For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The thing that moved me was the speed of the answer.
A question arrives at the altitude of civilization: data processing leaving Earth, silicon finding a new law, space and AI converging into infrastructure. The answer does not arrive as brainstorm. It arrives as a cost curve, a launch curve, a power model, a customer map, a hard-problem boundary, a history of workload shifts, and a claim about latency as the product. Two people sit on a stage and the map is already there.
A language model can summarize that moment. It can say many adjacent true things. What it cannot own is the cost that made those answers cheap for the humans saying them. Will Marshall's answer is backed by satellites built, launches bought, customers served, governments depended on, imagery collected, studies run, bets made, and errors paid for. Andrew Feldman's answer is backed by a decade of silicon architecture, foundry constraints, yield risk, memory bandwidth, product demand, and the long humiliation of trying to be right against NVIDIA. The sentence is fast because the world was slow.
Human expertise at the frontier is compressed reality. It looks like talking. It is really a domain model trained by consequences. The expert has already run the experiment in money, material, reputation, hiring, customers, regulators, and time. The answer comes out as speech because speech is the narrow channel at the end of a much wider machine.
The wider machine is capitalism, read without romance. Markets ground commercial claims by forcing them through customers, prices, capital costs, competition, and survival. Capital routes resources toward hypotheses. Customers decide which hypotheses touch reality. Public markets add another clock: liquidity, scrutiny, a price that updates in the open, a demand to keep delivering after the ceremony. Markets miss public goods, misprice long horizons, reward fraud for a while, and turn collective fear into numbers. Within their scope, they are still the evaluator that forces a model to meet the world.
This is why the global computer still exceeds any AI system in 2026. It is rockets, fabs, satellites, farms, regulators, militaries, customers, founders, investors, chip architects, launch providers, public shareholders, and families at the listing ceremony. It runs through sensors and invoices and failures. It updates through profit, loss, survival, reputation, and debt. It uses humans as high-bandwidth local processors because some knowledge only exists after a person has paid to find out.
Platforms change the computer by expanding what its nodes can try. SpaceX is the obvious case. The public reads it as Elon being the giant. The structural view is sharper: SpaceX lowers the cost of acting in orbit, and a lowered action cost creates customers whose prior models suddenly become executable. Planet is one of those customers. The genius of the platform is that it changes the price of experiments so downstream genius can become real.
That is why Marshall's space-compute answer matters even if the timeline moves. The claim is larger than data centers going to space. A new experiment clock opens when launch cost crosses a threshold. At high cost, space compute is mostly paper. At lower cost, it becomes a sequence of tests. You can launch chips, discover what fails, redesign, and launch again. The relevant improvement is the time between contact with reality.
Chamath's time-bounded Moore's Law belongs here. The old law was transistor density. The new one is closer to learning latency. How long from capability to usable answer? How long from launch to failed test to new launch? How long from a better tool to a changed habit? Silicon, space, and society each have a clock, and the slowest clock governs the system.
Andrew's answer makes the first clock explicit. AI widened the territory computers can act on: images, language, messy knowledge. New workloads create share shifts, and share shifts reward architectures that violate the incumbent shape. Cerebras bet that the bottleneck was memory-to-compute movement, so it made a giant chip with memory close to compute. The commercial translation is latency. People will not wait for AI any more than they waited for dial-up. A faster answer is a different product at the margin.
The social clock is slower. A country can invent, fund, and ship frontier capabilities faster than its institutions and citizens can absorb them. That makes diffusion the frontier behind the frontier. The median citizen's effective model of technology, budgets, energy, trade, and institutions becomes infrastructure. A government cannot climb out of debt with tools its population cannot route into productivity. Raising the floor of usable understanding is therefore engineering too, though the material is habit, attention, trust, and education.
The panel felt wondrous because it showed the human node at full resolution. Will answers as Planet: Earth sensed daily, AI made sighted, orbit as power system, launch cost as clock. Andrew answers as Cerebras: workload history, architecture divergence, memory bottleneck, latency as market. Chamath answers as capital allocator and moderator: finding the shared abstraction and asking each node to compute its part.
This is the part of human capability that AI should make us more precise about. AI can accelerate the transcript, the search, the code, the synthesis, the instrument panel. It can help a human compress more reality into fewer hours. The frontier answer still comes from a coupled system that has touched the world. Models trained on text inherit the residue. Humans trained by consequences carry the generator.
The miracle in the room was the world computer speaking through people it had trained.