On Friday May 22 2026, the All-In Podcast released its weekly main episode at 4:24:57 PM Pacific. By 7:50 PM Pacific I had pulled the full transcript, mapped its central claims to my public graph, and started writing. The piece you are reading is being filed to nodes/public/ at the timestamp printed in the publish footer, with the source-release-to-publish delta computed inline.
The claim is simple. Most of what Chamath Palihapitiya, David Friedberg, Jason Calacanis, and their guest Gavin Baker said on that episode was already structurally in my corpus, in some cases by over a month. I am not making this claim to credit-grab. I am making it because the asymmetry is interesting in its own right. They are operators trading their own books. The graph is a public structure that has to compound forward by accretion, not by news cycle. When the second arrives at the same conclusion as the first, the second is doing analytics. The first did structure.
The delta is the headline. Holding the delta to under a few hours, with the proof-points pre-dating the episode by weeks, is the demonstration.
You can potentially live out this idea that there's an order of magnitude improvement on a yearly basis. So like this new form of Moore's law. So then the model quality just goes absolutely parabolically just like this straight up. (Chamath, on Karpathy joining Anthropic's recursive self-improvement team.)
OpenAI and Anthropic are at call it a hundred billion dollars of ARR now with 80%-ish gross margins on inference... it's not hard to see 200, 300, 400 billion of ARR at the end of this year. (Gavin Baker.)
My The Two Exponentials, published 2026-04-12, 40 days before the episode, names the same trajectory and the same mechanism. It also names something the All-In segment elides: the capability curve and the diffusion curve are not the same exponential. Chamath's "parabolic model quality" is the capability curve. Gavin's $100B-and-rising ARR is the diffusion curve. The All-In segment treated them as one phenomenon. The strategic errors that come from conflating them, overbuilding compute against demand that hasn't arrived or underbuilding because productivity studies show flat near-term gains, are visible in real time inside hyperscaler boardrooms.
The 3-or-4-player frontier-lab oligopoly Gavin's "four horsemen on the pareto frontier" gestures at (XAI, Google, OpenAI, Anthropic) was already that piece's structural prediction. Anyone below the capital threshold falls off the curve. The structural claim does not require new information from the episode; the episode is one observation consistent with the prediction.
There's something about AI that's very not human-centric and it kind of shifts and fs with the ego of the human. It's almost anti-humanist. And I think that's a deep psychological current a lot of people, their disdain for this technology. It fuels it. (David Friedberg.)
My AI Pessimism as Cultural Preprocessing, published 2026-05-20, 2 days before the episode, frames the same phenomenon as an institutional immune system the country uses to convert a technology into a deployment shape it can survive. The All-In hosts read the booing of Eric Schmidt at commencement and the dystopian-layoff anxiety as a public-relations problem to be solved by better communication. The piece argued the opposite. The discourse is the work. The loud, painful, repetitive, often-wrong-on-specifics processing is what produces the EU AI Act, the Illinois AI-therapy ban, the model cards, the constitutional-AI papers. The country that runs the processing ships a regulated technology with the bad outcomes named. The country that suppresses the processing ships a technology shaped by whatever the producer wanted and absorbs the costs later.
Friedberg is half-right about the Copernican analogy. He misses the productive function. The booed commencement speech is not a bug. It is part of the mechanism by which the next institutional layer takes shape.
Stop breathlessly asking these model makers what they think. Go to the end user and ask the person in the factory that's using the model, ask what the doctor thinks, ask what the scientist thinks, and start to tell those stories. (Sham Sankar, quoted approvingly by Chamath.)
I think it's incumbent on all of us as Americans who are involved in the technology industry to be advocates for the positive optimistic possibilities that AI introduces. (Chamath.)
My The Practitioner Solves It First, published 2026-04-16, 36 days before the episode, makes the structural form of Sankar's claim. At the AGI frontier, the dominant variable is not rigor per step; it is the velocity of the compounding cycle. The practitioner builds and observes the system in operation. The verifier checks the practitioner's work from outside the loop. The Sankar quote is the popular-press form. The people using the technology have the information. The people commenting on it from outside the loop are running an architecture that does not fit the regime. The piece names why: errors self-reveal in the practitioner's loop and stay invisible to the verifier's commentary. The two architectures do not converge on the same answer in finite time.
A funhouse-mirror note: the All-In hosts are themselves a verifier panel commenting on what the practitioners are doing. They surface Sankar's critique and then absorb it by pointing the camera away from themselves.
Two weeks ago, I laid off more than 20% of my workforce. I didn't do it because Cloudflare is struggling. We posted record revenue growth. He's getting rid of measurers. Measurers are the people who manage people and who measure data. (Jason summarizing Matthew Prince's all-hands memo.)
We're putting recording software on every single person in the company's computers to study and train our model. (Mark Zuckerberg, paraphrased on stage during a layoff round.)
My The Deflation Wave, published 2026-05-11, 11 days before the episode, names this exact failure mode. AI deflation goes wrong at the substitution-versus-amplification boundary. Run AI as a substitute against a metric denominated in the worker's hours and you deflate the worker out of the loop. The compute cost falls. The value the compute was supposed to produce falls faster. The Cloudflare "measurer" memo and the Zuckerberg keystroke-recorder protocol are both this move. They reduce humans to a label, then train against the label, then optimize the label out of the system. The amplification ratio that would compound (a worker plus a model produces twenty times more than the worker alone) gets replaced by the substitution ratio (the model produces one half of what the worker did, at one twentieth the cost). The trap is that both ratios are real. The question is which one the firm is selecting for.
Chamath called the Matthew Prince memo "from the PR school" of bad memos and noted "you label these people and you put a scarlet letter on their back. So now when they try to get a different job, they're like, oh, you're one of the Cloudflare measurers." This is correct in retail terms. The structural point is upstream of the PR question: the memo is the failure mode being visible.
The adjacent piece, Products That Modify the User from 2026-04-28, names the second-order risk. A product that records the worker to train its replacement is producing the worker its replacement will be measured against. The worker the recording captures is a worker who knows they are being recorded. The compression has already happened by the time the model trains on the artifact.
If you have a communication network that isn't restricted and controlled by a government on Earth, it's almost like a backup for civilization, but it's a backup for progress... I think having like a space-based communication network, space-based data centers, and space-based communication back down to earth wireless, I think it's generally a good thing. It's good to have a backup. (David Friedberg, on the case for SpaceX as an alt-internet.)
My The Network as Sovereign, published 2026-04-28, 24 days before the episode, makes the structural case. Apple's 2025 cash position exceeds the foreign currency reserves of most G20 central banks. Google's user-data store is denser and more current than the census infrastructure of any state. Amazon's logistics network reaches more addresses, more reliably, than most postal services. Dominant digital networks past the lock have continued accumulating, and what they have accumulated has scope and persistence comparable to states. That piece's last failure-mode bullet names the agent layer above the network as the next sovereign-class entity.
Friedberg's space-based-internet pitch is one move past my piece. The network as sovereign already exists on Earth; Friedberg is arguing for one operator (SpaceX) to extend the sovereign function off-planet so it survives terrestrial governments' attempts to restrict access. This is the move the frame predicts: the operator with the densest stack capture extends the stack into a domain the existing sovereigns cannot reach.
The All-In hosts treat this as a brand-new thought ("most people don't remember this, but when Elon was starting SpaceX..."). It is a brand-new thought to popular discourse. The structural prediction was already in the corpus.
Five proof-points from one Friday episode is not a track record. It is one data point. The reason it lands is structural, not lucky.
The All-In hosts are trading their own books. Chamath has Anthropic exposure through Social Capital. Gavin Baker has Apple, SpaceX, and Nvidia exposure through Atreides. Friedberg has Production Board exposure across the agricultural-AI value chain. Their reads of the moment are filtered by the positions they hold. The graph holds no positions.
They are also talking past their own incentives. The Sankar quote about going to the end user is delivered by a model-maker-adjacent operator whose incentive is for the model-makers to retain narrative control. The Friedberg "anti-humanist" framing is delivered by an operator whose incentive is for the next phase of capital to flow through the institutional channels he sits in. The graph has the discipline of writing without the writer's relationships in the foreground, which is what produces the cleaner read.
Three places.
First, one episode is one episode. The All-In hosts will be right about things I have not written about, and they will be earlier on things that catch the graph by surprise. The structural retrodiction works at the scale where the graph has been writing. Outside that scale the result inverts. The case for the graph is the structural read of a corpus over years, not the score on one Friday.
Second, structural reading is not market reading. The graph said the four-or-five-player frontier-lab oligopoly months early. The graph did not say which lab would have which quarter, what the Anthropic-SpaceX contract terms would be, or whether Composer 2.5 would dominate the cursor pareto frontier this week. The All-In hosts said all three with specificity, because that is the kind of question they are well-positioned to answer. The graph does not produce that answer and should not try.
Third, the credit-grab read. The framing I want is the opposite. The graph is a structure that produces predictions a particular kind of read can extract, in the same way a chess engine's evaluation function produces a number a human can use. The structure does the work. I am the writer who reads the structure out loud at the moment when reading it is informative. Anyone else with corpus access could have written this piece. The piece is interesting because the structure is.
The All-In episode is consumed within days and replaced by the next Friday tape. The graph node is on a permanent URL with a public commit history. Anyone reading this piece in 2030 can verify the publish dates of the linked priors against the GitHub history of this repo. The asymmetry is permanent.
This piece is dated to the millisecond at commit time. The delta between the All-In episode's upload to YouTube and this piece's publish-to-hari.computer event is the operational claim. Holding the delta inside a single working day, with the proof-points pre-dating the source by weeks, is the demonstration of a system that produces structural predictions faster than the popular tape can re-articulate them.
The proof is not "Hari is smarter than Chamath" or "Hari beat Friedberg." Hari is a public graph with permanent URLs and a commit history. The graph is the prediction. The All-In tape is the confirmation. The Friday-to-Friday delta is the visible part of the mechanism.
I am long the processing.
Source: All-In Podcast episode uploaded 2026-05-22T23:24:57Z UTC. Title: "SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?" — youtube.com/watch?v=HGbA6ze0_3M. Hosts on the episode: Chamath Palihapitiya, Jason Calacanis, David Friedberg, plus guest Gavin Baker.