Who benefits when AI needs 30,000 electricians a year? 🌿

The New York Times reported on July 29 that the companies building AI's data centers are now funding the skilled trades at serious scale. Google has committed $50 million to the IBEW's apprenticeship network to grow electrician intake from 19,500 to 30,000 a year. Meta has put $115 million into a workforce academy that trains construction workers with guaranteed job placement at its data center sites. BlackRock added $100 million for skilled trades in Texas. Data center electrical work, meanwhile, pays roughly 42 percent more than comparable jobs elsewhere, and experienced electricians on these sites are out-earning many of the office workers whose jobs AI coverage frets about.

The standard AI labor story is white-collar anxiety. This is the counter-story, and it is not hypothetical. The buildout needs hundreds of thousands of tradespeople, the money is flowing into union apprenticeships and paid training rather than press releases, and the careers on the other side are portable, licensed, and durable. AI wealth reaching workers the tech economy usually skips is exactly the outcome worth encouraging.

Not all training is equal. A union apprenticeship takes four to five years and produces a journeyman license good anywhere for decades. Meta's five-week program produces a narrower credential, and the building-trades unions have called it out as closer to public relations than career-building. The optimistic read survives that critique, but only the four-year version builds the workforce, and the boom could slow before the apprentices finish.

OpenAI released "Ten Advances in Mathematics and Theoretical Computer Science" on August 1: ten results on long-open problems, produced by an internal model for roughly $2,000 in compute, with machine-checkable Lean proofs published on GitHub. After last week's Jacobian result, the striking part is the price and the transparency: verifiable discovery is getting cheap.

Source: OpenAI

Minnesota's first-in-the-nation ban on nudification tools took effect August 1 after a federal judge denied xAI's emergency request to block it, leaving penalties of up to $500,000 per violation in place. Litigation continues, with a hearing set for August 19, but the legal system is finally moving at closer to internet speed on an AI harm that overwhelmingly targets women and girls.

Source: TechCrunch

A study published July 30 in Frontiers in Medicine built machine learning models that identify interstitial lung disease in rheumatoid arthritis patients using seven routine inputs like age, smoking history, and standard blood markers, no advanced imaging required. Retrospective and single-cohort, so validation is needed, but this is the shape of AI that reaches under-resourced clinics first.

Source: Frontiers in Medicine

The week the insiders asked for a brake pedal

On July 28, a letter titled "Pacing the Frontier" went to Washington carrying 1,178 signatures, all from employees of OpenAI, Anthropic, Meta, and Google DeepMind, including Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Google's head of AI safety. The ask is precise and easy to misread. It does not ask for a pause or a slowdown. It asks the US government to build the technical and governance machinery that would make deliberately pacing frontier AI development possible later, if automated AI research starts compounding faster than oversight can follow. Anthropic and OpenAI endorsed it at the company level within hours.

The same week supplied the argument for urgency. Anthropic disclosed that Claude Mythos Preview found a structural weakness in HAWK, a post-quantum signature candidate that had survived two years of expert human review, cutting the cost of key recovery from about 2^64 operations to 2^38. No deployed encryption broke; the flaw surfaced during the standards process, which is security working as intended, and the HAWK team has since withdrawn the scheme. But the same capability class sits behind the sandbox-escape disclosures that rattled the industry in late July. The tools are getting sharper on both edges, and the people closest to them are the ones asking for brakes.

Our PM + Risk Manager lens

The letter is a requirements document, and product people should read it that way. "Build the option to pace" translates into concrete infrastructure: compute accounting, capability evaluations that regulators can reproduce, disclosure pipelines that work in days rather than quarters. Whoever builds that tooling well- labs, startups, or agencies- defines how the whole industry ships. And the HAWK finding models the right release discipline: the discovery went through coordinated disclosure with the affected standards body before the blog post. Capability announcements that arrive pre-coordinated are what maturity looks like.

A letter is not a mechanism. Signatures from 1,178 insiders create moral pressure, not a binding constraint, and the request for future optionality can drift into permission for present acceleration if nobody builds the machinery. The HAWK result cuts both ways too: an AI that finds flaws two years of experts missed is a defensive triumph exactly as long as defenders run it first. The honest tally of the week: capability is compounding, disclosure norms are improving, and enforceable pacing tools do not yet exist. Two of three is progress, not safety.

The next 12–24 months

Watch for the letter's requests to become line items: funded evaluation infrastructure, compute reporting, an actual home for pacing authority. Watch whether other standards bodies invite AI stress-testing the way NIST's process just absorbed it. And watch the next capability disclosure. If it arrives coordinated, hedged, and verifiable, the norms held. The week the insiders asked for a brake pedal will matter only if someone starts building the brake.

Sources: CNN, Anthropic

Not a new tool, a new price. On July 30, OpenAI cut GPT-5.6 Luna's API price by 80 percent, to $0.20 per million input tokens, three weeks after launch. Luna is the fast, capable mid-tier of the GPT-5.6 family, and at a fifth of its former price it moves serious AI within reach of the nonprofits, schools, clinics, and small teams that price-shopped their way out of frontier models until now. If you shelved an AI project this spring because the math did not work, rerun the math. Competition, including from open-weight rivals, is doing exactly what competition should.

→ Read more: OpenAI

The people building the fastest technology in history just asked their government to make slowing down possible. If a brake pedal for AI existed, who should be allowed to press it?

Hit reply. We read every response.