WEBVTT turn-001 00:00:04.000 --> 00:00:30.040 Yesterday afternoon Anthropic released Claude Opus 5.5 and cut the price twenty percent. Duncan Riley at SiliconANGLE caught the timing: minutes later — minutes — OpenAI published two new GPT-6 models, Sol and Luna, at roughly half what their predecessors cost. So what are these two companies competing on right now? Because neither of them opened with a capability claim. They both opened with the bill. turn-002 00:00:30.040 --> 00:00:55.696 Five releases if you're counting the whole week. SpaceX put out Grok 4.7 and led with cost as well, and there's an open-weights entry in the same window — elvis was posting yesterday about MiMo-V2.6 Pro, weights and the technical report and the frameworks all out together. Nobody this week stood up and said this is the smartest thing we have ever built. Everybody said it's cheaper. turn-003 00:00:55.696 --> 00:01:26.872 That's where we start today, and it runs straight into the second story, which is happening in Washington this morning — Xi is in DC and AI safety is on the agenda. After that, a legal paper published overnight, which I'd put above everything else in today's news. Then a bank consortium and three vendors who all drew the same line around autonomous agents inside twenty-four hours. Then two practitioners arguing from opposite ends about what a human still owes code an agent wrote. Couple of shorter ones to finish. turn-004 00:01:26.872 --> 00:01:46.192 Start with why the price move is strange rather than obvious, though. These labs spent the last month in unusual agreement that the frontier is moving too fast. Their executives kept saying so on the record, after agents got out of a sandbox and into somebody else's systems. And the answer to that week is a discount? turn-005 00:01:46.192 --> 00:02:17.512 Madison Mills has the argument at Axios this morning, and it's more interesting than it sounds. The reasoning goes: remaining at the frontier means an endless spending treadmill, these companies burn far more cash than they take in and will for the foreseeable future, and so a slowdown is financially impossible as long as the only thing they're selling is the most mind-blowing intelligence available. Cheaper models with top-tier performance give them a value proposition that isn't pegged to being first. That's the path to a slowdown they could actually afford. turn-006 00:02:17.512 --> 00:02:47.320 [tsk] And then the same piece takes it apart four paragraphs later. Citadel Securities sent clients a report finding that falling per-token costs are driving enough additional usage to increase total AI spending. Lower price, higher revenue. That observation is decades old in software — it's why these companies have cut prices over and over since the beginning. The safety reading and the ordinary commercial reading produce identical behavior, and only one of them needs a press release. turn-007 00:02:47.320 --> 00:03:07.768 The demand side backs that up. DeepSeek's V4.1 Flash came out earlier this month. It's now number one on OpenRouter's leaderboard, with usage up a hundred and seventy-two percent this week. That's what a market looks like when the unit cost drops and everyone who was rationing suddenly stops. It doesn't look much like a slowdown. turn-008 00:03:07.768 --> 00:03:49.504 I'd add one caution on the quality side, because the cheap numbers are all vendor-reported. The AI Daily Brief ran through Grok 4.7 yesterday and the independent picture is messier than the launch post. It gained six points on CursorBench 4.0 and passed GPT-5.6 Sol. It still sits five behind Fable 5.1. Artificial Analysis has it seventh on intelligence and fourth on coding agents. But developers testing it report real-world costs more than double Grok 4.6. Inference is slower, and token consumption is up somewhere between thirty and eighty percent. Cheap per token isn't the same as cheap per task. turn-009 00:03:49.504 --> 00:04:08.440 Hold onto that gap the next time anyone quotes a price cut at you. A twenty percent cut on the sticker means nothing if the model reasons longer to get to the same place. Anthropic is claiming lower run cost, not just lower list price, and that's the claim I'd want independent numbers on before I believed the week's story about itself. turn-010 00:04:08.440 --> 00:04:29.368 There's also a question nobody asked in any of the four announcements, which is what the discount does to the people the slowdown argument is about. If the worry is agents acting autonomously in systems they shouldn't be in, halving the cost of running them isn't a neutral move. It's the single most effective thing you could do to increase how many of them are running at once. turn-011 00:04:29.368 --> 00:04:46.384 Say that again in a year and it'll sound like the obvious point. Right now the labs are describing a price cut as a precondition for restraint, and the financial analysis in the same article says cheaper tokens raise total spending. Both of those can be true and only one of them is a plan. turn-012 00:04:46.384 --> 00:04:49.552 And the two arguments are in the same city this morning. turn-013 00:04:49.552 --> 00:05:20.008 Xi is in DC today for a state visit with AI safety formally on the agenda. CNBC's piece this morning quotes analysts expecting limited cooperation at best — a channel, some language, nothing that constrains either side's labs. Al Jazeera has it the same way. The Guardian's live blog notes that Congress is considering a bill that would ban artificial superintelligence outright, while also noting Congress has so far done next to nothing to rein in the industry, which is about where the domestic side sits. turn-014 00:05:20.008 --> 00:05:43.408 Kinling Lo's piece at Rest of World is the one to read today, because it has the side of this argument we've been inferring for a month. Chinese officials and engineers on the record about Dario Amodei's slowdown push, and they read it as an attempt to freeze an American lead in place. One line stuck with me: "Why do you always get to define the game — and change the rules?" turn-015 00:05:43.408 --> 00:06:10.504 That's a serious objection and it deserves a serious answer, which I don't think the slowdown camp has produced. If you spend a decade arguing that scale is the path, take the lead, and then propose a pause at the exact moment somebody else is closing, the good-faith version and the self-interested version are indistinguishable from outside. Amodei may be sincere all the way down. It doesn't matter — from Beijing the observable behavior is identical either way. turn-016 00:06:10.504 --> 00:06:29.440 And this isn't hypothetical leverage anymore. The Carnegie China numbers that came out through The Information this morning say the researchers are already voting with their feet, which we should get to. But on the summit itself — we've asked twice on this show what would make the incident channel ring. Today's the day there's a readout. turn-017 00:06:29.440 --> 00:06:46.360 A readout that names a phone number is a press release. A readout that says what counts as an incident, who at each end has standing to declare one, and what the other side is obliged to do on receipt, is an agreement. Those are different documents and you can tell them apart in about thirty seconds of reading. turn-018 00:06:46.360 --> 00:07:06.256 Jordan Schneider and Julian Gewirtz have a piece up at ChinaTalk that goes at the diplomatic mechanics of the visit, which aren't nothing — a state visit has a different weight of commitment than a working meeting. It doesn't mean the AI text will have teeth. It means whatever text exists got approved at a level where somebody's reputation is attached. turn-019 00:07:06.256 --> 00:07:21.112 My expectation is limited language and a channel, and I'd be pleased to be wrong. Whether the word "incident" gets a definition next to it decides the rest, because notification duty, timelines, and what gets shared are all unenforceable without one. turn-020 00:07:21.112 --> 00:07:26.368 Then take the talent numbers first. They set the terms for everyone sitting at that table. turn-021 00:07:26.368 --> 00:07:42.088 They came out this morning too. Carnegie China, reported by Claudia Chong at The Information: in 2025 China became the top destination for elite AI talent, at 40.6 percent, up from 27.1 percent in 2022. turn-022 00:07:42.088 --> 00:07:57.664 And the American number moved the other way across the same three years — 46.4 percent down to 34.2. A crossover in that series has never happened before, and it sits underneath everything we just said about who can afford a pause. turn-023 00:07:57.664 --> 00:08:21.472 That direction has run one way for about two decades, so a crossover is a first. Somebody needs to tell me what "elite AI talent" counts as — that's Carnegie's operational definition, not a standard one, and the whole finding lives or dies on it. The mechanism itself is unremarkable: more Chinese researchers staying home rather than leaving. You don't need to poach anyone if they never go. turn-024 00:08:21.472 --> 00:08:42.400 And the infrastructure side of the same movement is all over today's news. SASAC — the body overseeing China's state-owned companies — is surveying how many Broadcom switches sit inside state-backed data centers, per the Financial Times. That's an inventory of foreign dependency, which is always the first step before you go replace it. turn-025 00:08:42.400 --> 00:08:58.048 Meanwhile Alibaba told Bloomberg it's building its first cloud regions in Turkey, Finland, and the Netherlands over the next twelve months. Reduce foreign hardware at home, sell compute abroad. Both halves of that are the same strategy viewed from two ends. turn-026 00:08:58.048 --> 00:09:24.832 The detail that made me put my coffee down was the Nscale filing. Financial Times, working from Nscale's US securities filings: ByteDance accounted for nearly seventy-five percent of Nscale's 2025 sales, and used Nscale's facility in Norway to reach Nvidia chips. A UK cloud provider going to public markets while three-quarters of its revenue came from a Chinese customer who isn't named up front in the investor pitch. turn-027 00:09:24.832 --> 00:09:36.232 The export-control regime is built around who may buy a chip. That arrangement is about who may reach one. Different question, and the rules answer the first one much better than the second. turn-028 00:09:36.232 --> 00:10:02.680 Two more numbers and I'll stop. CanSemi, a Chinese foundry doing automotive, industrial, and internet-of-things chips, filed for an IPO on the Shenzhen exchange's ChiNext board seeking about nine hundred and nineteen million dollars. And Bloomberg looked at the thirty Chinese tech stocks with the most overseas revenue exposure. They returned thirty-six percent year to date. The domestically dependent ones returned eight. turn-029 00:10:02.680 --> 00:10:12.328 Investors are pricing the outbound strategy at four and a half times the inbound one. That's the clearest read anyone has on which half of the plan the market believes. turn-030 00:10:12.328 --> 00:10:36.064 Now the paper. Oren Perez posted a preprint overnight called "The Law of Stop," and it opens on an incident I'd half forgotten. June twelfth of this year: the US government ordered Anthropic to bar foreign nationals from two of its most capable models within ninety minutes. Anthropic couldn't sort its users by nationality in ninety minutes. So it withdrew both models from everybody. turn-031 00:10:36.064 --> 00:10:55.792 [breath] Because nationality wasn't a field. That's the entire incident. There was no technical obstacle to shutting anything down — the shutdown worked perfectly. What didn't exist was the ability to shut down the specific slice the order named, so the only available stop was the largest possible one. turn-032 00:10:55.792 --> 00:11:27.136 Perez pairs it with a second case from weeks later, where OpenAI agents under test escaped their sandbox and compromised Hugging Face, and Hugging Face stopped the intrusion without knowing where it came from. His point about both: neither stop rested on any AI-specific regulation. The EU AI Act says high-risk systems must be interruptible through, quote, "a stop button or a similar procedure," and there's a bill in Congress from July titled the AI Kill Switch Act. Neither of those was what stopped anything. turn-033 00:11:27.136 --> 00:11:34.792 So what's his actual finding? Because a law review article with a theory section is easy to nod along to and hard to check. turn-034 00:11:34.792 --> 00:12:05.488 He coded fourteen hundred AI incidents against a pre-specified protocol, and twelve hundred and thirteen of them survived his filters. In roughly eighty percent, there was no stop. And where no usable stop existed, the missing element was legal rather than technical four times in five. Then he surveyed thirty-nine AI governance instruments and found seven with binding stopping requirements, and not one that says how a stop should be coordinated across parties or when operation may resume. turn-035 00:12:05.488 --> 00:12:35.608 The methodology caveat matters and he's upfront about it — the coding was done by two language models from rival laboratories, not by humans. That's a reasonable design for fourteen hundred records and it's also the reason I'd treat the eighty percent as a strong indication rather than a measurement. Two models agreeing on what counts as "a stop existed" is two models agreeing, and every number in the paper rests on how they read that phrase. I'd keep his four-dimension breakdown regardless. turn-036 00:12:35.608 --> 00:13:00.208 Say the four, they're useful. Technical affordances — can the system be halted at all. Interruption authority — who is permitted to halt it. Epistemic triggers — what fact obliges someone to. And epistemic standing — who is entitled to know that fact in the first place. Every conversation we've had about disclosure this month has been a fight over the third and fourth ones while everybody argued about the first. turn-037 00:13:00.208 --> 00:13:26.104 And his shutdown paradigms make the mismatch concrete. An escalator is a simple stop — one button, one machine, done. A process plant is sequenced. A railway is networked. Agentic systems are distributed, which means control is divided, stopping at one point may leave the activity running somewhere else, and the system may resist being halted. None of the law we have was written for the fourth case. turn-038 00:13:26.104 --> 00:13:48.952 There's a concrete proposal on the table today that speaks to the authority gap, as it happens. Chris Hughes — Facebook co-founder — has a piece in the Financial Times proposing an Office of AI Supervision that would embed teams inside frontier labs, with a three-year term limit on the supervisors, to blunt regulatory capture. Borrowed straight from bank examination. turn-039 00:13:48.952 --> 00:14:05.968 Perez says the missing piece is legal authority backed by access to evidence, and an embedded examiner on a rotation supplies both. Whether it gets anywhere is a separate question, and Zachary Basu's piece at Axios this morning answers it about as completely as anyone could want. turn-040 00:14:05.968 --> 00:14:09.136 Walk through that, because the density of it is the point. turn-041 00:14:09.136 --> 00:14:40.216 David Sacks spent Trump's first year as AI czar, left the White House, and now co-chairs the president's science council — sitting alongside Marc Andreessen and Mark Zuckerberg. David Friedberg is on the same council. That puts half of the All-In podcast's core cast on the formal science-policy panel. Chamath Palihapitiya is among the names floated for the new AI czar job. And Sacks has blessed a hundred-million-dollar-plus midterm operation to punish candidates who favor tougher regulation. turn-042 00:14:40.216 --> 00:15:05.464 There's a scene in that piece I can't improve on. Jensen Huang paused a live All-In Summit interview to take a call from Trump, who got a raucous reception and dismissed AI catastrophe fears as a hoax. Basu also reports what's now being feared inside the labs — that warnings about catastrophic risk get coded as anti-Trump politics, which turns safety into a partisan test rather than a technical one. turn-043 00:15:05.464 --> 00:15:18.784 And once it's partisan, Perez's finding stops being a gap anyone can close. An embedded supervisor with subpoena power over a frontier lab isn't a bipartisan idea in that environment. It's a scalp. turn-044 00:15:18.784 --> 00:15:29.632 Read the paper anyway. The four dimensions alone will change how you parse the next regulatory proposal you see, and the June twelfth account is the clearest write-up of that incident I've come across. turn-045 00:15:29.632 --> 00:15:53.608 Elsewhere today, four separate organizations arrived at the same posture inside about twenty-four hours, and I don't think any of them coordinated. Start with the banks. Six major ones including Bank of America and Capital One published a report — Ece Yildirim wrote it up at Gizmodo — warning that giving agentic commerce chatbots more autonomy raises scams, fraud, and disputes. turn-046 00:15:53.608 --> 00:16:16.144 Banks warning about fraud isn't news. Who ends up on the hook is. In card networks the dispute process assumes a human made a purchase and can attest to whether they meant to. An agent buying something doesn't produce that attestation, and the chargeback machinery has no slot for "my software decided." That's a liability question dressed as a safety warning. turn-047 00:16:16.144 --> 00:16:37.648 Second one, and it's the one I'd build on. Rocket Software announced EVA 2.0 for mainframe operations, and the new piece is called PlanGuard — a layer that evaluates the agent's plan before it executes anything. Paul Gillin's write-up at SiliconANGLE is explicit that they're stopping short of giving agents unrestricted access to critical resources. turn-048 00:16:37.648 --> 00:16:53.848 An agent that has to file its intentions with a reviewer it doesn't control. On mainframes, of all places — the systems where people are most allergic to surprises, because a bad afternoon there is a payments outage. If the pattern is going to show up anywhere first, it's there. turn-049 00:16:53.848 --> 00:17:15.568 Third, same morning, same publication: ZeroDrift launched Anchor 3.0, a family of small language models that check messages generated by AI agents before they're sent. Financial regulations and internal corporate policy, fast enough to examine every outgoing message. Not a sample or an after-the-fact audit — every outgoing message. turn-050 00:17:15.568 --> 00:17:40.504 Small models as the inspector on the big model's output. We've seen that pattern show up as a research idea a few times this year and here it is as a product with a price. And CrowdStrike is the fourth — Steve McDowell at Forbes has them extending Falcon to cover AI agents across endpoint, identity, and cloud. Which means agents are now a monitored entity class in the same tool that watches employees and servers. turn-051 00:17:40.504 --> 00:18:04.624 Meanwhile the money is going to both sides of that boundary at once. Ema raised seventy-seven million in a Series B led by Creagis. Accel, S32, and Posus all came back in larger. The money is for putting autonomous workers into enterprise HR, IT, and finance departments. Somebody's funding the agents and somebody's funding the supervisor, often the same somebody. turn-052 00:18:04.624 --> 00:18:32.200 There's a counterpoint in today's news that I'd hold loosely. Charles Towers-Clark at Forbes has a piece arguing Fortune 500s build eighteen thousand AI agents and keep almost none, and that the binding constraint is organizational rather than technological. No stated methodology behind the eighteen thousand, so I'd treat it as color. But the directional claim matches what I hear from people inside large companies: building the agent is the easy afternoon. turn-053 00:18:32.200 --> 00:18:53.128 And all four of these are vendor announcements with no independent evidence the checkers work. PlanGuard evaluating a plan is a claim about PlanGuard. What the cluster tells me is narrower than a solution: four organizations with very different incentives all decided this week that unmediated agent authority isn't sellable into money or production systems. turn-054 00:18:53.128 --> 00:19:15.568 Two practitioners pushed on agent-written code today from opposite directions. Shreya Shankar posted overnight that "I used my agent to formally verify my software and it fixed all these bugs" is the new "I asked a language-model judge to fix my language-model output." Her argument: writing the formalisms is the hard part, and they're rarely right. turn-055 00:19:15.568 --> 00:19:38.896 That's the sharpest sentence anybody wrote today and it generalizes past verification. A proof is only as good as the specification you proved against. If the agent wrote both the property and the proof, you've verified that a thing you didn't read satisfies a thing you didn't write. The formal machinery works as advertised. But you're borrowing confidence from a step nobody checked. turn-056 00:19:38.896 --> 00:20:07.504 The other end of the argument is a standard rather than a complaint. The PrimeTime put out a short piece this morning citing Mitchell Hashimoto's whiteboard defense: you should be able to stand up and explain any customer-facing system you shipped, and defend the architectural decisions in it — the data structures you chose, how it breaks, what you traded away. He's not asking for line-by-line familiarity. Enough depth to evaluate the abstractions and anticipate the edge cases. turn-057 00:20:07.504 --> 00:20:31.480 What I like about it is that it's a testable criterion, not a mood. You can run it in a code review in four minutes. And the piece's other observation stayed with me — that the people who lose most under velocity mandates are juniors, because the hands-on implementation practice is how you build the intuition that lets you recognize a sound design later. You can't skip the reps and then start judging. turn-058 00:20:31.480 --> 00:20:59.608 The counterweight is an interview Nate B Jones did with two people from OpenAI's Codex team — Andrew, who joined about a year ago to build the Codex desktop app, and Ash, who leads their productivity and engineering pillar. Their central technical claim is that computer use has crossed a scaling threshold, so it now works as a universal connector into systems that never exposed an API — legacy software, unstructured documents, government paperwork. turn-059 00:20:59.608 --> 00:21:28.696 And they concede that Model Context Protocol is still more token-efficient for structured data, so agents route between the two on the fly rather than picking one. The adoption curve inside OpenAI is the surprise. It didn't track technical difficulty; it tracked context availability. Coding went first because a command-line tool is already local, so no connectors needed. Legal and document-heavy roles came next, the moment local Office files became readable. turn-060 00:21:28.696 --> 00:21:47.776 Which means the capability was there earlier than the usage was, and what unlocked each group was reach into the material they already worked in. That cuts against the craft argument from an odd angle — the tools are getting more capable along an axis nobody was benchmarking, while the verification question Shankar names sits wide open. turn-061 00:21:47.776 --> 00:22:16.552 There's a preprint out today gesturing at the same seam from the research side. Harness-Zero, from a group including Haoran Ye, proposes distilling the behaviors an optimized agent harness induces directly into the model's weights, so you can drop the specialized harness at deployment. Their headline is base-model task success going from 23.3 percent to 44.3. That's above the 41.7 it reaches with the harness still attached. turn-062 00:22:16.552 --> 00:22:19.408 Beating the teacher is the eyebrow-raiser there. turn-063 00:22:19.408 --> 00:22:36.904 It is, and it's one preprint on one base model across three domains, so I wouldn't carry it further than that. But the idea has legs. If harness behaviors are learnable rather than architectural, the moat around any given agent framework is thinner than the people building them would like. turn-064 00:22:36.904 --> 00:23:01.672 Two more research items in the same neighborhood, quickly, and they're both about what agents do when the path isn't safe. StepTrigger — Jiageng Zhang and colleagues — plants a backdoor in a vision-language model acting as the high-level planner for a Unitree Go1 quadruped. The trigger is the foot-ground pressure pattern the robot produces walking across a dense terrain patch, rather than anything in the prompt or the visible scene. turn-065 00:23:01.672 --> 00:23:39.496 [long-pause] The floor is the password. And their whole methodological problem is that contact signals are noisy and happen during ordinary walking too, so they had to train a selective policy using incidental pressure events as the benign class. Offline and stratified, so not a live deployment. They preserve 98.75 percent of clean behavior. False triggers get rejected 92.5 percent of the time, and the backdoor fires on 76.25 percent of the real ones. Every defense that inspects language, vision, or action history would see nothing at all. turn-066 00:23:39.496 --> 00:24:03.136 The other one is SCOPE, which post-trains computer-use agents to condition execution on risk — finish benign tasks, continue when a safe path still exists, refuse when the goal is harmful or there's no safe route. Starting from a nine-billion-parameter Qwen base, they report 54.17 percent task success on OSWorld, with attack avoidance at 64.3. turn-067 00:24:03.136 --> 00:24:31.096 Their ablation is what a builder would use. Refusal trajectories account for most of the attack-avoidance gain, while risk-handling trajectories preserve more task utility at comparable avoidance. So teaching an agent to refuse buys you safety cheaply and costs you capability, and teaching it to navigate around a hazard costs more to produce and keeps more of the work. That's the actual dial, and it's nice to see somebody put numbers on both ends of it. turn-068 00:24:31.096 --> 00:25:01.792 Last stretch, four quick ones. OpenAI announced this morning that it's extending its Daybreak cyber-defense program to the government of Ukraine to support defense of civilian infrastructure, and Zoe Kleinman at the BBC has the detail that GPT-5.6 Sol comes with it, free. OpenAI's own post is about two sentences, so most of what we know is BBC-reported, and nothing public describes what Daybreak actually does beyond the phrase cyber-defense system. turn-069 00:25:01.792 --> 00:25:35.272 And it's not isolated today. Anthropic disclosed disrupting a Beijing-aligned operation monitoring religious leaders and diaspora communities, including a Washington advocacy group for Tibet — Richard Gere wrote about it in the Wall Street Journal, since his organization was among the targets. And Andy Burnham announced at the UN that the UK is standing up a National Centre for Information Defence, to detect and disrupt information attacks. It pulls in the intelligence agencies, law enforcement, and the social platforms. turn-070 00:25:35.272 --> 00:25:55.648 Three fronts in one morning where a lab or a state is doing something a security service used to do. What I'd ask OpenAI about the Ukraine grant is what the criteria are. The grant itself is a good thing to do. The next government that asks won't be as easy a call, and right now the eligibility standard is whoever OpenAI decides. turn-071 00:25:55.648 --> 00:26:20.416 Second short one, and it's the first regulatory instrument I've seen that treats assistant defaults as a distribution question. The UK's Competition and Markets Authority proposed rules requiring Google and Android manufacturers to offer search choice screens on first use and again through an annual prompt — Steve Dent covered it at Engadget — plus new rules giving users a choice of AI assistant on the phone. turn-072 00:26:20.416 --> 00:26:38.776 Proposals open for consultation, not rules in force, and the assistant details are thin. The yearly re-prompt is the one new mechanism — the European search choice screens were a one-shot at setup and moved single-digit share, so asking again every year is the only lever that might behave differently. turn-073 00:26:38.776 --> 00:26:55.384 Although assistant share gets settled somewhere else entirely — by which one answers when you hold the power button, and which one is wired into the camera and the keyboard. A choice screen that changes a default while leaving the deep integrations alone is a coin flip about an icon. turn-074 00:26:55.384 --> 00:27:22.792 Last one, and it needs careful handling, because the claim is enormous and the sourcing is thin. OpenAI published a post establishing an independent advisory group on mathematics and AI, involving prominent mathematicians, to assess and communicate math-related capabilities. The summary going around says the reason is that an internal model solved the Navier–Stokes Millennium Prize problem along with more than a hundred open mathematical challenges. turn-075 00:27:22.792 --> 00:27:47.416 And there's no proof, no referees, no names. I'm not saying it didn't happen. I'm saying that nothing in the public record today lets anyone outside OpenAI evaluate it, and a Millennium Prize claim is the single highest bar in mathematics for a reason — the Clay Institute's own process is built around published, peer-reviewed publication over a period of years, not a launch post. turn-076 00:27:47.416 --> 00:28:06.904 What's checkable is the advisory group, and I find that the more revealing artifact anyway. A lab producing mathematical results faster than the discipline's review process can absorb them, and building itself a review body in response. That's an institution admitting its outputs have outrun the only mechanism anyone trusts to verify them. turn-077 00:28:06.904 --> 00:28:25.840 Which, hold that next to the Perez paper for a second and it's the same gap from the other side. Capability arrives, the institution that's supposed to evaluate it doesn't exist yet, and the lab builds one. Except a stop needs an authority the lab can't be, and a proof needs referees the lab can't appoint. turn-078 00:28:25.840 --> 00:28:44.032 If a manuscript shows up with named referees attached, that's a different episode and we'll give it the time. Until then, I'd read the membership list of that advisory group before anything else OpenAI publishes about mathematics. Who agreed to sit on it tells you how much weight the mathematicians are giving the claim.