Latest / Super Intelligence News and Report / Day 14
Transcript
- Super Intelligence News and Report — Day 14 Welcome to Super Intelligence News and Report. Today is Tuesday, November 3rd, here are the headlines. Theo: Today's throughline is the frontier business hitting physical and financial limits at the same moment it claims to be near self-acceleration, with record compute bills, emergency power deals and lawsuits over the data underneath. Maya: Start with chips. Marvell is buying Lumentum in an all-stock deal announced November 2nd. It is expected to close in the first half of twenty twenty-eight, subject to approvals. The implied value is about a hundred and fifteen billion dollars, among the largest semiconductor deals ever. Marvell makes custom networking chips, Lumentum makes the lasers that link AI processors, so this is really about moving data fast enough inside giant data centers that are still getting bigger. Theo: And the market clearly wants to see how Marvell digests a deal that size. Maya: Same buildout, different funding problem. Oracle said November 2nd it will sell common stock and senior notes to raise money for AI data centers, with the offerings expected to close November 4th. The total is twenty billion dollars. That matters because Oracle is not a startup hunting for its first check. It is a large cloud vendor, and it still has to go to public markets to keep paying for capacity at the pace customers are demanding. Maya: Salesforce is buying its way into voice. It announced November 2nd a definitive agreement for Cresta, whose voice agents handle calls and coach human agents in real time. Cresta will fold into Service Cloud, with closing expected in fiscal first quarter twenty twenty-eight. Salesforce did not disclose a price. About two point seven billion dollars, reported, not confirmed. The bet is plain: the phone call is still where a lot of expensive, messy service work happens. Theo: Which is exactly why automating it is valuable, and why mistakes there are so visible to customers. Maya: CrowdStrike is buying agent security before agents get looser in the enterprise. At Fal.Con Europe on November 3rd, it announced Dreadnode for automated AI security testing, and Fiddler AI for agent observability, both to land in Falcon in fiscal fourth quarter twenty twenty-seven. CrowdStrike disclosed neither price. The pair at about four hundred ninety-five million dollars, reported, not confirmed. Think pre-deployment testing, plus a flight recorder and a kill switch for an agent that starts doing the wrong thing. Maya: Now the data fight widens. The New York Times sued Brave Software November 2nd in Manhattan federal court, and the Daily News, Chicago Tribune and Orlando Sentinel filed a parallel suit the same day. They allege Brave copied journalism into its search index, resold access through an API, reproduced paywalled text in Ask Brave answers, and blocked paywalls while substituting its own ads. Brave had not responded by the morning of November 3rd. The target here is the middleman, not just a model lab. Theo: That changes the liability map for everyone buying search data, not only everyone training models. Theo: People on X are talking about OpenAI's LifeVal, posted November 2nd. It is a benchmark for everyday errands, built from American time-use data and a large public poll, covering childcare, money, shopping and government paperwork. GPT-6 Astra scored just over eighty percent on average reward, only slightly ahead of the next model, while much older models lagged far behind. Why it matters is the shift in what gets measured: not another exam, but the chores people actually open an assistant to finish. Theo: Contrast that with GDPx from SemiAnalysis, also released November 2nd. It uses expert tasks across economically chosen domains, including chip design, cybersecurity, biology, finance, legal and software engineering, with long time limits and strict grading. Claude Opus five point five led at twelve percent. That is about one in eight expert tasks completed properly. The top system also wins in some hard technical areas, but the headline is the ceiling. Everyday tasks may be saturating, expert agency is not close. Maya: One in eight is a useful reality check before anyone talks about hands-off research teams. Maya: Power is the next bottleneck. OpenAI announced its Grid Protectors Program November 2nd, committing future U.S. compute campuses it directly develops to cut draw when grids are strained. The pilot is Project Camellia in Effingham County, Georgia, under a long Georgia Power deal, with a large curtailable share built in. The relief on offer is a thousand megawatts in an emergency, described as roughly the air conditioning for seven hundred thousand homes. Not a daily saving, a promise that could decide whether a campus gets built at all. Theo: On X, OpenAI also posted November 2nd that its next internal security sprint will include Cathedral Systems researchers using OpenAI models alongside internal teams to challenge defenses. At the same time, its bug bounty now pays up to thirty thousand dollars for critical bugs in ChatGPT, Codex and the API Platform, with exceptional findings paid more. That is more than four times the prior critical midpoint. As AI gets better at finding flaws, the money for proving them has to move just as fast. Theo: Anthropic has a quieter research drop. It released FlashProve November 2nd, an Apache two point zero prototype that uses zero-knowledge proofs to show a response came from committed model weights, without revealing the weights themselves. It includes a prover, verifiers and a visualizer, and runs at a hundred and ten tokens per second on one H100 for Llama four Maverick in a test setup. Still heavy for production. But it points to a hardware-free way to audit that the model you paid for actually did the work. Theo: Meituan is now selling agent models, too. It made LongCat two point five generally available November 2nd, five weeks after preview. It is a natively multimodal mixture-of-experts model meant for terminal, browser, graphical interface, spreadsheet and design tool work, and it works with Claude Code, Codex and Hermes. The context window is a million tokens, aimed at long coding jobs that need to hold a whole project in view. A delivery giant offering frontier-length context at budget pricing will pressure everyone else. Maya: Especially teams already counting every long run against margin. Theo: Two agent infrastructure items to close. DeepSeek's Harness team has open-sourced a super-agent prototype under MIT, with the repo up October 31st and an essay on super-agent design. It routes approved GitHub and Feishu requests to isolated parallel workers, one repo per deployment, with no new model attached. Separately, Trajectory announced November 2nd it raised forty-five million dollars to build a post-training platform that learns from live app feedback for legal and medical specialists. One parallelizes coding tickets, the other sells continual retraining without an in-house research team. Maya: Next, we are going deep on Anthropic telling the SEC it may be nearing its own line for AI-accelerated AI development, and what that claim looks like against the compute bill behind it. Now for the Main Episode Maya: So here is the question we are left with after that tease. Is Anthropic genuinely approaching the point where AI sharply speeds up AI research itself, or is an IPO filing using safety language as both a warning and a sales pitch? Theo: And the reason that question is hard is that both readings fit the same paragraph. Maya: Exactly. So let me walk through what Anthropic actually put in front of the SEC, and then we can pull it apart. Theo: Please do, because the headlines made it sound like they said Claude is about to build the next Claude on its own. Maya: They did not say that. Not about any current model. Maya: On November 2nd, Anthropic filed an amended S-1. This is the version that adds third quarter results and new language in the Risk Factors section. Theo: And the timing matters here. They first filed publicly on October 26th, and that version only covered the first half of the year. Maya: Right. This update is the first one that puts full quarterly financials next to this particular warning about accelerated progress. Maya: The company is preparing to list Class A shares on Nasdaq under the ticker ANT. The expected listing is before Thanksgiving, on November 26th, reported, not confirmed. Theo: That date is doing a lot of work for the marketing calendar, presumably. Maya: We will get to that. First, the exact claim, because the wording is narrower than the reaction to it. Maya: Anthropic said its increasing use of AI in its own research and development may cause capabilities to advance faster than safety work can keep up. Theo: That part alone is not new. Every frontier lab says some version of that. Maya: The next part is newer. They said they have in the past failed to recognize a significant acceleration until after it happened. Theo: So they are telling investors, we might not see it coming in real time, and we have missed it before. Maya: And then they go one step further. They said internal estimates and expectations for future models, not current models, indicate they may imminently meet their own definition of dramatically accelerated progress attributable to AI. Theo: That definition lives in their Responsible Scaling Policy, which is their published system for monitoring and mitigating risk as capabilities rise. Maya: Right, and they were careful to add the limit right after. They have not designated any current model as over that line, and they continue to assess the data. Theo: That sentence is the whole ballgame for me. Future models, internal estimates, may imminently meet. That is three layers of not yet. Maya: It is also signed language in an SEC filing, which is why people reacted the way they did. Maya: TechCrunch reported the language on November 2nd, framing it as the first time a frontier lab put a recursive self-improvement threshold warning in an SEC filing signed under penalty of perjury. Theo: And conversation on X amplified it the same day. One post drew about a hundred and twenty one thousand views, which tells you how fast the sales pitch reading and the warning reading spread together. Maya: Anthropic added one more commitment in that section. They said they do not expect to be able to safely use Claude to fully autonomously design successive frontier generations without meaningful human supervision, if that becomes technically possible. Theo: And they said they are considering updates to keep humans in control. Maya: Considering. No date given, no draft shared, no metric published. Theo: So what do we actually not know here? Maya: Almost everything that would let you test the claim. We do not know whether any model is actually near the threshold. Maya: We do not know what metric they use internally to judge dramatically accelerated progress. We do not know what data would trigger a formal designation. Theo: We do not know whether auditors or the board would agree with that internal estimate. And we do not know what the human in control update will actually say. Maya: Plainly, that is what the brief says is still unknown, and it matters because without a metric, imminently is a judgment call. Theo: Okay, so why put a judgment call like that in Risk Factors right before an IPO? That is your beat. Maya: Because Risk Factors is both places at once. It is where you disclose what could hurt the business, and it is also what every investor reads to understand why the business might be extraordinary. Theo: So saying we might be close to AI accelerating AI research is a risk if safety cannot keep up, and it is also the most bullish sentence you could write about your research productivity. Maya: That is exactly the tension TechCrunch flagged. The report noted skeptics point out an S-1 functions as both disclosure and marketing, so claiming closeness to self-building AI can warn and advertise at once. Theo: Let us weigh the serious reading first, because I do not want to dismiss it just because it helps the pitch. Maya: The serious case, as TechCrunch reporter Julie Bort put it after reviewing the filing, is that the language deserves attention precisely because it sits in Risk Factors, where false statements carry liability. Theo: And on X, one user made the same point in shorter form. The SEC signature matters. Maya: Bort also pointed to the capability backdrop. Claude gains in coding, math and cybersecurity this year make faster self-assisted research plausible, not just theoretical. Theo: That backdrop is real from the build side. Labs have increasingly used coding agents in their own research, which is part of why recursive self-improvement became a popular term earlier this year. Theo: If your researchers are spending their days directing agents that write, test and debug research code, then a better model does not just help customers. It speeds up the next training run, the next evaluation, the next experiment. Maya: And that is where the money story and the capability story become the same story. But hold that thought, because the counterweight is in the same news cycle. Theo: GDPx. Maya: GDPx. Theo: We mentioned it in headlines, but it belongs inside this story because it tests the thing Anthropic is warning about, just from the outside. Theo: SemiAnalysis released it on November 2nd. Two hundred and six expert tasks across eleven economically chosen domains, built with fourteen data partners, with long time limits and strict grading. Maya: And Claude Opus five point five led the field. Theo: At twelve percent. About one in eight expert tasks completed properly. Maya: Say that again slowly, because it sits awkwardly next to imminently. Theo: One in eight. The best agent available, on tasks in chip design, cybersecurity, biology, finance, legal and software engineering, finishes about one in eight to the standard. Maya: And in finance and legal, every model scored zero on that strict grading. Theo: Zero. Not low. Zero. Maya: The brief also notes alignment tasks broke behavior rules in most runs. That is not a footnote if your worry is handing research over to the system. Theo: It is not. And the cost tells the same story about how far this is from effortless autonomy. Theo: Opus five point five averaged about fifty one dollars per GDPx task, with well over half a million output tokens. The runner up on cost was around sixteen dollars per task. Maya: So the leader costs about three times as much per attempt as the cheaper competitor, for a twelve percent success rate. Theo: Which is useful progress, and also expensive trial and error. That does not look like a system ready to design the next frontier generation without meaningful human supervision. Maya: Anthropic would agree with that sentence, to be fair. They said they do not expect to be able to do that safely. Theo: They would. My push is different. If expert agency is still at one in eight, what would dramatically accelerated progress even look like internally that we cannot see externally? Maya: That is the sharp listener question. It could be narrow. Coding help inside a lab is a much tighter loop than open ended legal or finance work. Theo: That is the concession I would make. Internal research acceleration does not require solving every expert domain. It requires getting much faster at the specific experiments a lab runs every day. Maya: And GDPx cannot see that internal loop. It can only put a ceiling on the broad claim. Theo: Right. GDPx says autonomous frontier development is not close. It does not say internal productivity is not compounding. Maya: So both things can be true. The external ceiling is low, and the internal speedup could still be real and still be hard to measure from outside. Maya: Now bring in the money, because Anthropic is asking public investors to fund the loop that would produce that speedup. Theo: And the numbers are genuinely large, even by frontier lab standards. Maya: Third quarter revenue was about seventeen billion dollars. A year earlier it was roughly one point four billion, so about twelve times higher. Theo: That is real enterprise demand. You do not get to seventeen billion in a quarter on hype alone. Maya: No. For the first nine months, revenue was about thirty three billion dollars. So the scale is not in dispute. Maya: What investors are pricing is the change in slope. Sequential growth slowed to forty seven percent in the third quarter, from a hundred and forty four percent in the second quarter. Theo: Forty seven percent quarter over quarter is still extraordinary for a company this size. It is just less extraordinary than what came right before it. Maya: Exactly. And the operating line flipped with it. Anthropic returned to an operating loss of about five hundred and sixty seven million dollars in the third quarter, after an operating profit in the second quarter of just under a billion dollars. Theo: That swing matters more to me than the headline net loss, because operating is closer to whether the core business covers its costs in the quarter. Maya: The headline net loss is much bigger, and it needs unpacking because it is misleading on its face. Maya: The third quarter net loss was about twenty three billion dollars, and about a hundred and four billion for the nine months. The filing attributes most of that to non-cash charges on convertible notes and SAFEs. Theo: So accounting revaluation as the expected valuation rises, not cash walking out the door that quarter. Maya: Right. That distinction was part of the argument online. One user countered that accounting adjustments turning losses into claimed profits would not pass a basic exam, which tells you how contested the presentation already is. Theo: And on the other side of that argument, you have investors looking past the accounting to the growth. Maya: Investor Gavin Baker of Atreides Management made the bullish business version. He said a hundred billion dollars in annualized revenue by year end could be possible, implying the growth justifies the spend. Theo: Annualized from a seventeen billion quarter gets you part of the way there, but you would need another strong step up in the fourth quarter to hit that pace. Maya: Which is why the fourth quarter operating result is one of the concrete things to watch. A renewed profit or a second straight operating loss will say more than another adjusted headline. Maya: Now the spending promise that ties this back to your world. Purchase commitments and leases in the filing total about eight hundred and ninety five billion dollars. Theo: Say that in human terms. Maya: It is the contracted computing bill stretching years ahead. With more estimated payments, the total could pass a trillion dollars. Theo: Compared with what they had disclosed just a week earlier, that is a huge step up. Maya: In the first filing, covering only the first half, commitments were in the five hundred and forty to five hundred and sixty seven billion dollar range. Now they are near nine hundred billion. Theo: So in one update, the future compute promise grows by roughly the size of a very large company, and investors are supposed to read that alongside the claim that AI may soon accelerate AI research. Maya: That is not a coincidence. That is the same bet stated two ways. Buy enough compute for years, run better models on it, use those models to build the next ones faster. Maya: It also explains why the other headlines today are not separate stories. Oracle raising twenty billion, Marvell paying for optics, OpenAI promising to cut power draw in an emergency. Theo: Because someone has to supply the chips, the networking, the cloud capacity and the electricity for commitments like Anthropic's, and pay for them before the revenue from the next generation arrives. Maya: And public markets are being asked to fund Anthropic's share of that directly now. Theo: What are investors actually debating on price? Maya: Investors are reported to be pushing for a valuation range of one point five to one point six trillion dollars, down from two trillion discussed earlier. Reported, not confirmed. Theo: Down, but still pricing in a lot of that acceleration actually happening. Maya: Very much. Business professor Daniel McCarthy is reported as praising other parts of the S-1, while D.A. Davidson research head Gil Luria called it a very speculative investment. Maya: New Constructs called the offering thesis overblown. That spread, praise for disclosure quality next to speculative and overblown on value, is the debate in miniature. Theo: And retail is part of that debate too. Maya: Underwriters including Goldman Sachs, Morgan Stanley and JPMorgan are reported to lead, with JPMorgan leading a United States retail slice of fifteen to twenty percent of shares. Reported, not confirmed. Theo: That matters for who bears the risk if imminently turns out to mean years. It would not just be institutions who can model a trillion dollars in commitments. Maya: It would be individual buyers reading the same Risk Factors paragraph as a promise. Theo: Let me press the positioning case harder, because I think you are being generous. Maya: Go ahead. Theo: If I am preparing to list before Thanksgiving, and my sequential growth just slowed from a hundred and forty four percent to forty seven percent, and I just flipped back to an operating loss, what sentence would I most want in the conversation? Maya: That you might be close to the thing that makes growth compound again on its own. Theo: Exactly. And I would want it in the one section no one can accuse me of putting in a marketing deck. Maya: I concede the timing helps the pitch. I do not concede that the timing proves the claim is invented. Companies disclose real risks at IPO precisely because they have to price them. Theo: But they choose how imminent to call it. They chose that word. Maya: They chose may imminently meet, based on internal estimates for future models. That is hedged in every direction except the headline it creates. Theo: And they added that they have missed an acceleration before until after it happened. That could be candor. It could also be a way to say, trust our internal signal even though we cannot show you the metric. Maya: Both readings survive, which is why I keep coming back to what would settle it. A published metric, or a Responsible Scaling Policy update with an actual trigger. Theo: Or a designated model. The day they designate a current model as over their own line, the debate changes completely. Maya: Until then, for the people building with these tools, what does this change in practice? Theo: Less than the headline suggests, and more than nothing. If internal acceleration is real, model capability and coding assistance could improve faster than planning cycles assume. Roadmaps built on annual step changes get stale. Theo: But GDPx says do not plan on handing expert work over. One in eight, zero in finance and legal, behavior rules broken in most alignment runs. That is a human supervised tool, not a replacement research team. Maya: And for investors, the stakes are concentrated rather than broad. This is really about Anthropic and its rivals, and whether the compute bill compounds into an advantage or into a fixed cost no one can escape. Theo: Because if everyone commits to roughly a trillion dollars in future compute, the advantage is not having compute. It is getting more research per unit of compute than the lab next door. Maya: Which is another way of stating the opening question. Recursive acceleration, if it arrives, is the thing that makes nine hundred billion in commitments look visionary instead of reckless. Theo: And if it does not arrive, those commitments are just a very large bill against growth that already slowed once. Maya: That is why I would not tour every affected group here. Developers will feel it through capability and price, but the real bet is Anthropic convincing public investors that its research loop justifies buying the machines for that loop years in advance. Theo: I land slightly more skeptical than you on the claim itself. Maya: Tell me where. Theo: I believe Claude is speeding up Anthropic's coding and experimentation. I do not yet believe we have evidence for dramatically accelerated progress attributable to AI, in their own sense of that phrase. Theo: Internal estimates for future models is not evidence I can weigh. Especially with no metric, no trigger, and an external benchmark at twelve percent. Maya: I am more willing to take the Risk Factors placement seriously, because that section has consequences, and because a company does not lightly tell the SEC it has failed to notice acceleration before. Maya: Where I join your skepticism is on valuation. Serious as a risk disclosure does not mean worth one point five trillion today, especially with growth decelerating and a second operating loss possible. Theo: So we agree it can be a real warning and still be doing marketing work. Maya: We do. And we disagree on how much weight to give the internal signal we cannot see. Theo: Fair. So what would change my mind is concrete. Publish the measure for dramatically accelerated progress, or designate a model against it. Maya: For me, the next filing does the first piece of work. The next amended S-1 should set a price range and a listing date, with marketing expected the week of November 9th and pricing before November 26th. Reported, not confirmed. Theo: And the Responsible Scaling Policy update they said they are considering. If humans in control means something specific, that document has to say what it means. Maya: And then the fourth quarter numbers, whenever they are disclosed next. Renewed operating profit would support Baker's argument that growth can carry the spend. Theo: A second straight operating loss, against that compute bill, would support Luria and New Constructs more than any argument about wording. Maya: So back to where we started. Is this genuine proximity to AI accelerating AI research, or safety language doubling as a sales pitch? Theo: My read is that the filing proves Anthropic wants investors to price the possibility now, before the evidence that would prove it exists in public. Maya: Mine is that the liability and the history they admitted make it more than a pitch, but not yet more than a possibility either. Theo: Which leaves us watching the same three things. A metric, a policy with teeth, and a quarter that shows the business can fund the loop without asking the story to do all the work. Maya: And until one of those lands, imminently is the most important undefined word in the filing. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.