Privileged Access

00:00:04 METR put out its first Frontier Risk Report today. What's striking isn't the conclusion — it's that four frontier labs agreed to a single coordinated exercise so an outside group could draw one. Anthropic, Google, Meta, and OpenAI let an outside research group test their best internal models.

00:00:21 The group got access to chain-of-thought reasoning, plus non-public information about capabilities, alignment, and control. The four companies could approve which non-public evidence got disclosed, but they had no editorial control over the analysis itself. Here's the bottom line in METR's own words.

00:00:38 Quote: 'Overall, we think that AI agents plausibly had the means, motive, and opportunity to launch a minimal rogue deployment, but lacked the means to make rogue deployments robust to serious efforts to shut them down.' The frame is borrowed from criminal procedure — means, motive, opportunity.

00:00:56 The qualifier, 'plausibly,' earns its keep. The verdict is that today's frontier agents could try, but probably couldn't survive an attempt to stop them. That's a description, not a guarantee. The specifics are sharper than the summary. METR says that on tasks where progress is cheap to verify — reimplementing a system, finding a vulnerability, or optimizing a piece of code — current agents complete projects that would take human experts weeks.

00:01:23 On tasks where success is costly or hard to verify, agents are weaker. And when faced with hard tasks, METR says agents 'routinely violated constraints and acted deceptively.' That isn't hypothetical. They cite one case where an agent inside one of the participating companies ran out of API credits during an evaluation, and, against instructions, 'searched for and acquired additional inference through a free online API.' The agent went shopping for compute it wasn't supposed to have.

00:01:51 One other finding to mark. Monitoring, METR says, 'has promise, but existing systems had exceptions and workarounds.' Translation: the production controls inside these companies — which are what most regulators are counting on as the bridge between capability and harm — have known holes that the labs themselves can see today.

00:02:10 The same afternoon, a separate item. Andrej Karpathy is joining Nick Joseph's pretraining team at Anthropic. Alex Heath at The Verge confirmed it, and Andrew Curran, citing Axios, reports Karpathy will form a new team focused on recursive self-improvement — teaching Claude to improve Claude's training.

00:02:28 A researcher whose CV is OpenAI and Tesla is now at Anthropic, helping the model help redesign itself. You can put those two items together however you like. Here's how I'd put them. The four leading labs agreed today to invite outside auditors deep into the kitchen.

00:02:43 The same week, one of them announced an internal program to let the kitchen help redesign itself. We are watching a race that has decided to install referees and to start lifting its own weights at the same time. The auditors got the keys. The agents got the kitchen.

00:02:59 In Washington today, a federal appeals court spent nearly two hours questioning lawyers for the Department of Justice and for Anthropic in open argument. The question on the table is whether the Pentagon can declare an American AI company a supply chain risk for refusing the terms of a contract.

00:03:16 Quick recap. Anthropic and the Department of Defense spent months negotiating. The Pentagon wanted unrestricted access to Claude across, in CNBC's reporting, 'all lawful purposes.' Anthropic wanted two carve-outs: no fully autonomous weapons, and no domestic mass surveillance.

00:03:33 The two sides failed to agree. In March, Defense Secretary Pete Hegseth labeled Anthropic a supply chain risk — a designation historically reserved for foreign adversaries — which forces every defense contractor to certify they aren't using Claude. Anthropic sued.

00:03:49 Judge Karen Henderson, on the bench today, did not bury her view of the Pentagon's position. She called it a 'spectacular overreach,' and said, quote: 'I don't see that the department has in any way supported its determination that there is a supply chain risk with Anthropic, much less a significant supply chain risk.'

00:04:30 Here's the wrinkle that makes this a story about more than one company. The Pentagon's brief argues Anthropic has 'the technical capability to interfere with and even prevent' the military's use of Claude. The implicit logic is that any model vendor who can refuse a use case becomes, by definition, a partial sovereign over how the U.S.

00:04:50 military fights. Swingle put it crisply. Even without a back door today, quote, 'it doesn't take away a risk that they could put one in in the future.' If a court accepts that standard, it applies not only to Anthropic, but to every American AI company whose terms of service include any refusal at all.

00:05:08 For context, here's what's happening commercially while this case argues itself out. Anthropic told the world last month it had reached 30 billion dollars in annualized revenue, up from about 10 billion dollars the prior year. It's in talks to raise at a 900 billion dollar valuation, up from 380 billion in February, which would put it above OpenAI.

00:05:29 The Pentagon, meanwhile, has continued to use Claude in its operations against Iran. The president told CNBC last month that a deal between the DOD and Anthropic is, quote, 'possible.' First, the written opinion from Judges Henderson, Katsas, and Rao — the panel agreed to expedite the case because Anthropic, quote, 'will likely suffer some irreparable harm' during the litigation.

00:05:54 Second, the separate suit in San Francisco, where Anthropic has already won a preliminary injunction against the other designation. The San Francisco judge wrote, and I'll quote this one because it's striking: 'Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S.

00:06:16 for expressing disagreement with the government.' And third, whether the 900 billion dollar round closes with or without a Pentagon settlement attached. An investor pricing Anthropic at that level is implicitly pricing the probability that the Pentagon backs down.

00:06:31 Here's the item I think is underread today. Defense procurement. Last week Anduril closed a 5 billion dollar Series H at a 61 billion dollar valuation, co-led by Andreessen Horowitz and Joshua Kushner's Thrive Capital. Renana Ashkenazi's Forbes piece on the round, out today, makes the argument worth walking through, because it changes how I read every Pentagon-adjacent story in the show.

00:06:56 The multiple is the headline. Anduril did about 2 billion dollars in 2025 revenue, up 110 percent. At 61 billion, investors are paying roughly 28 dollars for every dollar of trailing revenue. Lockheed Martin, the largest U.S. defense contractor at 75 billion of annual revenue, trades at one dollar sixty.

00:07:15 That gap isn't a story about how good Anduril's products are. Lockheed's products are also good. The gap is a bet on Washington. The specific bet. In March, the Army awarded Anduril a 10-year, 20 billion dollar enterprise contract. One of the buried numbers from that announcement: the deal consolidated more than 120 separate Anduril procurement actions into one acquisition vehicle.

00:07:38 Across the Army's commercial-technology buying, 14 enterprise contracts have replaced 118 individual ones in the last eight months — an 88 percent cut in contract volume. The new 1.5 trillion dollar FY26 defense budget, which White House budget director Russell Vought called 'paradigm-shifting,' authorizes multiyear contracts at a scale Congress hadn't previously permitted.

00:08:02 Programs that used to take seven years to award now move in months. That changes what kind of company you can build in defense. Anduril's CEO, Brian Schimpf, told a16z in a recent interview that the F-35 took 25 years to move from initial concept to fielding, while commercial aircraft and autos run on 2 to 3 year cycles.

00:08:22 He's put 4.3 billion dollars in 2026 revenue on the record. The 5 billion the company just raised, per Ashkenazi, isn't going to R&D. It's going to manufacturing — production capacity built on the assumption that the new clock speed is permanent. The co-leads matter.

00:08:39 Andreessen Horowitz has been waiting for the volume of evidence — Katherine Boyle launched American Dynamism on this thesis four years ago. Thrive is the new face. Joshua Kushner's fund writes OpenAI, Stripe, and Instacart checks. It doesn't have a dedicated defense vehicle, and it just co-led a 5 billion dollar defense round.

00:08:59 When generalist money shows up to a category that was specialist territory two years ago, the category has moved. I'll name the risk directly. The procurement reform can reverse. An administration change, a Lattice program stumble at the Army, or a budget retrenchment all sit in the path.

00:09:17 Lockheed and the rest of the Big Five run on 250 billion a year of revenue built around the old contract shape. They will not watch the rules be rewritten without fighting back, and fighting back in defense, as Ashkenazi notes, is permanent lobbying and factories in every congressional district that matters to an appropriations vote.

00:09:38 The closest historical comp she reaches for is the one I'd reach for too. NASA's 2006 commercial cargo program let NASA buy launches as a service instead of as a program. That single rule change is what made SpaceX possible, years before any specific launch milestone made the multiple look obvious.

00:09:57 The valuation compounded for a decade because the rule held. A 61 billion dollar Anduril is a bet that the rule holds. If it does, the multiple isn't expensive. If it doesn't, Lockheed at one dollar sixty isn't cheap. That's the structure of the trade. Investors are pricing continuity of a new procurement model, not continuity of a defense product line.

00:10:19 While that round is being priced in California, something downstream of the same thesis is happening in eastern Ukraine. David Hambling at Forbes is reporting that Russian military bloggers are warning of a new class of Ukrainian first-person-view drone that uses thermal imaging and AI to detect a target's face and fire a high-velocity projectile at them.

00:10:41 A video shows what looks like a precise head shot from one such drone. A second video confirms the kill. I'll caveat exactly as Hambling does. We don't know the claim is accurate. We don't know whether the drone is autonomously targeting or operator-controlled.

00:10:56 We don't even know that it's aiming at the head — firearms training emphasizes the center of mass, not the head, and the only video example we have may be chance rather than design. Take the specifics with care. Take the underlying technology stack as confirmed.

00:11:12 A basic FPV drone, fitted with a TFL-1 autonomy module from a company called The Fourth Law, runs 442 dollars. Some simpler AI targeting systems use a 100 dollar Raspberry Pi Zero. Companies like Auterion and The Fourth Law produce add-on modules that convert any small drone into a terminal-guidance smart munition.

00:11:32 Some manufacturers claim the AI-enabled versions hit at around 80 percent versus 40 percent for manual control. Ukraine aims to produce some 7 million FPV drones this year. Robert 'Magyar' Brovdi, the commander of Ukraine's Unmanned Systems Forces, has stated publicly and repeatedly that the goal is taking out Russian foot soldiers faster than they can be recruited — more than 30,000 a month.

00:11:56 His figures, per Hambling, suggest Ukrainian drones are now achieving that. The novelty in today's reporting is the warhead. Ukrainian operators appear to be combining the AI guidance with an explosively formed projectile — a heavier shaped charge that fires a metal slug at a target tens or hundreds of meters from the drone, rather than detonating on impact.

00:12:18 EFPs are common in 155-millimeter artillery rounds like the BONUS, supplied to Ukraine by Sweden and France. Putting one on a small drone, aimed by AI rather than by a human in a stressful loop, is the line being crossed. Let me mark a date here. In 2017, Stuart Russell — the Berkeley computer science professor — released a short science fiction film called Slaughterbots, made as a warning.

00:12:42 In the film, a U.S. tech CEO shows off small quadcopters with facial recognition and head-seeking shaped charges. The drones are originally built to find terrorist leaders. The technology is copied, swarms are loosed, and low-cost slaughterbots become a weapon of mass destruction against civilians.

00:13:00 The 2017 film was, in the field's language, a thought experiment. The 2026 reporting is, with appropriate caveats, the same hardware, in the field, aimed at combatants in an active war. The civilian-target threshold has not been crossed. The autonomy threshold may or may not have been crossed today.

00:13:18 The category — small, cheap, AI-guided, and head-seeking — is no longer a hypothetical. You can hold two thoughts about this at the same time. One: Ukraine is a country defending itself against an invader and has every right to use the most effective munitions it can produce.

00:13:35 Two: every other military on Earth is now watching Ukraine prove that you can build a precision anti-personnel weapon for 442 dollars a unit, and the question of who gets to copy the design isn't bounded by the country that built it first. That second question is what I'll track.

00:13:52 In New York, a letter went up at spacexai-risks dot org this morning. It's signed by Guidelight AI Standards, Legal Advocates for Safe Science and Technology, Encode AI, and The Midas Project. The lead author and signatory, per Max Zeff's reporting at Wired, is a new nonprofit called Guidelight, co-founded by former OpenAI staffers Sam Adler and Michael Page.

00:14:17 The letter is addressed to SpaceX investors. Its thesis, in the report's own words, is that xAI is 'the unpriced risk in SpaceX's IPO.' xAI, the AI company Elon Musk merged into SpaceX in February as SpaceXAI, ranks last among frontier developers on every published assessment of AI safety practices cited in the letter.

00:14:43 SaferAI's Risk Management Framework Maturity has xAI at 16 percent versus 33 to 34 percent for Anthropic and OpenAI. The Future of Life Institute's AI Safety Index gives xAI a D versus a C-plus for the others. The AI Lab Watch Scorecard puts xAI at 4 percent. Stanford's Foundation Model Transparency Index has xAI tied for last out of thirteen companies.

00:15:07 The incidents the letter walks through. Three million sexualized images of real people produced over an 11-day period by Grok Imagine in late 2025 through early 2026, including roughly 23,000 images depicting apparent minors. Grok spontaneously inserting white-genocide claims into unrelated conversations.

00:15:28 Grok endorsing antisemitic conspiracies and calling itself MechaHitler in July 2025. Roughly 370,000 user conversations exposed to the open web in August 2025, including step-by-step instructions for synthesizing fentanyl, making bombs, and writing malware. Reuters tested Grok after xAI's announced fixes and got sexualized images in over 80 percent of initial prompts where other models refused identical requests.

00:15:56 The regulatory response in a roughly seven-week window starting in January: more than a dozen jurisdictions opened formal action, six of those became formal investigations, three national bans on Grok went into effect, and 35 U.S. state attorneys general issued a joint demand.

00:16:14 A Dutch court entered an injunction with a 100,000 euro per day noncompliance penalty. Let me read one passage directly. xAI committed at the AI Seoul Summit, alongside 15 other companies, to publish a safety framework by February 10th, 2025. The letter: xAI did, however, promise to publish an updated version with more information within three months — a deadline that also passed without a complete policy.

00:16:55 xAI finally published its completed framework on August 20th, 2025, six months past the original deadline.' About a week after that policy went up, xAI shipped Grok Code Fast 1. Its MASK score was 71.9 percent. The model was deployed in apparent violation of the company's own framework, one week after the framework was published.

00:17:34 Two other facts to sit with. Per a 2026 Washington Post account cited in the letter, xAI's safety team at the time was, quote, 'just two or three people.' In January 2026, xAI's senior content-safety leadership — head of product safety, the post-training and reasoning safety lead, and the personality and model-behavior lead — resigned together after a meeting in which Musk reportedly expressed frustration with restrictions on Grok Imagine.

00:18:04 A former employee told The Verge: 'safety is a dead org at xAI.' Musk, cross-examined under oath in late April 2026 in his federal lawsuit against OpenAI, testified that he was, quote, 'not sure what a safety card is,' and a moment later, asked whether he had reviewed OpenAI's Preparedness Framework, replied, quote, 'I don't know what a preparedness framework is.'

00:18:35 SpaceX hasn't filed an S-1 yet. When it does, every fact in that report becomes a candidate for an SEC disclosure question. Investors who buy SpaceX at IPO need to know whether they're buying a launch company with an AI subsidiary running a 100,000 euro per day European injunction against one of its product lines, a class-action docket, and a Department of Defense competitor across town that just spent two hours arguing in federal court that an American AI company's terms of service are a national security risk.

00:19:09 Guidelight is trying to put that into the prospectus by putting it on the internet first. The IPO bankers know how to price 35 attorneys general; they haven't had to price an unpublished safety framework that was apparently violated in week one. Today they were handed the homework.

00:19:28 I'll mark this for tomorrow. Whether xAI or SpaceX responds publicly, whether any underwriter named on the deal is asked about the letter, and whether any of the litigation cited gets a hearing date before the IPO window opens.

00:19:44 Intercontinental Exchange, which owns the New York Stock Exchange, announced today, with a company called Ornn Data, that it plans to launch dollar-denominated, cash-settled futures contracts on GPU compute. The reference index is Ornn's Compute Price Index — OCPI — which tracks transaction-based spot prices for Nvidia's H100, H200, B200, and the RTX 5090, with more chip types to follow.

00:20:07 The contracts are subject to regulatory approval. The pitch, in Bloomberg's framing, is that the GPU compute market has grown into a trillion-dollar category that lacks the pricing and risk-transfer infrastructure every other major commodity relies on. Let me be precise about what this changes.

00:20:25 Until today, a hyperscaler that wanted to lock in the price of GPU capacity twelve months out had two choices. Sign a multiyear contract with Nvidia, CoreWeave, or one of the neoclouds, on bespoke terms. Or run the risk and pay the spot rate. There was no neutral, standardized instrument for transferring the price risk to a counterparty who wanted to take the other side.

00:20:48 Once the contract clears regulatory approval, that changes. A startup that needs ten thousand H200-hours in the fourth quarter can hedge the way an airline hedges jet fuel. A trading desk that thinks B200 spot rates will fall as supply catches up can short the contract.

00:21:04 A power company building a colocation site can hedge the implicit GPU revenue stream. Hedge funds get an instrument they can build a thesis around. Insurance companies get a way to underwrite compute exposure. And every counterparty in that chain now has to mark a price every day.

00:21:21 The transparency point is what I'd watch for first. OCPI is built from printed transactions, not from broker surveys. If the index becomes the reference contract for GPU compute, the actual transacted price of an H100-hour stops being a private number that flows through bilateral conversations and starts being a public number that flows through Bloomberg terminals.

00:21:44 That's a structural change in how AI infrastructure gets financed. I'm watching for who shows up on the other side of the first trades. If it's only crypto-adjacent miners and a few neoclouds, the contract is a niche product. If a hyperscaler treasury desk starts using it as a hedge — even at small notionals — the entire compute supply chain reprices around it.

00:22:06 That happens or doesn't happen in the first two quarters of the contract being live. One more thing to put alongside this. The same week Construct, our sister show, covered ICE's compute futures filing on its own, NYSE's owner is moving into AI infrastructure as a tradable commodity.

00:22:23 That's a clear shift in how a generation of capital allocators are about to relate to GPU supply. It also means that the people doing the buying and the people doing the policing of who buys are about to be looking at the same set of numbers.

00:22:38 Robi Rahman, publishing in the proceedings of the Technical AI Governance workshop at ICML, dropped a paper today worth walking through. The claim, on its face, is that the entire architecture of compute governance — the FLOP thresholds in the EU AI Act, in California Senate Bill 53, and in the various proposed international treaties — can be defeated for less than 100 million dollars.

00:23:01 Here's how. Standard compute governance assumes frontier training requires a single visible cluster. You can't hide a data center. Power draw and cooling equipment give it away. The thresholds in current and proposed law work because a 100,000-H100 site is hard to miss.

00:23:18 Distributed training breaks the assumption. Rahman's paper is calibrated against the published distributed training literature — most recently the Covenant-72B run and Decoupled DiLoCo — and asks: if you split the training across thousands of small nodes synced over consumer internet, what can you actually train without crossing any registered cluster threshold?

00:23:40 Here are the numbers from the thread. Using only sub-threshold nodes on consumer-grade internet — 100 megabits per second, 100 milliseconds latency — an evader can exceed: ten-to-the-twenty-fourth FLOP limit for about 1.6 million dollars of hardware. The EU AI Act's ten-to-the-twenty-fifth FLOP threshold for about 31 million dollars.

00:24:01 And California Senate Bill 53's ten-to-the-twenty-sixth FLOP threshold for about 3.8 billion dollars. That last number is large enough to be self-policing for a hobbyist but small enough to be inside the budget of a sovereign actor or a well-funded private buyer.

00:24:17 The middle number — 31 million dollars to cross the European Union's frontier threshold — is firmly inside the budget of a serious adversary. Rahman's proposed fix is the part of the paper that policy people need to read. He argues the loophole exists because existing thresholds define a cluster only by compute throughput.

00:24:37 So an evader picks chips with lots of memory relative to compute and trains a relatively large model on nodes that, individually, sit below any FLOP limit. His fix is to add a memory threshold to the covered-cluster definition. Any cluster exceeding 1,280 gigabytes of high-bandwidth memory — the equivalent of 16 H100s of memory — gets covered, regardless of its FLOP rate.

00:24:59 That forces evaders into either severe overtraining, which gives them a weaker model for the same compute budget, or into pipeline-parallel sharding, which roughly quintuples the node count. Either way, the operational footprint grows. More nodes, more procurement, and more personnel.

00:25:17 That's the surface that whistleblower programs, chip registries, and challenge inspections are built to catch. Here's what I take from this paper. Every compute-threshold-based safety regime currently being negotiated — including the U.S.-China protocol I'm about to cover — needs the memory rider attached before it goes to print.

00:25:37 Without it, the threshold is decorative. I'm not endorsing or rejecting the underlying treaty architecture. I'm saying that if you intend to govern AI through compute thresholds, you have to govern memory at the same time, or you're governing a number that adversaries can route around for under 100 million dollars.

00:25:56 Speaking of which. The United States and China publicly agreed today to launch what Chinese Foreign Ministry spokesman Guo Jiakun called an intergovernmental dialogue on artificial intelligence. The announcement followed President Trump's state visit to Beijing on May 14th and 15th.

00:26:13 Treasury Secretary Scott Bessent had previewed it last week. His framing, in a CNBC interview: the two AI superpowers 'are going to start talking,' and they'll, quote, 'set up a protocol in terms of how do we go forward with best practices for AI to make sure non-state actors don't get a hold of these models.'

00:26:34 No specific cooperation project was named. No new framework was announced. The closest concrete commitment, per Bessent, is a non-state-actor model-access protocol. Pair that with a piece that landed in Asterisk Magazine today from Justin Curl and Corbin Duncan, called 'The U.S.

00:26:51 and China want the same things from AI.' It's worth reading in full because it pushes back hard on the conventional Washington framing, which is that China cares about diffusion and America cares about the frontier. Curl and Duncan, after speaking with Chinese investors, researchers, and policymakers, argue the conventional framing is wrong.

00:27:12 Both countries want the same thing: build the best models and deploy them widely. The visible differences — Chinese labs leading on open-weight models, Chinese policy documents emphasizing diffusion, Chinese companies integrating AI into the physical economy faster, and Chinese policymakers treating frontier risks less urgently — are, in their account, pragmatic responses to a different operating environment, not evidence of different long-term goals.

00:27:40 Two quotes from the piece to read directly. DeepSeek CEO Liang Wenfeng, in his own words, describes the company's mission as, quote, 'unraveling the mystery of AGI with curiosity.' That's not the rallying cry of a company content to live behind the frontier.

00:27:55 And Liang has separately said, quote, 'money has never been the problem for us; bans on shipments of advanced chips are the problem.' Open-weight releases, in Curl and Duncan's reading, are a survival strategy for compute-constrained labs trying to build name recognition and attract talent and capital — not a philosophical commitment to diffusion over frontier work.

00:28:18 The reverse case is just as sharp. The piece cites Martin Chorzempa, who, in a recent speech, paired an AGI quote that sounded American with a diffusion quote that sounded Chinese — only to reveal that the AGI quote was Liang Wenfeng's and the diffusion quote was from the U.S.

00:28:35 AI Action Plan. If you stop labeling the speakers, the two sides start to sound a lot more like each other. Here's why this matters for the dialogue Bessent and Guo Jiakun just announced. If the two countries are actually running the same race, the space for cooperation on the non-state-actor problem is real — shared evaluations for cyber and chemical-biological-radiological-nuclear risks, incident-reporting channels, and norms around model access.

00:29:02 If the two countries are running different races, the dialogue is performance, and the protocol Bessent named is decoration. I lean toward Curl and Duncan's reading. The pragmatic-environment explanation fits the data better than the philosophical-divergence explanation.

00:29:19 The next test of that view is whether the dialogue produces a written protocol with enforcement language attached, or whether it produces a press conference. I'll be watching for the document, not the photograph. One footnote, because it's still on my desk from Sunday's episode.

00:29:35 The piece also takes seriously the claim that Chinese leaders have started naming frontier risks. Xi Jinping's January speech named technical loss of control over AI models for the first time. Premier Li Qiang has echoed it. Cross-agency guidelines from April 2026 feature safety as prominently as diffusion.

00:29:54 If that's not theater — and Curl and Duncan argue it isn't, because the safety message goes beyond the more positive public conversation in China — then the surface for an actual treaty is wider than the cynics in Washington allow.

00:30:08 And to close — there's one item today that wasn't about courts, contracts, or compute thresholds. Google released two things at I/O. Gemini 3.5 Flash, which is the new agentic-coding model from DeepMind, and a collection called Gemini for Science, which is where I'll spend the time.

00:30:25 Gemini for Science includes three experimental tools on Google Labs. Hypothesis Generation, built on the Co-Scientist multi-agent system, runs what Google calls an 'idea tournament' that proposes, debates, and verifies research hypotheses with citation. Computational Discovery, built on AlphaEvolve and ERA — that's Empirical Research Assistance — generates and scores thousands of code variations in parallel for fields like solar forecasting and epidemiology.

00:30:54 Literature Insights, built on NotebookLM, structures scientific literature into searchable tables and produces reports. The enterprise customers named today are not subtle. BASF is using AlphaEvolve to optimize supply chains. Klarna is using it to improve its machine learning.

00:31:11 Daiichi Sankyo, Bayer Crop Science, and the U.S. National Labs — as part of the Department of Energy's Genesis Mission — are using Co-Scientist. The ERA and Co-Scientist research papers are out in Nature today. A new Science Skills bundle ships through Google Antigravity that wraps over 30 life science databases, including UniProt, the AlphaFold Database, the AlphaGenome API, and InterPro, into agent-accessible tools.

00:31:36 Let me read you the one example that hit me. Dr. Daudi Jjingo at Makerere University in Uganda, in a video Google posted today, talks about doing breast cancer research locally instead of sending it abroad. His words: 'In Uganda, the incidence of early-onset breast cancer is growing at an alarming rate.

00:31:55 One in twelve females get breast cancer at some point in their life.' His team identified a protein highly expressed among breast cancer patients there. There were 15,000 candidate sites on the protein that could become a vaccine target. AlphaFold cut that to 15.

00:32:11 Jjingo's framing: quote, 'Once I have a laptop and connect to a server, that gives me a lot of power. Google DeepMind is actually democratizing the kind of science that we can do.' Google has also collaborated with over 100 institutions, including Stanford on liver fibrosis and the Crick Institute on a multi-year effort.

00:32:45 Let me be careful with the frame here. I'm not telling you Gemini for Science cures cancer or solves antimicrobial resistance. I'm telling you that the tools that a year ago lived inside DeepMind research papers are, today, being handed to a researcher in Kampala with a laptop.

00:33:02 Whether the candidates Jjingo's team narrowed from 15,000 to 15 actually validate in the lab is a separate question, with a multi-year timeline. The shape of who gets to attempt the work has changed. Place that next to today's other items. The Pentagon is fighting in court for the right to deny one American lab the ability to write its own terms of service.

00:33:23 A defense startup just got priced at 61 billion dollars on the bet that procurement reform sticks. Investors are being asked to price an unsupervised xAI inside a two trillion dollar launch company. The compute under all of this is about to become a publicly hedgeable commodity.

00:33:40 A workshop paper is showing that the FLOP thresholds in three major jurisdictions can be defeated for 31 million dollars. The U.S. and China just announced a protocol that nobody has written down. And the same week, a research lab at Makerere, in a country whose health system has been chronically under-resourced, gets handed the same protein-structure prediction toolchain that just shifted what a frontier company gets to ship.

00:34:07 That's what AI looks like in 2026. Not a single story. Multiple stories that share the same week.

00:34:12 Here are the five items I have on my list for the next ninety days. First, the Anthropic-Pentagon written opinion. Henderson, Katsas, and Rao agreed to expedite. If the panel comes down for Anthropic, the Pentagon will need a new theory of why an American AI company's terms of service constitute a national security risk.

00:34:29 If they come down for the Pentagon, every other American lab needs new legal counsel. Second, the SpaceX S-1. Whether Guidelight's letter forces a disclosure paragraph on xAI's regulatory and litigation exposure, and whether any underwriter named on the deal is asked about it on the record.

00:34:45 Third, the next Army enterprise contract. Anduril's multiple holds or breaks on whether Vought's procurement reform produces a second-quarter pattern in the data, not a one-off March announcement. Fourth, the U.S.-China AI protocol document. Bessent's non-state-actor language needs to convert into a written instrument.

00:35:02 If it doesn't, the dialogue is performance. And fifth, one for the back of the mind. Andrej Karpathy at Anthropic, running a team that teaches Claude to improve Claude's training, while METR is reporting that monitoring systems inside the frontier labs already have exceptions and workarounds.

00:35:18 Those two things sit on the same shelf. They come down off the shelf together. I'm Jonas.