Apple Sued OpenAI: This Week in AI (Jul 10 -17)

I read Apple’s lawsuit against OpenAI on Friday night instead of watching the football I had queued up. Forty-some pages about two ex-Apple guys, a $6.5 billion acquisition, and a company accusing its own former staff of walking out the door with the blueprints. Genuinely more entertaining than most Netflix originals.

Same week, a lab out of China put out a model with 2.8 trillion parameters, gave away the weights for free, and by Friday morning people who get paid to know better were saying it closes the gap with GPT and Claude.

Both things happened in the same seven days. That is not a coincidence, that is the whole story right now. Some companies are pulling up the drawbridge. Others are giving away the thing everyone else is charging for. Here is what actually happened, and what I think it means if you are building anything on top of this stuff.

Apple sued OpenAI, and it is not about code

Apple filed suit on July 10. It names two former Apple hardware leads, now working at io, the hardware unit OpenAI built around Jony Ive. The claim: they took CAD files, supplier relationships, even Apple’s own internal security playbook, on their way out. Apple’s own filing says over 400 of its former employees now work at OpenAI.[9to5Mac]

When a company sues over departing staff instead of just out-building the rival that hired them, it is telling you where it thinks the real value lives. It is not the model. It is the fifty people who know how to build the thing.

Anthropic just claimed an entire profession

Claude for Teachers launched July 14. Free premium Claude access for every verified K-12 teacher in the US, mapped to actual state curricula, built to be FERPA compliant from day one. [Anthropic]

Same week, a checklist going around showed Claude already covering Code, Design, Finance, Science, and now Teaching, with Marketing and Sales still empty. The labs are not shipping features anymore. They are shipping a branded front door for every job that exists, one at a time. If you work in one of the categories still empty, that is either a warning or an opening, depending on how you use it.

Claude’s chat window just turned into something you can build on

Claude Code artifacts can now call your own MCP connectors. Which means a dashboard Claude draws for you can also go fetch live data and act on it, for whoever is looking at it, with no deploy step.

I keep coming back to this one line: the scarce skill used to be “can you code it.” Now it is “can you actually say what you want.” That is a bigger shift than it sounds like on a Tuesday.

Mira Murati’s lab gave away its first model

Thinking Machines shipped Inkling on July 15. Full weights, free, on Hugging Face. 975 billion parameters, 41 billion active, a million tokens of context. [Thinking Machines]

It is not the strongest model out there, open or closed, and the lab says so itself. That is the point. It was built to be taken apart and rebuilt, not to win a leaderboard screenshot.

NotebookLM got a new name and an actual job

Google renamed NotebookLM to Gemini Notebook on July 16 and gave it a secure cloud computer, so it can now write and run code against your own sources instead of just summarizing them back to you. Rolling into the Gemini app now, AI Mode in Search soon. Google says over 30 million people already use it. ([Google, The Keyword]

The thing you used to dump your reading into is starting to think back at you. Worth watching what that does to how you use it.

A free model out of China just closed the gap

Moonshot AI released Kimi K3 on July 17. 2.8 trillion parameters, which the lab calls the world’s largest open weight model. Top story on Hacker News the day it dropped, 1,254 points. Early testing already has people saying it is the closest an open model has gotten to the US frontier.

Free, and closing the distance. That combination is exactly what should worry anyone whose pricing model depends on staying ahead.

X and YouTube both stopped paying for copies

On the same day, with zero coordination between them, X started routing monetized impressions to the original poster no matter what watermark or edit got slapped on top, catching 1.5 million stolen posts this cycle and handing back over a million dollars to the people who actually made the thing. YouTube, hours later, named generic and templated content as a new target for demonetization. [Social Media Today on YouTube]

Two platforms, same day, same instinct. Stop paying for the reshare. If your content strategy leans on repackaging someone else’s stuff, this is the week that got harder.

Four things worth adding to your stack this week

Not news, just useful. All free, all real, all checked before I put them here.

Guizang PPT Skill - Point it at any brief and it builds a genuinely presentation-ready HTML slide deck, two full visual systems to pick from, no design tool needed. 21.6K stars on GitHub, the most-used content skill on my radar right now. [GitHub]

Hallmark - Stops your AI-built site from reading like an AI built it. Picks a real structure, applies one of 20 themes, runs 57 checks against the tells that make a page look templated. Went from 8,962 stars to 11,500 in about a day this week. [GitHub]

LM Studio Bionic - A coding and document agent that runs entirely on your own machine, with local voice dictation so nothing you say ever leaves your laptop. [LM Studio]

specification.website - An open checklist, built by the guy who founded Yoast, for whether your own site is actually legible to the AI crawlers now reading it on your behalf. If you have not asked this question about your own site yet, this week is a good week to. [GitHub]

So that is the week. One side locking things down, the other side giving the frontier away for free. Both bets are real, and I do not think you get to sit out either one much longer.

If you want the swipeable version with all seven stories and the four tools laid out as a carousel, it is on Instagram too.

See you next week.