the next trillion-dollar AI business is implementation, not models. The context for all of it: a confidential S-1 already sits at the SEC, with an October listing the reported target.
Our take: the frontier fight is moving down the stack. Apple is suing over hardware, OpenAI is shipping hardware, and Anthropic is taping out hardware — because when every lab has a frontier model, the durable margins live in devices, chips, and deployment.
🌮 Fail of the week
Gemini 3.5 Pro: the do-over. Three issues ago it slipped "to July." Last issue it was hiding in limited preview. This week we learned why: Google reportedly scrapped the base model entirely and rebuilt it from scratch after finding structural failures the old architecture couldn't fix — a redo measured in months and hundreds of millions of dollars. The rebuilt model is reportedly targeting launch tomorrow, July 17. We'll believe it when we can prompt it.
⚡ Rapid fire
Every AI lab got graded — nobody passed — the Future of Life Institute scored nine frontier labs on 37 safety indicators. Best grade in the class: a C+ (Anthropic). The report's sharpest finding: the top three labs have all weakened or dropped their own red-line pledges.
Big publishing sued Google — Hachette, Cengage, Elsevier, and author Scott Turow filed a class action alleging Gemini trained on their books, including copies from "known pirate sources," and that Google abused its own Books platform to do it.
New York froze new AI data centers — the first statewide moratorium in the country: up to a year, anything over 50 megawatts, with 39 applications caught in the pause. Watch how many states copy the template.
Humanoid robots performed surgery — a first: teleoperated humanoids completed a gallbladder removal on animal models in a UC San Diego proof-of-concept. Your surgeon keeps their job; their hands might not.
ChatGPT is back on WhatsApp in Europe — because EU regulators ordered Meta to reopen the door to rival bots or face fines up to 10% of global revenue. Six months of blackout, undone by Brussels.
You know ChatGPT and Claude. Here are five more high-impact tools flying under most people's radar:
Browser Use — the open-source framework behind a huge share of the web-browsing agents shipping right now; point it at a site and it clicks, types, and extracts.
Gamma — describe what you need and get a finished deck, doc, or landing page; the conversational agent edits it with you like a designer on call.
Hedra — talking-avatar video that doesn't look haunted; give it a photo and a script and get a presenter.
Warp — the terminal rebuilt around AI: explains errors, writes commands, and runs agents where developers actually live.
Fireflies — the meeting notetaker that transcribes, summarizes, and answers questions across every call your team has ever had.
This week's spotlight: Tom Blomfield. He built Monzo into Britain's favorite bank and GoCardless before that, then spent three years at Y Combinator teaching founders how to survive. Now he's taking a leave from YC to join Anthropic's compute team — arguably the least glamorous, most consequential job in AI. Compute is the industry's rent, its biggest cost, and the thing that decides whether a $47B run rate ever becomes profit. Hiring a man who built a regulated bank to run it tells you exactly how Anthropic is thinking about the next chapter: less lab, more institution.
Big names in the AI Graveyard this week — including two from the company that just shipped its first gadget:
DALL·E 3 — for two years, the way millions of ChatGPT users made pictures; fully replaced by GPT Image and retired.
OpenAI Operator — OpenAI's first computer-use agent lasted seven months as a research preview before being absorbed into ChatGPT's agent mode.
Cursor — the fastest-growing dev tool in history (~$2.6B annualized revenue) ends its independent run, acquired by SpaceXAI. The editor lives on; the startup doesn't.
Sana Labs — Europe's admired AI learning platform, acquired by Workday for ~$1.1B.
Want the autopsy? We dug into what actually kills AI tools — funding, acquisition, or just a dead domain — in our data report.
The adults are grading now
For three years, AI safety ran on the honor system: labs wrote their own pledges, marked their own homework, and asked for trust. This week showed what happens when the honor system expires. An independent institute graded all nine frontier labs and handed out nothing better than a C+ — noting that the top three have quietly weakened the very red-line commitments they once volunteered. Sixteen Nobel laureates signed a warning that the world isn't ready for what's coming. The UN pushed for global guardrails against "catastrophic harm." And New York stopped waiting for anyone's pledge, freezing new hyperscale data centers until the rules catch up.
None of these actors can train a model. All of them can now shape who gets to. That's the shift worth noticing: oversight is no longer something the labs grant — it's something arriving from outside, with report cards, moratoriums, and subpoenas instead of blog posts.
The uncomfortable question for the industry is simple: if the self-imposed red lines were real, why did they need to be walked back the moment they became inconvenient? Whoever answers that well — in public, with receipts — earns the trust everyone else is about to lose.
Prompt of the week: grade your own AI stack
The safety institute gave every frontier lab a report card this week, and nobody beat a C+. Fair's fair — run the same exercise on the tools you depend on:
Grade my AI stack like a safety auditor. Here are the AI tools and vendors my team relies on: [list them]. For each one, grade A-F on: (1) data handling — what happens to everything we type into it; (2) reliability — outages, silent model swaps, or deprecations in the past year; (3) exit cost — how painful leaving would be. Then give my overall stack a GPA, name the single riskiest dependency, and tell me the one change that would raise my grade the most.
On the calendar:
AMD Advancing AI — July 22–23, San Francisco (Moscone West). MI450 accelerators and "Venice" EPYCs expected.
Black Hat USA 2026 — August 1–6, Las Vegas (Mandalay Bay). The agentic-ransomware era gets its first big stage.
Ai4 2026 — August 4–6, Las Vegas. North America's biggest industry AI gathering.
Until next time 👋
Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything.
— The ToolDirectory.AI team
