Super Intelligence News and Report — Super Intelligence News and Report: Day 27 Lead, GPT-6 Astra's Physics Claim Under Scrutiny Welcome to Super Intelligence News and Report. Today is Monday, November 16th, here are the headlines. Maya: Nvidia's crisis has left the boardroom and landed in Washington, right before Tuesday's earnings. So let's start there. On Sunday, November 15th, Treasury Secretary Scott Bessent said he wants answers from Nvidia's board and its outside counsel within ten days. The question is whether Nvidia pursued what he called seemingly illicit chip sales to China to boost revenue, with chips allegedly routed through Megaspeed, a Singapore company. Axios reports he asked whether Nvidia goosed its earnings. He also urged the board to hire an outside CEO, and he took another swipe at Commerce Secretary Howard Lutnick. Nvidia didn't comment. Theo: And that ten-day clock runs right past Tuesday's report. Maya: It does. The bigger point is that a sitting cabinet secretary is now publicly steering a company's governance. That's not normal. Theo: Over to the model side, where the story is also partly about checking claims. People on X are picking apart Alibaba's Qwen4-Max, released Monday, November 16th. It's the flagship of the Qwen4 series: a huge mixture-of-experts model with native vision and a million-token context. It's closed, with no open weights, though Alibaba says smaller open versions are coming. The coding headline is that it edges out Claude Opus 5.5, but that rests on one of Alibaba's own benchmarks, DeepSWE. On Terminal-Bench, Opus is clearly ahead. Maya: So where does it land on independent rankings? Theo: Artificial Analysis puts it ninth, well behind Opus. But the average cost per task is roughly a third of what Opus costs. That's the real story. China's top closed model is competing on price, not on the frontier. Maya: Back to Washington, and this one is reported, not confirmed. The Wall Street Journal, single-sourced, says the administration will name more than two dozen companies as AI national champions. Nvidia, Amazon, Microsoft, OpenAI and Anthropic are on the list. The idea is to bundle American chips, models and cloud into packages that foreign governments and businesses can buy, as a counter to Huawei. Reported perks include faster export licenses, government support overseas and federal financing. Lutnick is expected to unveil it, possibly within a week, and people on X have plenty to say about it. Theo: So Nvidia gets investigated by Treasury and crowned by Commerce in the same weekend. Maya: Pretty much. And Anthropic makes the list despite the February order telling federal agencies to stop using its products. The two halves of the government aren't exactly in sync. Theo: Which brings us to the Jensen comeback talk. Maya: Right. CNBC reported Monday that calls from the tech world for Jensen Huang's return are growing. Reports say he's weighing a comeback, and that allies are setting up a lobbying war room. Reported, not confirmed, as are the petition numbers. The Stand with Jensen petition is the thing everyone on X is sharing. Signers include Elon Musk, Amjad Masad and Tim Sweeney, though its authenticity has been questioned. About seventy-five hundred signers say they're current Nvidia employees, which nobody has verified. Meanwhile the FT says the board approached Arm CEO Rene Haas. Nvidia says it hasn't formally engaged any candidates and is still picking a search firm. Theo: And bringing Huang back would be a direct no to Bessent. Maya: Exactly. An outsider is what Treasury asked for, and Huang is about as inside as it gets. Next, Megaspeed itself. On November 15th, the company said Alice Huang left effective August 30th, 2025, and has no current role. That's about fourteen months before this blew up. It's also hired outside counsel to review its export-control compliance. A May short-seller report from Culper Research, though, quoted an employee calling her still our boss. Theo: Convenient timing on that exit date, if it holds up. Maya: Separately, and this part is reported, not confirmed, Semafor says Lutnick will fly with Trump to China this week. Bessent has been pressing him on why Megaspeed isn't on the Entity List. So whether Megaspeed gets blacklisted is now tied up with the Trump–Xi meeting. Theo: And the markets noticed. Maya: Hard. On Monday, November 16th, Korea's Kospi fell nearly four percent. SK Hynix and Samsung both took big hits. In Japan, Advantest, Lasertec and Tokyo Electron all slid. Nvidia reports fiscal third quarter on Tuesday, November 17th, after the close, with interim CEO Colette Kress on the call. Consensus revenue sits just above Nvidia's own guidance of about a hundred and eight billion, and that guidance assumes zero data-center revenue from China. Theo: First earnings call without Huang. That's a strange one to listen to. Maya: It's the biggest test of the AI trade this week, and now it comes with a leadership question attached. And right behind it is Anthropic. Quick update: pricing is still set for Wednesday, November 18th, with trading on Nasdaq under ANT starting Thursday. The FT reports bankers are weighing a higher range on strong demand. That's reported, not confirmed, and there's no amended filing yet. Morningstar puts fair value at ninety-two dollars a share, about a quarter above the midpoint of the filed range. Its argument is that the range reflects AI sentiment more than fundamentals. Theo: And they call that undervalued, even with the range possibly moving up? Maya: That's their call, though their bear-to-bull spread is enormous. Either way, the biggest AI listing yet prices one day after Nvidia reports. No pressure. Theo: Last one, and it's a quieter gem. On November 15th, Apple launched the Apple Health AI Usage Study app with OpenEvidence. It's open to US iPhone users twenty-one and up, and it runs for more than a year. Participants grade AI health answers that link to peer-reviewed research, and they fill in monthly surveys. It's aiming for up to two hundred thousand people. That's a big deal for OpenEvidence, whose own app is otherwise limited to clinicians. Apple says the data is anonymized and won't be used to train models. Reports differ on which iPhones qualify. Maya: So Apple is doing its homework before shipping anything? Theo: That's how I read it. They're collecting evidence on consumer medical AI at scale before it shows up in Health or Siri. Very Apple. And speaking of evidence, up next we're going deep on OpenAI's claim that GPT-6 Astra cracked a decades-old physics problem, and who can actually check it. Now for the Main Episode Theo: So here's the question. OpenAI and a respected physicist say a frontier model did the hard proof work on a problem that's been stuck for about half a century. Is that a real scientific advance, and how would anyone in the field actually know? Maya: And it comes the same weekend as Qwen4-Max, where a lab's own benchmark did a lot of the talking. So I'm coming in a little suspicious. Who's making the claim? Theo: Nathaniel Craig. He's a particle theorist at UC Santa Barbara, also affiliated with Kavli, and he's a visiting scientist at OpenAI. Keep that last part in mind. The model is GPT-6 Astra, OpenAI's frontier model. Maya: So the human author has a formal link to the company announcing the result. Theo: He's a visiting scientist there, yes. It's disclosed, and it doesn't disqualify him. But it's context. Now the problem itself, because it's a good one. Physicists test the Standard Model by putting it on a lattice, basically a grid of points in spacetime, so a computer can work with it. That works fine for the strong force and for electromagnetism. Maya: And then the weak force spoils it. Theo: Exactly. The weak force treats left-handed and right-handed particles differently. Physicists call that chiral. Put a left-handed particle on a grid using the standard rules, and an unwanted right-handed twin shows up. That's fermion doubling. Because of it, nobody has had a rigorous definition of the full Standard Model on a lattice that doesn't lean on approximations. OpenAI and Craig say it's been open about fifty years. Some coverage says twenty-five-plus. Either way, it's old. Maya: Has anyone made real progress on it? Theo: In 1998, Martin Lüscher solved a simpler version, the abelian case. Craig's work extends that approach. Peter Woit, the Columbia mathematical physicist who writes the Not Even Wrong blog, says this line of research was built up in the eighties and nineties and then mostly dropped as unpromising. Maya: So an abandoned method, picked up again by someone new to it. What did Craig actually do? Theo: There are two moments, and the order matters. On October 6th he posted a preprint on arXiv called "A gauge-invariant measure for lattice chiral fermions." Per Woit, it runs a hundred and twenty-three pages, with sections written by Astra mixed in with sections Craig wrote. Craig wrote that the model was essential to the result. Maya: A hundred and twenty-three pages. How many people can check that properly? Theo: Not many, and not fast. It's a niche inside a niche. Woit also points out that Craig hadn't published on lattice gauge theory before. Astra played the mathematician on the team, producing rigorous arguments, and some of those can be formalized in Lean, the proof-checking language. Maya: Interesting split. The physicist has the idea, the model does the rigor. Theo: That's the pitch. Then the second moment. On November 15th, OpenAI launched a science blog, where researchers will explain their findings and methods and give practical guidance. Craig's write-up is the first guest post. He says Astra took his refinement strategy and turned it into a concrete construction with a detailed proof. That's a strategy he couldn't get working in every case he needed. Maya: What did the model actually contribute, mechanically? Theo: According to a DEV Community summary of the post, Astra suggested mapping field configurations onto finer and finer lattices, worked through phase corrections, and built out the detailed proof. Craig also thanks three lattice theorists for their input: Maarten Golterman, Andreas Kronfeld and Yigal Shamir. Maya: And that's input, not peer review. Theo: Right. Being thanked in a paper is not signing off on it. What I respect, though, is how Craig frames it. He calls it a candidate construction, and he says it's more an answer in principle than an answer in practice. Maya: That's a lot more careful than "AI cracks fifty-year-old physics problem." Theo: Yeah. There's a real gap between what the scientist says and what the headlines say. Maya: From the business side, why is this the very first post on a brand-new science blog? That's a choice. Theo: It's a statement. Research collaboration is the flagship use they want to be known for. And it follows August 1st, when OpenAI published "Ten advances in mathematics and theoretical computer science," crediting Astra, unreleased at the time, with ten results that came with machine-checkable proofs. Independent review was still pending then. Maya: So a summer of math claims, and now physics. Theo: And physics is harder to verify. With the math results, a checker either accepts the proof or it doesn't. Some of Craig's arguments can go into Lean too. But the questions still open here aren't the kind a proof checker can answer. They're about physics. Maya: Go through those, because that's where "in principle" versus "in practice" really shows. Theo: There are three, and Craig and the coverage list all of them as open. First, reflection positivity at finite lattice spacing. Roughly, does the construction behave like a physically sensible quantum theory. Second, does it connect to a continuum interacting theory, what you get as the grid spacing shrinks to nothing. Third, does it give any practical route to running simulations. Maya: So you could have a proof that's airtight in Lean and still describe something that isn't physically meaningful. Theo: That's the risk. A correct theorem about the wrong object. Maya: Which brings in Woit. He's the main skeptic? Theo: He's the only named outside physicist we have on the record. In his October 8th post, he called Craig's preprint a modest advance on Lüscher's 1998 work. Far from a full definition of the continuum theory, and it doesn't enable practical electroweak calculations. He also argued the program was abandoned for good reasons, and that today's models are far from the deeper insight you'd need on non-perturbative chiral gauge theories. Maya: That last one is a claim about the models, not just this paper. Theo: It is. And one more thing. One summary of his post says Woit noted Craig hadn't tried to follow the most complicated, Astra-written parts in detail. Maya: Wait. So the human author hasn't fully checked the hardest sections? Theo: That's how that summary describes Woit's point. It's secondhand, so hold it a little loosely. But if it's accurate, that's the core of the verification problem. The hardest part is the part a person has checked least. Maya: For me, that's the whole story. Not whether the model is smart, but who carries responsibility for the proof. In a normal paper, the author vouches for every line. If he openly relied on the model for the parts he couldn't do, accountability moves somewhere new. Theo: I'd push back a little. Big collaborations already work like that. Nobody on a large paper checks every section. You trust your co-authors' parts. Maya: But a co-author can defend the proof at a conference. They have a reputation on the line. The model doesn't. Theo: Fair. You can ask it to redo a step, but it isn't accountable the way the field means. I'll give you that. Maya: So what's the strongest version of the other side? Theo: An established particle theorist saying, under his own name, that the model did the detailed construction and proof he couldn't finish, on a problem that's beaten people for decades. That's a working scientist putting his credibility on it. And he's not overselling. He's the one who said "in principle." Maya: Being modest isn't the same as being right. Theo: No. But it changes the argument. Woit is disputing whether this is a big deal. He's not really saying the math is wrong. Maya: Good distinction. "Modest advance" is still an advance. Theo: And then there's the twist that changes how the whole thing reads. Edward Witten, probably the best-known living theoretical physicist, has posted a sole-author note on the same problem. It's called "A Note On Lattice Definition of Gauge Theories With Chiral Fermions," thirty-six pages. This part is reported, not confirmed. Secondary write-ups, FourWeekMBA among them, say it credits Anthropic's Claude Opus 5.5 in three places and thanks T. Hartman for pointing Witten to Craig's paper. Maya: So a rival lab's model turns up on the same frontier problem. In a Witten paper. Theo: If the write-ups are right. One says Witten's method can in principle be computed, but isn't practical yet with standard Monte Carlo methods. And to be clear about what we don't know: the arXiv page itself wasn't opened in our research, and it's unclear whether Witten is building on Craig's result or taking a different route. Maya: That matters. If he's extending Craig, that's a big endorsement. If he's going around him, that's a quieter vote of no confidence. Theo: Or just a different idea. Theorists do that. Either way, Witten writing on it at all tells you the problem is live again. Maya: And from the money side, look at the backdrop. OpenAI opens a science blog to plant a flag. There's a reported Witten note with an Anthropic model in the acknowledgments. And Anthropic prices its IPO this week. I'm not saying any of it was timed. But both labs are selling investors and enterprise buyers the same story: our model does real research. Right now, only one of them is the one announcing it. Theo: That's fair. And for people actually building with these tools, I think that's the real shift. Not "AI solved physics." It's that top theorists are using frontier models as working collaborators, across labs. Maya: Is anyone outside physics weighing in? Theo: Lots of attention, not much expertise. OpenAI's post on X was at about three hundred ninety thousand views by Monday. And an informal poll on the daily.dev aggregator came out about two-thirds skeptical. Maya: Which tells us what? Theo: Mood. Developers are tired of breakthrough headlines. It says nothing about whether the construction is right. Almost nobody voting has read a hundred and twenty-three pages of lattice gauge theory. Maya: Including us. Theo: Including us. Which is kind of the point. Most people can't check this, so they fall back on trust, and trust in lab announcements is pretty low. Maya: So who can check it? Theo: A small group. Golterman, Kronfeld and Shamir, the three Craig thanked, are exactly who I'd want to hear from. They know Lüscher's approach well. Responses on arXiv or their blogs, or formal follow-ups, would be the real review. And so far, beyond Woit, we found no named outside physicist reacting to the November 15th post. Maya: Here's my read, and it may differ from yours. The scientific claim is honest and small. The marketing claim is big. Craig says "candidate" and "in principle," and then it gets packaged as the flagship launch of a new blog. The danger isn't that the math is wrong. It's that the framing gets ahead of the review. Theo: I mostly agree, but I'd weigh one thing more. If the model really did the part the human couldn't, even a modest advance is a milestone in how research gets done. Modest advances on fifty-year-old problems are how fields move. Maya: Even if the human didn't follow the hardest parts? Theo: That bothers me too. But I see it as a norm we have to work out, not a reason to toss the result. Put what you can into Lean, and let the specialists argue about whether it's the right object. That's a healthy split. Maya: If it happens. Right now we have one skeptic, three acknowledgments, and a reported Witten note nobody on our end has opened. Theo: Fair. And the bigger lesson is that verification has changed. The checker only gets you partway. Reflection positivity, the continuum limit, whether it's ever simulable, those take people. Maya: And the labs are now fighting over scientific credibility, on blogs. I'll be watching whether OpenAI's next science posts hold up, and whether Anthropic answers with write-ups of its own, given the reported Witten paper. Theo: Plus whether Witten's note builds on Craig or goes around him. That one detail would tell you how the top of the field reads this. Maya: So, back to the question. Is it a real advance? Theo: Probably real, and probably modest. Which is what the physicist himself said. Maya: And how would the field know? Slowly. A few lattice specialists, a proof checker for the parts it can handle, and time. Not a view count. Theo: Which means the first real signal is whether Golterman, Kronfeld or Shamir say anything on the record. If one of them posts a response, that's the story I want to bring you next, more than anything OpenAI puts on its blog. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.