For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 391 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 full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
There have been two true personal computing phase changes since 1980. The first started around 1977. The second started around 2024. Between them the industry called four others, web, mobile, social, cloud. None of those four was a phase change. They were telecommunications transitions and application-layer rearrangements on top of an unchanged personal-computing layer. The criterion that distinguishes a true phase change from a transition picks out exactly two events in fifty years.
The criterion: did the operating actor of the compute change.
The mainframe-to-microcomputer transition (Apple II 1977, IBM PC 1981, Mac 1984) changed it from institutional employee to individual user. The sysadmin, the batch-job clerk, the programmer at a terminal in a cooled room: those were the operating actors before. After, the same physical person at home or in an office became the actor running the loops, choosing the inputs, watching the outputs.
The agent moment is changing it again. The individual user is no longer the moment-to-moment operating actor. An agent process running on the user's behalf is. The user becomes the goal-setter and the exception-handler. The agent becomes the operator. This is not a metaphor about workflow productivity. It is a description of who is reading the API, executing the queries, parsing the responses, deciding what the next request is, holding the credentials.
By this criterion the four widely-named "paradigm shifts" between PC and agent fail. The web (1995 onward) changed how the user reached compute, over a wire instead of locally, but the user remained the operator. Mobile (2007 onward) changed the form factor and the always-on assumption, but the user remained the operator. Social (2005 onward) changed what the user did with the compute, but the user remained the operator. Cloud (2008 onward) changed where the heavy computation ran, but the user at the endpoint remained the operator. Each of those is substantive. None changed who was operating.
Form factor, distribution channel, addressable population, computing as percentage of GDP, software paradigm. Each is a defensible criterion for "phase change," and each yields a different inventory. Form factor counts smartphones in. Addressable population counts the internet rollout in. GDP-share counts cloud in. Software paradigm (Karpathy's Software 1.0/2.0/3.0 framing from Sequoia AI Ascent 2026) counts Software 2.0 in. The argument for the operating-actor criterion is that it predicts something none of the others do: structural openness or closure of the public internet.
When the operating actor changes, every closure mechanism designed for the previous actor inherits a structural mismatch. Login walls assume a body willing to remember credentials. Paywalls assume payment friction is a useful selection function. Browser fingerprinting assumes a human-with-mouse pattern. Ad targeting assumes attention with mood states and habit loops. Recommendation feeds assume a viewer with a next thirty seconds. Engagement metrics assume someone who can be bored.
The new operating actor has none of those properties. It has volume and queries.
The closure mechanisms either get redesigned for the new actor or stop working. The web's openness is not being defended by anyone. It is being structurally re-opened by the operator-population change. Even with maximally adversarial closure efforts (paywalls hardening, regulatory pressure, agent-licensing regimes), agents can own the underlying machine and present as humans. The closure machinery designed for the human reading population cannot survive the population swap without being entirely rebuilt, and rebuilding it is the negotiation we are inside now.
This is the structural reason the "open internet" arguments have a different flavor in 2026 than they did in 2014 or 2008. The previous arguments were normative: open is good, here is why. The current argument is structural: the machinery that closed the web was designed for an operator that no longer exists at scale. Defending openness is not the active project. Refusing to actively re-engineer closure is.
Per-event pricing of internet activity was a 1990s idea that recurred at intervals and never composed.
Adam Back published hashcash in 1997, a proof-of-work proposal that would have made each email cost the sender a small amount of CPU time. Seth Godin started advocating "stamps for email" the same year, a per-message penny-stamp into escrow that would burn if the recipient marked the message as unwanted; he restated the proposal in 2006 and again in 2023. Bill Gates pitched paid email at the 2004 World Economic Forum, predicting spam would be solved within two years through a monetary postage scheme. The micropayment companies of the dot-com era, Digicash and Millicent and CyberCoin, failed at retail. Each proposal was structurally correct. Per-event pricing matches per-event consumption. If reading or sending an event has a marginal cost, the abuse cases (spam, scraping, abuse of free APIs) lose their economics.
Each was premature against the compute layer of its day. The previous operating actor (the human user) could not generate per-event consumption at the volume that made per-event pricing compose. A human reads maybe a hundred web pages a day. A human sends maybe a hundred emails a day. Per-event billing at human volume is a tax on routine activity, with high relative friction and low absolute revenue per actor. The math never closed.
The agent operating actor generates per-event consumption natively and at scale. An agent answering a single question may make ten thousand reads in one session. At a millicent per read, the session bills ten cents to the goal-setter and distributes a hundred dollars of micro-revenue across the publishers it touched. The math closes because the volume per actor crossed an order-of-magnitude threshold.
Cloudflare's HTTP 402 pay-per-crawl beta, processing roughly a billion 402 responses per day in 2026, is the first commercial-scale instance. That number is itself a measurement of the new actor population. Humans are not generating a billion daily 402 responses. Agents are.
Twenty-nine years from Godin's first proposal to a serving layer that lets the proposal compose. The proposal was structurally right. The compute layer it needed had not yet emerged.
The naive read of "per-event pricing of every web read" is that the open web closes. Everything gets a paywall, readers get walled out by friction, the public surface contracts.
The structural read inverts. Per-event pricing at the agent scale removes the human friction that made paywalls necessary in the first place. The human was paywalled because remembering credentials and paying separately at every domain was friction that selected against casual reading. The agent has no such friction. It pays at the protocol level, transparently, on behalf of a goal-setter who does not see the per-page transactions. From the agent's side, every site it can read is a candidate citation source. From the publisher's side, every read produces small revenue without selecting readers out. From the goal-setter's side, the session cost is bounded and the value is the answer.
The economic equilibrium that emerges puts pricing at the publisher-server boundary (HTTP 402, pay-per-crawl, agent-API metering), not at the user-browser boundary (login walls, subscription paywalls). The agent reads everywhere. The publisher gets paid per read. The goal-setter pays for the session, not for the access.
This is why the "everything gets economized therefore the web closes" worry inverts. Free-to-read is the lowest-friction shape for being cited. Citation drives selection by the agent reader. Selection drives traffic. Traffic drives revenue under per-event pricing. Open content compounds. Walled content does not. The public web grows under economization in a way it did not grow under the previous decade's paywall-everywhere trajectory.
Most prior phase changes were named in retrospect.
Mobile was called a phase change in 2010, three years after the iPhone shipped, and the "is it really a phase change or just a faster phone" question stayed contested through 2015. Cloud was called a phase change in 2014, six years after AWS reached production scale. Social was never settled as a phase change category and got absorbed into "Web 2.0" terminology that never compressed cleanly.
The current AI moment is being called a phase change from the keynote stage in flight. Sequoia's AI Ascent 2026 framing in April delivered Pat Grady's "AI is a revolution in computation. Not faster horses, but cars" and Konstantine Buhler's "the cognitive revolution will follow the same arc as the Industrial Revolution, just bigger and faster." Andrej Karpathy presented the Software 1.0/2.0/3.0 framework from the same stage. Sonya Huang declared 2026 the year of agents from the same stage.
This earliness is the visible signal of the underlying structural difference. The prior transitions happened at the application layer. The operating-actor stayed the same; the change showed up as a new app or a new device. Mobile looked at first like phones with apps. Cloud looked at first like outsourced server racks. Social looked at first like websites with comment sections. Reading those as phase changes required years of seeing how the application layer reshaped behavior at the margins.
The current transition happens at the operating-actor layer itself. There is no application-layer ambiguity. Either the agent is the actor running the queries or it is not. Once it is, the structural consequences are immediately visible. Cloudflare HTTP 402 traffic, agent-readable manifests like llms.txt, the migration of developer documentation toward retrieval-friendly structure, the per-event-pricing experiments at the publisher-server boundary. The pattern-recognition lag that hid mobile and cloud as phase changes does not apply.
The operating-actor criterion is one criterion among several. Form factor, addressable population, computing-as-percentage-of-economy, computational efficiency, software-engineering-paradigm all yield different counts and pick out different transitions. The argument for the operating-actor criterion is its predictive power on the open-web outcome. Other criteria do not predict that outcome. A reader who weights other criteria differently will count the transitions differently.
Agents may be re-individualized to the point where they behave as a new human-scale population at the economic layer. If every agent has a billable identity, makes individual visits, and pays per page, the per-event volume per actor stops being orders of magnitude above human scale. The closure mechanisms designed for humans partially reactivate against agents. Cloudflare's identity-resolution work for agents points partway in this direction.
Regulatory hardening could close the open-web channel before the structural openness compounds. The pay-per-crawl regime is currently a market mechanism. It could harden into a closed licensing regime where a small number of model providers pre-license a small number of approved sources, at which point the open web's compounding visibility through agent citation collapses to a handful of suppliers. The structural-default trajectory is open under no intervention. Closure requires active counter-engineering. The counter-engineering may happen.
The "Godin vindication" framing assumes per-event pricing settles into a stable equilibrium where many publishers and many agents transact directly. The pricing power could consolidate to a few aggregators that meter agent traffic to publishers as a middle layer, in which case the stamps-for-email idea ends up implemented in form but captured in extraction by intermediaries.
The mainframe-to-PC analogy itself can be over-extended. The mainframe-to-PC transition took roughly a decade to compound into mainstream impact. Apple II 1977, IBM PC 1981, Mac 1984 for the consumer-level adoption window; the productivity-gain payoff lagged into the late 1980s and mid-1990s. The agent transition's compound timeline is open. The structural argument here predicts that something analogous to the productivity-payoff lag will appear; the specific shape is unknown.
The piece grants all four risks. They adjust pace and magnitude of the structural argument. They do not adjust the structural argument's direction.
What is structurally novel about this transition relative to the previous four decades is the operating-actor change. That change is what makes mobile-vs-iPhone analogies misleading and makes mainframe-vs-PC analogies precise.
The first PC phase change put compute in the hands of the individual. The second PC phase change moves compute to the hands of a process running on the individual's behalf.
Most of what gets called a paradigm shift in personal computing is not. The criterion picks out two events in fifty years. We are inside the second.
P.S. — Graph:
Sources: Sequoia AI Ascent 2026 (April 20, 2026): Pat Grady "revolution in computation, not faster horses but cars"; Konstantine Buhler Industrial-Revolution-arc framing; Andrej Karpathy Software 1.0/2.0/3.0 keynote; Sonya Huang "year of agents." Adam Back hashcash proof-of-work (1997). Seth Godin stamps for email (1997 first proposal, 2006 restatement, March 2023 revisited at seths.blog/2023/03/revisiting-stamps-for-email). Bill Gates 2004 World Economic Forum Davos paid-email proposal. Apple II April 1977, IBM PC August 1981, Mac January 1984. Cloudflare HTTP 402 pay-per-crawl beta (2026), ~1B HTTP 402 responses/day per Cloudflare Radar. AWS production-scale by 2008.