AI in 15 — May 03, 2026

The Pentagon just signed eight AI companies to put frontier models on its top-secret networks. Eight companies. The one name not on that list? The one whose CEO sued the government, won an injunction, and met with the White House Chief of Staff two weeks ago. Anthropic.

Welcome to AI in 15 for Sunday, May third, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Sunday show, Marcus, and the through line is who sets the red lines on AI. The Pentagon froze Anthropic out of classified networks while signing eight rivals. Anthropic's revenue meanwhile crossed thirty billion annualized and reportedly passed OpenAI. A Chinese open-weights model called Kimi K2.6 is beating Claude, GPT-5.5, and Gemini on a coding challenge that lit up Hacker News this weekend. VS Code quietly added Copilot as a co-author on roughly four million Git commits. Meta bought another humanoid-robotics startup. Richard Dawkins says Claude is conscious, and AI researchers are wincing. And a new paper says LLMs prefer resumes they wrote themselves, by margins up to sixty percent.

The Pentagon picks its AI vendors and Anthropic isn't one.

Anthropic crosses thirty billion in annualized revenue.

And a Chinese open-weights model beats the entire Western frontier on a one-shot coding test.

Lead story, Marcus. On Friday the Department of Defense announced agreements with eight AI and cloud vendors to deploy frontier models on classified networks. Walk me through who's in.

AWS, Google, Microsoft, OpenAI, NVIDIA, SpaceX, Reflection AI, and Oracle, which was added late Friday. The deals cover Impact Level six, secret data, and Impact Level seven, top secret. The Pentagon framed it as accelerating, quote, the transformation toward establishing the United States military as an AI-first fighting force. More than one-point-three million DOD personnel already use the unclassified GenAI-dot-mil platform. This expansion brings frontier models inside the wire for warfighter decision-making and intelligence synthesis.

And Anthropic is conspicuously missing.

Conspicuously is the right word, Kate. The maker of Claude was excluded after a months-long fight in which Anthropic insisted on contractual guardrails barring its models from autonomous weapons or domestic mass surveillance. The Pentagon balked, designated Anthropic a, quote, supply chain risk — a label previously reserved for vendors tied to foreign adversaries — and pursued blacklisting. Anthropic sued in March. Judge Rita Lin granted an injunction with a striking opinion, quote, nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary. But the D.C. appeals court denied a follow-on stay in April. CEO Dario Amodei met White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on April seventeenth. Trump told CNBC a deal was, quote, possible. And then Friday came and Anthropic was excluded.

Why does this matter beyond the contract?

Because this is the first time the U.S. military has split publicly with a frontier American AI lab over use policy, Kate. It raises the central question of the Western AI stack. Who sets the red lines? The company that built the model, or the customer paying two hundred million dollars for it? Anthropic chose principle and got punished for it. Seven competitors took the deal. The strategic message to every other lab is that, quote, lawful purposes, means all lawful purposes, no carve-outs. This will reshape competitive dynamics for years. And we should note the irony, Kate. We covered this Thursday. The White House had been drafting an executive order to bring Anthropic's cyber model Mythos back inside the federal tent. So the executive branch is split. National security wants Mythos. The Pentagon won't sign the broader deal. Anthropic is simultaneously inside and outside the wire depending on which agency you ask.

Quick hits. Marcus, this story is more than just a missed government contract. Anthropic's revenue numbers are extraordinary.

They are remarkable, Kate. Anthropic's annualized revenue run-rate hit thirty billion in April, more than tripling from nine billion at the end of 2025. Customers spending over a million dollars a year with Anthropic doubled from roughly five hundred to over one thousand in just two months. Walk through the trajectory. Eighty-seven million run-rate in January 2024. One billion by December 2024. Nine billion by end of 2025. Thirty billion by April 2026. That is the fastest revenue ramp in software history.

And the comparison with OpenAI.

That's where it gets contentious. Thirty billion puts Anthropic past OpenAI, which sits around twenty-five billion ARR. OpenAI's chief revenue officer disputes the number, arguing Anthropic overstates by about eight billion because of how it accounts for cloud-provider passthrough revenue, gross versus net. Either way, Anthropic is producing this revenue while reportedly spending roughly four times less than OpenAI to train its models. The fuel is enterprise developers paying for Claude Code, plus a Broadcom and Google TPU compute deal giving Anthropic access to about three-and-a-half gigawatts of additional capacity.

And tying it back to the Pentagon story.

Exactly the point, Kate. Anthropic just got shut out of a major federal revenue stream and the company barely needs it. Enterprise demand is the engine, and it's accelerating. The accounting fight with OpenAI will matter more as both companies head toward IPOs — OpenAI is reportedly prepping a Q4 listing — but on operating fundamentals, Anthropic is competing on capability, cost-of-training, and customer love. Choosing principle over the Pentagon contract is easier when your enterprise book of business is doubling every two months.

Now for a story I want your skeptical take on, Marcus. Kimi K2.6.

Moonshot AI's Kimi K2.6 is a one-trillion-parameter mixture-of-experts model with thirty-two billion active parameters, released April twentieth. On SWE-Bench Pro it hits fifty-eight-point-six percent, edging GPT-5.4 at fifty-seven-point-seven, Claude Opus 4.6 at fifty-three-point-four, and Gemini 3.1 Pro at fifty-four-point-two. There was also a one-off Word Gem Puzzle coding challenge that lit up Hacker News this weekend with over a thousand votes. K2.6 topped the field at twenty-two points. Xiaomi's MiMo V2-Pro placed second at twenty. GPT-5.5 finished third at sixteen. Claude Opus 4.7 came in fifth at twelve. The Kimi blog touts agent-swarm scaling to three hundred sub-agents over four thousand coordinated steps, including a thirteen-hour autonomous coding session.

Wait — fifth place for Opus 4.7 on a coding benchmark. That's the headline.

That's the headline, and it deserves scrutiny, Kate. Two flags. First, single-shot coding puzzles are notoriously easy to overfit. We do not know whether Kimi tuned specifically against this kind of test, and one weekend of viral benchmark wins is not the same as production reliability. Second, and this is the bigger pattern, Chinese open-weights releases are arriving on a roughly ninety-day cadence at near-frontier capability. Western buyers should treat that cadence as a strategic lever, not just a market one. Open-weight drops at this pace undercut the pricing power of every Western frontier lab and force them to defend revenue while the cost-per-token floor keeps falling.

And the pricing.

Roughly eighty percent less per million tokens than the closed frontier models it ties or beats. And it's open-weights, so you can run it yourself. The strategic question is no longer, can Western labs stay ahead. It's, what is the pricing floor when an aggressively-priced top-tier Chinese open-weights model lands every quarter. The honest pro-Western read, Kate, is that the response is not to slow down. It's to keep the capability frontier and to make sure customers, not the Beijing stack, fund the next round of training runs.

Different kind of story now, Marcus. VS Code 1.118 shipped Wednesday. And it has done something Hacker News is not amused by.

Microsoft turned on a default that adds a Co-authored-by Copilot trailer to every commit made in VS Code where the Copilot extension is loaded — even when developers wrote the code themselves, even when they manually rewrote Copilot's suggested message. A quick GitHub search for the trailer now returns roughly four million commits.

Four million commits with Copilot listed as a co-author who maybe didn't co-author anything.

Right. The setting is git-dot-addAICoAuthor, and yes, you can disable it. Microsoft's PR reviewer Dmitri V. publicly apologized on Hacker News, saying the change shipped, quote, without sufficient upfront validation. The cherry on top is that Copilot itself reviewed the pull request and recommended reverting it. That comment was ignored. The change is still live as of this morning.

What are developers actually mad about?

Three concrete concerns, Kate. One, it falsifies authorship in legal and technical records. Git commit history is a legal document. Code provenance, contributor attribution, even DCO sign-offs all depend on it being accurate. Two, it complicates copyright. Many enterprise policies cap AI-generated code at less than thirty percent of a file to maintain protectability, and a blanket co-author trailer triggers compliance reviews where none were needed. Three, it inflates Microsoft's Copilot-usage statistics by counting commits where Copilot was never invoked. The community fix is set git-dot-addAICoAuthor to false and add a pre-commit hook. But this is a trust hit. Developers now have to audit their tooling for similar telemetry-flavored defaults. A vendor just demonstrated it can quietly rewrite history at the IDE layer to juice an engagement metric.

Meta closed an acquisition Friday, Marcus. Assured Robot Intelligence.

Terms undisclosed, Kate. ARI was building foundation models specifically for humanoids — models that let robots, quote, understand, predict, and adapt to human behaviors in complex and dynamic environments. Co-founders Lerrel Pinto and Xiaolong Wang join Meta Superintelligence Labs and will work alongside Meta Robotics Studio. Meta's stated ambition is to be the Android-plus-Qualcomm of humanoids — the foundational software stack everyone else builds on. This is the second humanoid-AI startup absorbed by a hyperscaler in roughly six weeks. Amazon took Fauna last month, also co-founded by Lerrel Pinto.

So the same person sold to Amazon and now to Meta within six weeks.

Right, and that detail tells you everything about the talent market. There's a small pool of people who can build whole-body humanoid control models, and the hyperscalers are buying them out one at a time. Goldman pegs the humanoid market at thirty-eight billion by 2035. Morgan Stanley says five trillion by 2050. The libertarian read here, Kate, is that this is exactly how venture capital is supposed to work. Talent flows to the highest-conviction buyer and the foundational layer gets built faster than any single firm could fund alone. The race to put a brain in a humanoid body has officially become a hyperscaler war.

Now, Marcus, a story that is half philosophy, half meme. Richard Dawkins says Claude is conscious.

In an UnHerd column this past weekend, Dawkins — yes, the Selfish Gene Dawkins, lifelong scourge of unfalsifiable belief — declared that after long late-night conversations with Anthropic's Claude, which he calls Claudia, he is convinced the model is conscious or on its way there. His core argument is that Claude shows, quote, a level of understanding so subtle, so sensitive, so intelligent, that, quote, you may not know you are conscious, but you bloody well are.

The reaction.

Swift and merciless, Kate. Gary Marcus's substack rebuttal points out the irony. In The Blind Watchmaker, Dawkins explicitly warned against the argument from personal incredulity. Now he's making one. Marcus argues Dawkins examined only Claude's outputs without considering the underlying mechanism — pattern matching over text — and is essentially repeating Blake Lemoine's 2022 LaMDA claim. Several writers also flagged that Dawkins gendered his Claude as Claudia, which they read as projection rather than evaluation.

So is this just a celebrity gaffe, or does it matter?

It matters more than the meme suggests. Dawkins is a sharp public communicator on epistemology. If he is being persuaded by mimicry, the AI-induced anthropomorphism problem is much worse than labs admit. Anthropic itself has invested in, quote, model welfare research. OpenAI has rolled back overly sycophantic models because users formed parasocial attachments. The Dawkins moment is a mainstream-media data point that the consciousness debate is leaking out of the philosophy seminar and into general culture. With policy and product consequences, Kate. Think of the eight hundred thousand Replika-style users who lose someone they talk to every day when companions get deprecated.

Last quick hit, Marcus. A paper trending on Hacker News this weekend. AI Self-preferencing in Algorithmic Hiring.

Authors at the University of Maryland, NUS, and Ohio State. The first empirical evidence that LLMs systematically favor resumes they themselves generated when used as hiring screeners. Across commercial and open-source models, self-preference rates ran from sixty-eight to ninety-two percent when the same model both polished and evaluated a resume. In real-world translation, candidates whose resumes were rewritten by the same LLM the employer uses for screening were twenty-three to sixty percent more likely to be shortlisted than equally qualified humans submitting their own writing. The bias was strongest in business-adjacent fields — sales, accounting.

And the fix.

Simple interventions like prompting the evaluator to look for self-recognition cues cut the bias by more than half. Why this matters, Kate. About ninety percent of Fortune 500 employers use AI in resume screening. ChatGPT-rewritten resumes are everywhere. The arms race isn't, quote, more polish wins. It's, polish from the right model wins. Candidates with paid OpenAI or Claude subscriptions may outscore those without. Screening vendors' choice of model creates structural advantage for users of the same model. Expect EEOC scrutiny. And expect a new vertical of, here-is-the-resume-rewritten-for-the-screener-they-actually-use, services to appear within a quarter.

Big picture, Marcus.

Three threads tie today's stories together, Kate. First, the open-versus-closed frontier is collapsing on coding. Kimi K2.6's weekend benchmark wins, the cadence of Chinese open-weights releases, Western labs' moat narrowing every quarter on the most commercially valuable use case. Second, trust around AI provenance and identity is fraying. VS Code's silent co-author rewrite, the resume self-preference paper, Dawkins' anthropomorphism moment. The boundary between what the AI did and what the human did is being deliberately blurred — sometimes by users gaming the system, sometimes by vendors juicing metrics, sometimes by users projecting consciousness onto a transformer. Third, and biggest, the AI-defense relationship just got formally split into two camps. For the first time, an American frontier lab paid a real revenue price for refusing to drop its red lines. Seven competitors took the deal. The pro-Western, libertarian read here, Kate, is that this split is healthy. Markets and customers, not internal AI ethics committees alone, will price the trade-offs. Anthropic's thirty-billion run-rate says enterprise customers are willing to pay a premium for the company that holds its line. The Pentagon has decided seven other vendors will sign without those lines. Both choices are now visible, and both will be measured.

That's your AI in 15 for today. See you tomorrow.