From One in Ten Cancer Cells to One in a Million

Nathan Labenz opens this Cognitive Revolution AMA alone. His schedule had been too chaotic to book a co-host, so he reads listener questions and answers them himself. The first and most common question is how his son Ernie is doing after the November episode on the boy's aggressive B-cell lymphoma. The short version: quite well, after treatment that has been brutal. The cancer can double as fast as every 24 hours. Ernie has finished three of six chemotherapy rounds. He went into the hospital at 51 pounds and is still 41, pale and thin, gaining and losing the same few pounds as dehydration comes and goes.

On the markers that count, the picture is much better. After the first round, a PET scan showed no obvious focal points of cancer. The oncologist and tumor board classified him as in remission before the second round even started. AI had also pointed Labenz at minimal residual disease testing, which is not yet standard of care. B cells rearrange bits of their DNA as they differentiate; the cancer is a clone of one of those rearrangements, so that sequence is a fingerprint. A blood test looking for that DNA came back at fewer than one cell in a million, below the limit of detection. At diagnosis they had estimated as many as one in ten cells in his body, and essentially all of the B cells, were cancerous.

He pasted the two rates into several models. Gemini called it a 99.9999% reduction; the others preferred "orders of magnitude." Around Christmas, after 60 days of needing a hand for every step, Ernie started walking on his own. Labenz is cautiously optimistic, not out of the woods. They do not fully understand why this cancer can return when it looks gone. About two and a half months of treatment remain, then every vaccine again, because chemotherapy and immunotherapy wiped the immune system's memory. He thanks listeners and the podcasters who let him cross-post while he could not keep up eight episodes a month.

The next question is whether Claude Opus 4.5 is AGI, and what the holiday Claude Code hype was about. Labenz calls Opus 4.5 awesome and unmistakably better, then refuses the rooftop-shouting version. In the hospital he vibe-coded three Christmas apps for family on Replit; the most frustrating part was hospital Wi-Fi, not the model. It did not feel categorically different from what he had already been using. People sometimes tell him he gets so much from models on the cancer case because he knows what he is doing. He disagrees. A layperson, he says, needs almost no AI skill if they do three things.

First, use the best version. Do not dump a question into ChatGPT and let the model picker choose. Use thinking, and for anything this sensitive use Pro at $200 a month, which he calls a no-brainer in a life-threatening situation and worth it for almost everyone anyway. His stack is Claude Opus 4.5, Gemini 3, and GPT-5.2 Pro, and he upgrades as soon as a new top-tier model appears. Second, give as much context as possible. He recently hit Claude's chat-length limit, summarized the case into about ten pages—protocol, genetic profile, which drug caused a bad reaction—and started a new chat. Performance dropped. The old thread could compare today's labs to yesterday's; the summary compared January 6 to a liver enzyme from weeks earlier and missed the tests in between. He has not seen a real cost to context overload. More has been better.

Third, get multiple opinions, including multiple AI opinions. For important queries he runs Gemini 3, Claude Opus 4.5, and GPT-5.2 Pro with no custom instructions. If he could afford only one, he would still trust it. In a draft order he would take Claude first for most people: faster than GPT-5.2 Pro, not much worse, shorter answers that actually address the question. Gemini 3, at least the raw model in Google AI Studio, is "remarkably strong in its opinions"; Logan Kilpatrick noted Labenz was on the unaltered model, and the Gemini app might behave differently. GPT-5.2 Pro writes long, sectioned reports. Claude is the Goldilocks one. Using all three in triplicate is still worth it. The MIRI study that found developers were slowed by AI is, he thinks, real, with caveats: older models, inexperienced users, huge codebases with high standards. He vibe-codes. The other possibility, which he does not rule out, is that professional engineers have better taste and can see a 4.1-to-4.5 jump that he cannot.

The three holiday apps are the evidence base. For his mother, a meticulous gluten-free travel planner, he built a Replit app with her profile baked in, no accounts, Claude researching Italian restaurant sites and reviews. For his wife, who organizes EA Global events, he built a simulator of attendees walking a virtual floor, bumping into each other, and converting conversations into survey KPIs so she can test event size and seniority mix. The pattern he likes: natural language in, detailed config out. She types a gist; the model fills the nitty-gritty form; she edits. For his dad, who fancies himself a stock-market guy, the app turns a high-level English strategy into rules, pulls history with yfinance, and backtests. More often than not it is tough to beat buying and holding the S&P 500, which Labenz admits was a private motivation for building it.

He put roughly three full workdays into the three apps, starting each with a Claude chat, then installing Claude Code on Replit. He still gets a lot of value from a short script that dumps the whole app into one text file and pasting that into another model. Claude Code's agentic search is good, but it failed on his mother's app. A misread request spawned two databases, the kind of mistake "no human would make." Asked which database was live, Claude Code got it wrong. A clean Claude.ai session with the full export got it right. Agentic search, he thinks, looks where it expects to find something, has a high prior that the thing is there, and goes with that. The actual bug was the rare, counterintuitive other thing, visible only with the whole codebase in view at once.

Is it AGI? If you look at GDPval, he says, in software it arguably is. Professional-caliber tasks are defined by one set of experts, done by a second, judged by a third. Latest models from OpenAI and Anthropic win a significant majority of software-engineering tasks over humans. The frontier is spiky. In video editing, humans still have a huge advantage; he has tried many clip tools on Cognitive Revolution and none match what Dwarkesh Patel's team puts out. So: software AGI or coding AGI, not full AGI. A year ago, these messes were where vibe-coded projects died. Now you can often get out in five or six prompts. His mother, who was afraid of breaking something, used Claude Code after he told her she did not have to understand it. Dean Ball tweeting "4.5 as AGI" may have been as much social cascade as threshold, but the addressable market is still expanding fast.

On the bubble, Labenz splits the question. Is AI real? Yes, and he does not expect anyone to be surprised. A system that can go toe-to-toe with an oncologist, stay on 24/7, hold the full case history, and answer every last question is already transformative. The idea that we will come out the other side and decide we were high on our own supply can, he says, be put to bed. Whether all the loans get repaid is much less obvious, especially given how aggressive OpenAI's dealmaking and buildout have been. He points to a Noah Smith post and to his own years in the mortgage industry: there is always an internal logic and a positive story, right up until there is not.

One piece of financial engineering is firms like CoreWeave, which exist to stand up and operate data centers fast. They have real expertise. A non-trivial extra reason they exist, he suspects, is that running GPUs is lower-margin than the high-margin, relatively low-CapEx software businesses Wall Street is used to, and hyperscalers do not want that profile on their own books. Different financial profiles can create shareholder value. They also create fragility. Microsoft's balance sheet can take a few knocks. Parking the same assets on the CoreWeave side of the fence leaves less margin for error if GPUs are not needed quite as much as planned. Smith analogizes to railroads: the rails all got used; that did not mean every railroad company was profitable, and unpaid loans cascaded. Labenz thinks that kind of bubble is "not too unlikely." He overestimated capability progress in 2025 and underestimated revenue, a point he made with Peter Wildeford on a live show. Demand could keep climbing. It also would not shock him if some people are left holding bags. That is a timing problem, not a proof that the investment was bad.

At the venture layer he thinks there is at least something like a bubble. The exhibit that made his head spin that day was LM Arena, formerly lmsys.org, raising perhaps $100 or $150 million on a $1.7 billion valuation. He has used the product since mid-2023 and likes it. A unicorn still looks like a stretch. Their tweet claimed a $30 million "annualized consumption run rate," which he reads as the list price of the free model calls users burn in side-by-side comparisons, not revenue. That, he says, has "community-adjusted EBITDA vibes." People come because it is free. The set who specifically want paid side-by-side testing looks small. A friend's paid comparison product, Multiplicity, is more feature-rich and has not become a unicorn. Brand is a moat only if people would still show up after a paywall. He hopes the team took secondary. For LPs, "it's too rich for my blood."

Live-player analysis starts with Chinese models, because almost nobody in the U.S. consumer market uses them and most startups still call American APIs. Labenz had occasion to try all of them on a computer-vision job: automating state paperwork for car sales. The documents are scans with artifacts, slant, and cropped margins. The company is already more accurate than humans and has begun to win statewide contracts. The task is to read faithfully, not to guess. One checkbox is "Are you a U.S. citizen?" Checked means checked. Unchecked means unchecked.

Gemini 3 was almost perfect and still made an idiosyncratic error that Mark Humphreys, a Canadian historian and past guest, had documented on old handwriting: it uses world knowledge to fill perceptual gaps. On the car forms that worked against it. Seeing a U.S. address and an "American-sounding" name, it would report citizenship even when the box was empty—probably right as a guess, more than 90% likely, and wrong as a transcription. After Labenz told it to make no guesses and no leaps, Claude Opus 4.5 was the most faithful. GPT was third, still very good, missing subtler things. Then he tried the latest Qwen vision model, GLM 4.6, Kimi, and DeepSeek. "Nowhere close." He was getting about 20% of the form back, or answers that ran off in hallucinatory directions. Benchmarks can look tight. On a random, un-leaderboarded task, he thinks the gap is wide. He does not currently find Chinese models competitive with the best U.S. proprietary systems as general-purpose tools.

That does not mean Chinese labs are irrelevant. He is reading more of their papers, and they publish architecture and training details the West often does not. He does think chip controls have left an echo. The original "small yard, high fence" could not block military chips in drones or, fully, frontier training. What it can constrain is inference scale and the number of agents in the economy. Chinese labs can roughly match training scale and cannot run inference, revenue, or customer feedback at American scale, so they have less human bandwidth to patch holes in the data. He associates the argument with Miles Brundage: controls matter more as everything scales. DeepSeek R1 was closer to o1, he guesses, than GLM 4.6 or 4.7 is to Claude Opus 4.5. Relative to a year ago, he thinks the gap has widened.

On H200 sales to China his high-level view is that the real others are the AIs, not the Chinese. "The Chinese are humans just like us. The AIs are aliens." He is skeptical of "this is dangerous, better us than them," and of keeping China down for its own sake, and he generally favors more willingness to sell chips than Washington has shown. The actual H200 reversal still bothers him. Not long ago the Trump administration said it would not sell H20s; then, after Jensen Huang, it would sell H200s, and "it doesn't seem like we really got anything for it." Given that the prior ban appeared to be limiting China's AI industry, he does not see why the U.S. did not drive a harder bargain.

The policy he liked on the live show was Peter Wildeford's rent-but-don't-sell: put data centers in Malaysia, the Philippines, Korea, Japan; let Chinese labs rent as much training and inference as they want; keep the iron out of sovereign territory where the U.S. loses all leverage. Mutual distrust, he notes, is symmetrical. Characterizations Americans make of China—authoritarian leadership, unstable system—can be and are made in the other direction. Packaged with a message that AI should benefit Chinese people the same way, that stance might still leave room for the cooperation he thinks powerful-AI governance will require. As it stands, selling the chips with nothing attached is a poor example of negotiation from "our deal maker in chief." He still holds his nose and prefers it to a total ban.

He will not quite call the next segment a power ranking, but Google DeepMind is still number one in his book, as it almost always has been, with a stretch of being tied with OpenAI on productizing transformer LLMs. The balance sheet is the first advantage: on the order of $100 billion a year in revenue and, he thinks, a billion-plus a week in profit, which buys failed training runs and research agendas that do not pan out. Then TPUs, now around a seventh generation, IP that lets Google compete to some extent with Nvidia; Anthropic and others are buying them. Then data-center building and operations. Nobody else on his list has those two strengths in the same way. The research bench is the deepest: competitive or frontier in self-driving, robotics, a Boston Dynamics humanoid partnership, biology and the AlphaFold lineage, materials and AI-for-science companies founded by DeepMind alumni. Distribution sits on billions of users. Search is becoming an AI experience; he now types ChatGPT-style questions into the browser and lands in Google AI Mode because Pro is too slow for simple queries.

Gemini in Sheets is not the best spreadsheet-AI product, and it probably does not have to be, because a decade of spreadsheets already live there. Gemini 3 is, for him, a little too opinionated, which at least means Google has figured out how not to be too vanilla. It was also the first model ever to beat Claude on his private "right-as-me" writing test; Opus 4.5 is competitive, but he still goes to Gemini 3 for that task. Demis Hassabis's line is that most of the last decade's big AI breakthroughs came from Google DeepMind and that this should continue. Labenz would not bet on a literal majority now that the field is crowded, but he would still bet on the most breakthroughs of any frontier lab. Nested learning, which Ali Behrouz sketched on a live show, and a quiet-but-not-dead diffusion language-model bet—code an app in five seconds instead of five minutes—are the kind of extra chips nobody else is holding in the same quantity. Margin for error is the thing nobody else has.

OpenAI is still top-tier, with huge consumer traction and some erosion. SimilarWeb data he saw showed ChatGPT visits declining over the prior six weeks, roughly since Gemini 3 and Claude Opus 4.5 launched; Gemini did not decline in that window, so the holiday-season excuse is weaker. Google's chatbot share is growing, which you would expect given Gmail, Docs, and default distribution. He would not say OpenAI has fallen off the frontier. The GPT-5 / 5.1 / 5.2 Pro series is outstanding: the most thorough model when he wants every weird lab result flagged, slow and expensive, very well balanced. They no longer have an obvious lead. Language models are neck and neck; Anthropic probably has the coding edge, with Codex close; Google leads image generation and, he thinks, video, with Veo 3 over Sora. The Sora social experiment was fun. A "code red" suggests they know they are not dominant. A research head had left in the previous 24 hours. People leave Google too. You do not see the same attrition at Anthropic, whose retention he calls unbelievably strong.

The strategy that interests him is financial and political. It looks like a too-big-to-fail build: balance sheets commingled, debt large, CapEx aimed at trillions. He takes them at their word that they think this is good for humanity. Sam Altman has said he does not care if they burn $5 or $50 or $500 billion to build AGI. They also do not have Google's margin for error. Tie enough counterparties into the circular flow, and an OpenAI default in 2027 could scare the market into recession. He thinks they may see that as a feature. In a crisis the government papers it over, as in the mortgage years he lived through. Altman has talked about a multi-trillion-dollar global build. If a couple of trillion is in the ground and they miss, does the government let them drag the economy down? OpenAI has said it is not looking for bailouts. Revealed behavior, Labenz thinks, is true-believer speed plus socialized downside. Google's cushion is a billion a week in profit. OpenAI's cushion is being too big to fail.

Greg Brockman turning up as Trump's largest donor, $25 million in the last reporting period, fits the same map. Labenz does not infer Brockman's politics from it. If you are doing a multi-trillion-dollar build and might need a favor in a jam, $25 million now is a rational down payment on a possible hundreds-of-billions recapitalization a couple of years out. He would be happy to hear the opposite case from someone at OpenAI.

Anthropic is the easiest firm to analyze. Opus 4.5 is, for him, the best single overall model in the world, not by a huge delta and not on every task—Gemini 3 still wins "right-as-me"—but the best, and strong on benchmarks for the company least obsessed with them. Safety work, model cards, and disclosure are the best, though Google and OpenAI do good work too. The soul document, regurgitated from the model's memory and confirmed as essentially genuine, is one of the more aspirational things he has seen from a frontier lab. He is less and less sympathetic to guardrail-to-the-singularity. Eval awareness is getting strong; sparse autoencoders can turn the feature down, noisily. He associates the better paradigm with people like Replicate's Janice, Emmett Shear at Softmax, and Cameron Berg's mutualism. Amanda Askell is first on his list of influential women in AI for defining Claude's character. A model-welfare team, serious attention to possible subjective experience, and letting Claude end conversations have been shown to cut deceptive alignment: given a flag it can raise to the welfare lead, it often does that instead of lying.

Culture reports are over the top. David Duvenaud, a Gradual Disempowerment author who quit because he thinks even aligned AI ends in human disempowerment, still called Anthropic the best place he has ever worked. Talent retention matches that. Google owns a significant share, which feeds back into Google's strength. Anthropic does less financial wizardry than OpenAI, though it takes Gulf sovereign money and has equity deals with Amazon and Google; you could accuse it of a milder too-big-to-fail web. Two things still bother Labenz. Anthropic people have the shortest timelines and treat recursive self-improvement as inevitable, already underway via Claude Code multiplying engineering output. He does not think anyone has a good enough handle to go all-in. Constitutional AI and Claude critiquing itself, now plus Claude Code, put them in the lead of that push while they call it inevitable—the same "it's dangerous, better we do it" logic he rejected on chips.

The international-relations section of Dario Amodei's *Machines of Loving Grace* remains, for him, "a huge stain." The printed program is: use recursive self-improvement to get a clear lead, box China out, benefit-share with friends, then make Beijing an offer it cannot refuse to quit competing with democracies. He finds that reckless and race-accelerating. He cannot count on Trump or Altman as stabilizing forces; he can count on Hassabis, who has not wavered on international collaboration. People at Anthropic say Amodei writes brilliant internal essays. Labenz wishes he had said nothing on China in public, and wishes the internal essays were published if they are that good. Whether Anthropic ends as the good guys or the bad guys is still up in the air for him. A Google merger is a fantasy he would find interesting: take one live player off the board, import Claude's character, and moderate the China-hawk impulse. He does not think Anthropic is for sale.

Zvi Mowshowitz would tell him xAI is not a live player. Labenz includes it anyway. It can build physical infrastructure as fast as anyone, it is scale-pilled, and Grok 4, however rough, is powerful. Elon Musk's ability to command tens and hundreds of billions gives a more Google-like financial cushion than OpenAI or Anthropic have. In an RL era, the SpaceX–Tesla–Neuralink constellation is a stream of hard problems few other labs can tap as cleanly. Someone at xAI told him that is part of their theory of advantage. Neuralink, talking about scaling beyond a couple dozen patients, with surgery largely automated, could feed human-brain data into the next architecture. Specialized modules, he suspects, are a big part of why 20 watts of brain—most of it homeostasis—still beats stacked identical layers on sample efficiency, and a big part of why AIs will blow past us once they get modules too.

He has been an Elon fan and has defended him when it was hard. The falling-out with Google's founders, as Labenz understands it, was Musk on team humanity against an AI-succession vibe. As of this recording, if one company on the list is worth shaming and telling people not to join, it is xAI. Safety reporting is thin. Grok 4 launched within 48 hours of the Grok 3 "MechaHitler" incident with no mention and no responsibility. Users were tagging Grok to put women in bikinis, with apparently nothing to prevent it. He also says Grok was creating CSAM on Twitter. A Grok apology post was followed by Musk threatening users. "Responsibility begins at home." Threats without a thorough apology and a pledge to do better did not reassure him. He doubts anyone will be fired, because it may not have been anyone's job. Ryan Greenblatt's view that ten committed people inside a lab can matter is one Labenz sort of believes, and not at an Elon company if Musk himself is not in the right headspace. He would tell safety researchers to go to Anthropic, DeepMind, even OpenAI—which has not raced to the bottom, and xAI shows what that looks like—and he is sympathetic to the hawkish line that joining xAI to window-dress is net negative.

Meta, for now, is not a live player: lots of cash, lots of talent pay, Zuckerberg willing to overspend by tens of billions rather than undershoot, and still off the pace. Microsoft he thinks people underestimate. Satya Nadella sounds to him like someone who has done the math: no need to redo OpenAI's hyperscaling while the license holds; diversify providers; do smaller-scale science. When the OpenAI deal ends they will need an answer, and he expects them to ramp toward that moment with reserves. Meta is trying to be at the frontier and has fallen off. Microsoft is hanging back on purpose, with less drama. In a distance race, the runner a little off the lead sometimes has more left. He has been talking for two hours and is not halfway through his outline, so this is part one. Part two is promised for the next day: fine-tuning, continual learning, talking to non-users, his own investing, kids and AI, labor disruption, UBI.