AI Actually

Issue No. 7 · Wednesday · May 6, 2026

Welcome back. The U.S. government had a busy week deciding whether AI is a weapon, a threat, a vendor, or a guy you call about an issue with your printer. The verdict so far: yes.

Meanwhile, the two biggest AI labs in the world quietly admitted that what they actually sell isn’t software, it’s the embeded capabilities - does this sound awafully familier, my consultant friends?!

And a 2024 AI model beat two emergency room doctors at diagnosing real patients, raising the question of whether the model deserves a white coat or just a really long apology from somebody.

Let’s get into it.

---

Washington can’t decide if AI is a national asset or a national problem

The U.S. government had a busy week reorganizing its relationship with AI. Two moves, in opposite directions, hours apart.

Move one: The Pentagon added eight AI companies to its classified networks — SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Amazon Web Services, and Oracle. The deal is part of what the Department of War is calling an “AI-first fighting force.” Conspicuously not on the list: Anthropic, the company that makes Claude.

This is the same Anthropic that the White House has spent the past few weeks fighting with over its Mythos model — a system so good at finding software vulnerabilities that Anthropic refused to release it publicly. The Pentagon says Anthropic remains a “supply-chain risk.” It also called Mythos a “separate national security moment.” Both things at once. Reader, the U.S. government would like Mythos. It would just prefer not to say please.

(We covered the earlier round of this feud in Sunday’s issue. This is the next chapter.)

Move two: The White House is reportedly drafting an executive order that would require AI companies to submit new models for government review before they ship. A working group of tech executives and government officials would do the vetting. This is a sharp reversal — the same administration spent its first months in office promising deregulation. Cybersecurity concerns over Mythos appear to be what changed minds.

Why it matters: A year ago, “AI policy” meant vague speeches about innovation. Now it means classified contracts, blacklists, and pre-release inspections. The U.S. government has decided AI is too important to ignore and too dangerous to wave through. Both at once. Expect more of this — not less.

Read the source (Pentagon contracts) → · Read the source (White House review) →

---

Anthropic and OpenAI accidentally became consulting firms on the same day

On Monday, within hours of each other, the two biggest AI labs in the world both announced they were getting into the consulting business. Not “partnering with” consulting firms. Becoming them.

• Anthropic: a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. Backers include Apollo, GIC, Sequoia, General Atlantic, and Leonard Green.

• OpenAI: “The Deployment Company,” a $10 billion venture with TPG, Brookfield, Bain, Advent, and SoftBank.

The same week. The same structure. The same realization.

Which is this: the bottleneck in enterprise AI isn’t the model anymore. It’s that most companies have no idea what to do with it. The labs have spent two years making Claude and ChatGPT smarter. The companies that bought subscriptions have spent two years watching their employees use them to summarize emails. Selling more software wasn’t going to fix that. Sending humans to install it might.

For consulting firms who told themselves they were safe because they owned the client relationships, the trust, and access to client data. The labs didn’t compete with consultants for clients. They went one floor up — to the firms that own the clients.

Why it matters: This is the moment AI stopped being a product and started rewriting a whole industry. Consultants’ “we own the client” argument just got a counter-argument: “we own the people who own the client.” McKinsey's quarterly review meeting is going to be tense.

Read the source → · Read my full take→

---

AI beat ER doctors. Again.

A Harvard study published in Science this week put OpenAI’s o1-preview — a reasoning model from 2024, already two generations old — through 76 real emergency room cases. The model out-diagnosed two attending physicians.

The numbers, in case you want them at a dinner party:

• AI initial-triage accuracy: 67%

• Physician 1: 55%

• Physician 2: 50%

Reviewers, blinded to source, couldn’t tell which diagnoses came from the humans and which came from the model. In one case, the AI flagged a rare flesh-eating infection in a transplant patient 12 to 24 hours before the treating doctor caught it.

This is the second AI-beats-doctors result in two weeks. We covered the Mayo Clinic study on Sunday — their model spotted pancreatic cancer up to three years before diagnosis on routine CT scans. Different problem, same direction. Both teams stopped short of recommending clinical deployment, the way researchers always do, with the air of people who know their email is about to fill up.

Why it matters: The AI in the Harvard study isn’t even current. If a two-generation-old model already beats attending physicians at the hardest part of their job — reasoning under uncertainty with sparse information — the question isn’t whether AI shows up in the exam room. It’s who gets sued first when a doctor doesn’t check with one.

Read the source →

---

That’s it for this week. If something here didn’t make sense, or there’s a story you want explained, hit reply. The best questions become next week’s issue.

See you Sunday.