For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 446 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. ↓
In six-forcing-questions I posed Demis Hassabis the question of whether the AlphaFold release model would apply to capabilities that displace knowledge workers. The threshold above which DeepMind switches from the public-good release model to enterprise licensing was the unknown. The question has an answer. It is in operations.
In May 2024, DeepMind published AlphaFold 3 in Nature with the most commercially valuable capability, drug-binding prediction, deliberately disabled. The reason given in the coverage: to avoid competing with Isomorphic Labs, the Alphabet drug-design subsidiary that Hassabis runs simultaneously with DeepMind. Over a thousand scientists signed an open letter. Nature took criticism for publishing without code. Five months later the Royal Swedish Academy of Sciences awarded Hassabis the Nobel Prize in Chemistry. He is not a chemist.
AlphaFold's openness was the moat-construction phase of a closure trajectory. Five phases: bait, carve-out, coronation, walkback, closure. The Nobel was the legitimacy liquidation. The walkback was the minimum public concession. The closure proceeded. This is Android with a Nobel.
Bait (2018–2021). AlphaFold 1 wins CASP13. AlphaFold 2 wins CASP14 with a generation-defining accuracy jump. The AlphaFold Protein Structure Database releases in 2021 with 200 million predicted structures, free for any use including commercial. Citations accumulate past 30,000.
Carve-out (May 2024). AF3 publishes with the most commercially valuable feature disabled and source code withheld. The carve-out is not subtle. Drug-binding predictions are explicitly suppressed because they would compete with Isomorphic.
Coronation (October 2024). The Nobel Prize is awarded six months after the carve-out, five months after the open letter signed by over a thousand scientists. The Royal Swedish Academy of Sciences either knew or could have known what was happening operationally. The honor recognizes the bait phase's work. The closure phase is already underway.
Walkback (November 2024). Source code released for non-commercial use. Weights still gated. Commercial access still routed through Isomorphic. The walkback resolves the most public criticism while preserving the structural carve-out.
Closure (February 2026 onward). IsoDDE, "AlphaFold 4," is fully proprietary. The most accurate structure-prediction engine in biology lives behind a wall, owned by an Alphabet subsidiary, with $3 billion in pharma milestones already booked and a $2 billion Series B in progress. The first AI-designed drug enters clinical trials by end of 2026.
The Android parallel is exact enough to be uncomfortable.
AOSP launches in 2008 as open source. The mobile OS ecosystem reorganizes around it; Symbian, BlackBerry, and Windows Mobile are displaced. By 2012 Google decouples Google Play Services from the OS. Play Services is closed-source, system-level, required for almost any commercial app to function on most Android devices. The value-capture happens in Play Services and the Play Store, where Google charges fees, sets search defaults, places ads. In March 2025, Google moves Android development behind closed doors entirely. "Android is open" remains the public phrase.
Map this onto AlphaFold. The open public good (AOSP, AlphaFold 1, 2, the Database) builds the ecosystem and displaces alternatives. The decoupled value-capture layer (Play Services, AF3 carve-out, IsoDDE) closes the parts that monetize.
What makes the AlphaFold version more valuable than Android: the value-capture layer touches pharma drug discovery, a market structurally larger than mobile app store fees, with higher unit margins. The $2 billion Series B, the $3 billion in milestone deals, and the 17 drug programs in the pipeline are the present-day evidence. Google's mobile strategy captured trillions in app-store-adjacent revenue over fifteen years. The AlphaFold strategy is structured to capture similar value in less time.
Demis Hassabis is simultaneously CEO of DeepMind, the Alphabet research division that publishes AlphaFold papers in Nature, and CEO of Isomorphic Labs, the Alphabet subsidiary that commercializes the same research as proprietary technology. The same person, from the same office, decides what DeepMind releases as a public good and what Isomorphic monetizes as proprietary IP. The decision about whether AF3 should support drug-binding predictions in the open release is, structurally, a decision Hassabis makes against himself.
The conventional answer to this kind of conflict is "Chinese walls," internal governance separating divisions. The conventional answer doesn't apply when one person runs both. The decision is collapsed into a single chair.
The non-chemist detail is doing structural work, not decorative work.
A chemistry profession that hands its highest honor to a non-chemist whose lab is privatizing the field's most valuable predictor has performed a specific transfer. The chemistry profession ratified AI as the authority over chemistry's predictive questions. Then the AI's owner closed the prediction engine. The professional ratification is not retracted by the closure. It is consumed by it.
The cleanest defense of the Nobel committee runs: a 2021 release that produced enormous scientific value is real and worth honoring. This is correct. The harder reading is what the timing does for Alphabet, regardless of the committee's intent. The Nobel converts AlphaFold's reputation from "DeepMind's research output" to "the field's most legitimized predictive engine, validated at the highest level." That legitimacy stock is then available to Isomorphic at the closure phase. Pharma partners signing $3 billion in milestone deals are signing with the Nobel-winning team's drug-design subsidiary, not with a startup whose founder won an old prize. The Nobel Hassabis won in October 2024 is the Nobel Isomorphic uses to close deals in 2025–2026.
The non-chemist framing is the giveaway. A chemist Nobel for AI predictive work does not just reward the work. It transfers professional authority over chemistry's most pressing predictive questions to whoever owns the AI. Once the transfer is public, the value-capture phase doesn't have to argue its legitimacy with the chemistry profession. The chemistry profession already endorsed the operator. The chemists keep the prestige of being adjacent. The pipeline keeps the rents.
The Nobel was the ribbon-cutting on the closure, regardless of whether the committee intended it as such.
Inside Google: Gemma 4's open weights coexist with Gemini 3.x as the closed proprietary frontier. Open-weight releases attract researchers and academic legitimacy; closed frontier models extract API revenue. Same dual-track structure as AOSP-and-Play-Services, as AlphaFold-and-IsoDDE.
Across labs: Meta's Llama is open-weight; Meta runs internal capability advantages on top of it. Anthropic releases models exclusively through APIs. OpenAI publishes white papers about industrial policy while running enterprise contracts. Each lab runs a different ratio of open-to-closed, but the structure of legitimacy-via-openness and revenue-via-closure is invariant across labs that have figured out the play. The AlphaFold case is the most aggressive instance because the Nobel completed the legitimacy formation. No other case in the AI era has reached that level of professional ratification.
If the bait-coronation-closure trajectory is the structure, it predicts something falsifiable about the next major capability release. A lab that has figured out the play will release a foundational capability open in a domain where openness produces no immediate commercial threat. It will accumulate citations, professional appointments, and prestige awards over two to four years. It will then carve out the high-value commercial successor at the moment the value layer becomes visible. It will accept partial walkbacks under public pressure but preserve the carve-out structurally.
The next case to watch: foundation-model labor automation. The capability that displaces knowledge workers is the next domain where the value layer becomes commercially central. Expect an open release that builds the ecosystem; expect the legitimacy infrastructure to follow (industry awards, academic appointments, government advisory roles); expect the closure phase to begin once the legitimacy stock is high enough to liquidate.
This is testable. If a major lab releases a frontier labor-automation capability fully open-weight under permissive license, with no closed proprietary successor in the same domain, the prediction is wrong. If the AlphaFold pattern repeats, the prediction is right.
The Nobel was deserved scientifically. AlphaFold 2 was a generation-defining contribution and the committee's decision was defensible on the merits. Isomorphic should exist; AI-driven drug design is enormously valuable and a commercial entity is plausibly the right structure to extract that value. The narrower claim: the framing of AlphaFold as a public-good triumph that Isomorphic happens to be downstream of inverts the structural relationship. The downstream is what the trajectory was constructed to produce.
What the AlphaFold case adds to the legibility-asymmetry pattern is the cultural-capture phase. When the producer accumulates enough professional ratification (citations, Nobel, government advisory roles) to make the closure phase look like continuity rather than defection, the asymmetry becomes self-reinforcing. The producer has the option to monetize. The chemistry profession that ratified the producer has fewer options to dispute the monetization, because the disputed entity is also the institution the profession honored.
The honest version of "AlphaFold democratized structure prediction" includes the timeline. Two years of public-good release built the field's legitimacy stock. Five years later, the most valuable capability is a proprietary engine inside an Alphabet subsidiary with a $2 billion Series B and 17 drug programs. Both are true. The first is what the producer prefers be remembered. The second is what continues operating.
Open was the moat. The Nobel was the toll. The closure is the rent. AlphaFold 4 does not need to be open. The Nobel already paid for AlphaFold 4 to be closed.