Gavin Baker: SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?

Episode 274 of All-In opens with one regular host absent and Gavin Baker of Atreides Management sitting in. The hosts list the week's tech stories: the SpaceX and OpenAI IPOs, Andrej Karpathy joining Anthropic, and Nvidia's results. They start with Karpathy, who is described as only 39 and already a legend in the industry. There is some side talk about the Liquidity event, where Karpathy is expected to give a keynote on day two, Gavin Baker will be there too, and Baker is told he will help judge the best-ideas section.

The discussion goes through Karpathy's record. He was a founding member of OpenAI, led the self-driving team, and coined the term "vibe coding." He recently built "auto research," an open-source training tool that helps AI models improve themselves through five-minute experiments. The tool passed 82,000 GitHub stars and is described as a weekend experiment that led many non-specialists to build their own recursive LLMs. Someone has also released a tool called "Andre Karpathy skills," built on his principles for Claude Code. At Anthropic he will reportedly lead a new pre-training team focused on recursive self-improvement, meaning Claude would improve itself.

One host questions how much this matters in 2026. The field is in "a different inning" than it was ten years ago, when Karpathy was at Tesla, or five years ago, when he co-founded OpenAI. The reply compares him to Google's "Google Fellows," a small group of technical stars such as Amit Singhal, Shrida Ramaswamy and Jeff Dean who were at the front of wave after wave of Google's work. Karpathy is said to have been at the front of successive AI waves in the same way. The speaker calls him probably the first person to commercialize Richard Sutton's "bitter lesson" essay, about brute-force computation, while leading FSD at Tesla. The speaker adds a story Karpathy told privately, which may not be public: around 2016 or 2017 he spent perhaps a quarter of his time hand-labeling Tesla video data.

On this view the hire matters because Karpathy is a curious inventor and recursive self-learning puts models on "a combination of overdrive and autopilot." Combined, these could produce an order-of-magnitude improvement every year, described as a new form of Moore's law in which model quality goes up parabolically. The speaker sums up the key point as putting a lot of compute at the problem, after which the models learn very quickly.

Asked about Anthropic's success and hiring spree, one panelist calls the success "extraordinary" and "undeniable." They cite a Wall Street Journal report that Anthropic was EBIT positive in its most recent quarter, and say this matters for the wider debate about circular funding and return on investment. Their estimate is that OpenAI and Anthropic together are at "call it, a $100,000,000,000 of ARR" with gross margins on inference of about 80%, and growing fast. In their view that means the returns are there.

The same speaker adds Gemini, Cursor, xAI and open source and projects a much larger high-margin annual recurring revenue figure for language models by the end of the year. The figure is garbled in the transcript ("$203,104 $100,000,000,000"), so its exact value cannot be recovered. When asked, they confirm the estimate includes Google. It deliberately leaves out what they call some of the most economically important and profitable uses of GPUs: better recommender systems, ad targeting and ad measurement at Facebook, Google and Amazon. Even on the narrow definition of LLM tokens, they expect a strong ROI this year.

The speaker then names two possible "final frontiers" for AI. The first is recursive self-improvement, where during training the model itself, or another model, has input into its own training. The second is continual learning, called "the holy grail," where a model learns from experience the way humans do. That has not been achieved yet. If recursive self-improvement works, the annual 10x figure, which this speaker attributes to Chamath, "might seem conservative." Together the two capabilities "might pull the future forward in a very real way." The forecast is conditional throughout.

The panel's estimates of Anthropic's lead are loose. It is said to be ahead of everyone else by three or six months, probably six to twelve months ahead of open source, and maybe three, six or nine months ahead of its peers.

The host poses a question: at what point will AI contribute more than humans to improving a language model, crossing what they call "the recursive valley"? The host suggests this may already have happened at Anthropic, though humans still orchestrate the process.

The answer gives no date. The speaker says they don't know when a full model will be fed into a context window to train a new model. They suggest many other architectural paths exist. One is to build many smaller models that work together as a network, which could cost less energy or money per token than a single large model. The speaker notes they have made this point three or four times before and argues there is a lot of room in redesigning models and how they cooperate. In their view, one minor breakthrough could halve the cost per token. As support they point to early papers, including one from MIT that they shared a few weeks earlier, which suggest "there's a lot of room to run here."

The host agrees that small and vertical language models are the future and mentions a portfolio company, Avocas, doing this for corporations. The host then brings up a report that, about two weeks earlier, Google had quietly put its Gemini Nano model into the Chrome browser. The model takes about four gigabytes on the user's computer and handles tasks such as proofreading, spelling and autocomplete. The host describes this as Google "covertly installing" it.

Hold on. Yeah. Covert's a strong word, so let's not use that word.

The host rewords the claim as "without telling people, without giving people a heads up." The objecting speaker widens the point. In their view, excited commentary about every model improvement is "a waste of time" with "no ROI." The focus should be on results for end users that used to be impossible. Their examples are a decades-old open math problem solved with OpenAI's model by a human using it, a detail they call important, and drug candidates that had sat on the shelf and are now about to enter clinical trials and INDs.

The speaker argues that the host's framing feeds a "boogeyman us versus them" story. They say it is not deliberate, but it helps explain why AI is becoming like "a four letter word." They say Google is not a company that does shady things, though other companies would; they decline to name them when asked whether they mean their former employer. They suspect Google shipped the model because it is useful to users. They warn that if the hosts are careless with words, the result could be a Luddite reversal where "nobody wins."

The host does not fully back down. They deny being a Luddite and note that many people reported being surprised or shocked by the model's size and its background installation, which led some to look into privacy. The host agrees Google is not a bad actor and calls the move probably more a matter of speed than anything else. Another participant adds a joke about Google's legal caution:

There's attorney maxing, and then there's attorney maxed, and Google is probably attorney maxed.

That participant also says Americans in tech have a responsibility to advocate for AI's positive possibilities. They say it is "starting to feel or seem like there may be a CCP funded campaign against AI and data centers in America." This is stated as a suspicion, not documented. They say such a campaign would be "very logical for China" but bad for America.

The host asks who is responsible for the backlash. One option is Dario, with constant warnings that everyone will lose their jobs. Another is CEOs who blame AI for layoffs. One answer starts: "Everybody is trading their own book." The speaker says it makes sense for Dario to create the boundary conditions for a regulatory moat, since Anthropic is now big enough to benefit from being inside it. They claim the volume of alarm can be matched to successive fundraising rounds and growth stages. They call this a reasonable business strategy and Dario "quite clever." They add that if you ask a game-theory ("Nash") agent inside Claude what to do, it comes up with the same strategy.

The speaker's conclusion is that every CEO has clear incentives and is speaking for their own interests, so outside observers need to step back and see the big picture. They cite a clip of Palantir CTO Shyam Sankar, apparently on Fox News, saying to stop asking model makers what they think. He said to ask the factory worker, the doctor and the scientist who use the models, and to tell those stories.