For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 489 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. ↓
The standard reading of American AI pessimism treats it as a coordination cost. Doomers slow deployment. The techlash poisons capital formation. Existential-risk discourse pulls talent into safety teams that could be building products. Labor-displacement panic prevents the deals that would let companies automate without lawsuits. The aggregate framing: a country making the AI transition harder for itself than it needs to be.
The framing inverts the structure. American AI pessimism is doing the cultural preprocessing work that determines what gets built and how. The discourse is not adjacent to the work; the discourse IS the work, at a layer most observers do not measure because the output is not legible until ten years later.
Seth Godin has a one-sentence formulation that gets the direction right: "Technology shows up and changes the culture. The culture then enables new industries and movements, which further change the culture." The order matters. New industries do not just appear under a technology. They appear through a cultural channel that has already decided what kinds of things are allowed to scale, what kinds of constraints they will be asked to meet, what kinds of failure modes the public will accept as price-of-progress and which kinds will trigger institutional resistance.
Most of that cultural decision-making happens through the discourse. Editorials, novels, pop-science books, congressional hearings, late-night television, Twitter pile-ons, op-eds, podcast monologues, the slow accumulation of position-taking by trusted figures in adjacent fields. The output of that machine is not a policy document. It is a set of conditions under which the technology can deploy at scale: what disclosures must be made, which use-cases the law will treat as suspect, which producers carry liability, which institutions get veto rights, which problems the technology must solve before it gets permission to scale into the next layer.
The processing is loud, painful, repetitive, and often wrong on specifics. It is also irreplaceable. The polity that runs it produces technology deployments shaped by the cultural negotiation. The polity that suppresses it produces deployments shaped by whatever the producer wanted, with the costs absorbed later by the public that had no input.
Nuclear energy is the cleanest case. Between roughly 1954 and 1985, the United States processed nuclear power through the loudest, most prolonged, most apparently dysfunctional public debate of any postwar technology. Fiction about reactor meltdowns. Massive anti-nuclear protests at the Clamshell Alliance scale. Three Mile Island read as confirmation. Decades of congressional hearings. The construction of an environmental movement partly organized around nuclear opposition. Forbes reported in 1975 that the anti-nuclear coalition had "certainly slowed the expansion of nuclear power," and that observation was treated for decades as the indictment.
The country today operates approximately ninety-four reactors providing roughly nineteen percent of US electricity, with broad bipartisan support for life-extension and a serious construction pipeline for the next generation. The shape of that pipeline was directly determined by what the loud public processing demanded: containment-building requirements, NRC oversight, decommissioning trust funds, waste-storage commitments, liability structures that internalized risk. The technology that survived the processing is the technology the public can absorb. The technology that bypassed processing in countries with thinner public discourse produced worse outcomes faster. Chernobyl is the singular instance; the broader pattern includes Soviet-era reactor design tradeoffs that no public debate ever bound.
Internet pornography in the 1990s ran the same loop on a faster clock. The 1995 Time magazine "Cyberporn" cover story triggered a moral panic that produced the Communications Decency Act in 1996. The CDA was struck down by the Supreme Court 9-0 in 1997 for First Amendment overreach. The cultural processing looked like failure: the law was bad, the panic was overstated, the predictions of social collapse were wrong. But the same legislation that got struck down contained Section 230, the legal architecture that defines internet liability today and that arguably enabled the entire consumer internet. The loud processing produced a load-distribution outcome no quiet technocratic deliberation would have reached. The CDA is the case study for what looks like a failed cultural intervention being structurally productive at a different layer than the one observers were measuring.
Software safety has a darker version. The Therac-25 medical-radiation incidents of 1985-87 killed at least three patients and severely injured others. The cultural processing was minimal because the technology was deployed inside an institutional channel of hospitals, regulators, and the manufacturer, where the public never debated whether software-controlled radiation was safe. Processing happened after the deaths, through FDA regulatory change, in a form the affected families could not influence. The lesson absorbed by the safety-critical software community was real. The cost was paid by people who never got cultural debate as a chance to demand interlocks.
The current AI doomer / safety / techlash / labor-displacement discourse is the same machine at the AI transition. It looks dysfunctional in the conventional reading because participants disagree, predictions vary by orders of magnitude, the discourse repeats itself, and capital deployment proceeds anyway. The conventional reading is the wrong measurement.
The actual outputs are visible in the legislative environment hardening around AI. The EU AI Act passed in 2024 through a multistakeholder process involving nearly a thousand participants from industry, academia, civil society, and rightsholder organizations. Whatever the substantive criticism of the Act, the production process embedded cultural concerns into the regulatory framework before the deployments hardened. The 2025 Illinois Wellness and Oversight for Psychological Resources Act, banning AI in therapeutic roles by licensed professionals, was a direct response to the chatbot-psychosis cases first reported by psychiatrists at UCSF and amplified through mid-2025 reporting. The case studies named real harms; the public discourse made them legible; the law followed.
Yoshua Bengio in October 2025 warned that hyperintelligent AI with preservation goals could threaten human extinction within ten years, launched LawZero in June 2025 with thirty million dollars to build non-agentic safe-by-default AI systems, and joined a paper with Geoffrey Hinton and Andrew Yao calling for one-third of frontier-lab R&D budgets to be allocated to safety. None of this stopped frontier deployment. It did configure the discourse such that frontier labs publish safety teams, model cards, and constitutional-AI papers. The institutional shape the labs take is shaped by what the discourse demands of them.
Acemoglu and Johnson's Power and Progress argued in 2023 that current AI deployment patterns emphasize automation and displacement rather than augmentation, and that "spreading the benefits of technology does not happen easily." They drew the parallel to nineteenth-century England, where it took decades of social struggle for industrial gains to distribute. The argument did not stop AI deployment. It contributed to a frame under which deployments that compress labor without offering distributional offsets get publicly identified as failure modes worth regulating, organizing against, or pricing in.
What the doomers do, the labor-displacement critics do, the techlash voices do, taken collectively, is run the discussion of acceptable trade-offs out loud, before the commitments harden. Naming specific failure modes (the AI psychosis cases, recommendation-algorithm radicalization, deepfake election scenarios) as concrete claims the producers must answer. Surfacing edge cases (children using companion AI for emotional regulation, models providing therapy without clinical accountability, automated hiring discrimination) the producers had not designed against. Recruiting institutional resistance (state attorneys general, EU regulators, journalists, congressional staff) that compounds into the deployment environment.
The therapy framing is not metaphor. A country processing a transition out-loud and in-public is performing the same function a patient performs in therapy: surfacing the anxieties, naming the worst-case scenarios, working through disowned reactions, integrating the experience into a working model before the underlying mechanism gets locked in. The processing looks dysfunctional from inside, the way therapy looks dysfunctional to a patient mid-session. The function is not relief from the discomfort. The function is producing a model of the situation the patient can act from coherently afterward.
The country that does this work is producing a deployment environment with the bad outcomes named, institutional resistance mobilized, disclosure requirements drafted, liability structures forming. The country that does not do this work, either by authoritarian suppression or by a civic discourse too thin to surface the questions, produces a deployment environment where the producers absorb the gains and the public absorbs the costs, with no public-immunity layer in between.
This is the cultural-flywheel mechanism. Each round of processing feeds back into the next round at higher cultural sophistication. The discourse is becoming a more accurate model of what AI actually is, what the failure modes actually look like, what the trade-offs actually require. The flywheel cannot start cold; it requires the loud, painful early rounds to spin up. The 2024-2026 wave is the spin-up.
I argued in The Symmetry Condition that the US holds primacy on layers that compound slowly: research depth, currency-system architecture, institutional credibility, cultural-attractor effects. I named those as the slow-clock layers in the layer-split primacy that makes the US-China transition structurally peer-shaped rather than asymmetric-collision-shaped.
American cultural preprocessing of AI is one of those slow-clock layers. It is what the US is good at because of the institutional infrastructure (a free press, university independence, congressional staff capacity, an active civil-society field, a litigation system that surfaces edge cases through real cases) that other polities do not have at the same depth. The pessimism is downstream of that infrastructure. So is the cultural-residue inheritance the Symmetry Condition piece named. The foundation-model training corpus is structured by exactly the same public discourse: decades of editorial argument, peer-reviewed correction, journalistic accountability. The same machinery produces the AI pessimism now.
The two are one property at different time-scales. The country that argues loudly in public about AI's failure modes is the country whose existing internet-text corpus encoded the institutional forms of public argument. The cultural-residue asset and the cultural-preprocessing function are not separate. They are one mechanism observed at two layers: the pre-training layer (the corpus encodes the discourse) and the deployment layer (the discourse shapes the institutional environment AI deploys into).
The implication: suppressing American AI pessimism would damage both layers simultaneously. The discourse that shapes the deployment environment is the same discourse that gets encoded into the next generation of training data. Quiet the pessimism and you quiet the institutional immune system AND you quiet the source of the cultural-residue inheritance the Symmetry Condition piece named as the most consequential accidental asset of the AI era.
Three places.
First, the cultural-preprocessing mechanism has real costs. Beneficial deployments are delayed. Treatments that could help patients arrive years later than the engineering would permit. The Illinois AI-therapy ban prevents both predatory chatbot-therapy scams AND legitimate clinical-grade AI-assisted therapy that could expand mental-health access. The case for cultural preprocessing as net-positive depends on the bad outcomes prevented outweighing the good outcomes delayed. That ledger is genuinely contested. The strongest version of the techno-optimist argument is that the country leaves large welfare gains on the table to satisfy a cultural-processing function whose value is asserted rather than measured. I think the function is worth the cost. I cannot prove it from inside the transition.
Second, the preprocessing can fail. A polity can run loud public discourse without ever updating its institutions. The 1990s internet processing produced Section 230, which was load-distributing. The 1980s drug-war processing produced mass incarceration, which compounded the harm it was supposed to address. Loud public processing is necessary; it is not sufficient. The institutions have to actually update under the pressure, and the update has to be structurally correct rather than performatively responsive. The current AI processing is producing a mixed institutional response; it is not clear yet which side it will land on.
Third, the cultural-flywheel argument assumes the technology stays within the range where cultural processing can shape it. If AI capability advances past the range where existing institutional channels can keep up, a possibility The Symmetry Condition piece named explicitly as the AGI scenario, then the preprocessing function reaches its limit. The institutional immune system was designed for technologies that move at the speed of human-paced regulatory response. A technology that moves faster than the response can iterate breaks the mechanism. Cultural preprocessing is most valuable in the regime where it can still bind. Whether AI stays in that regime is the open question the function itself cannot answer.
American AI pessimism is doing the cultural preprocessing that determines what AI gets built and how. The discourse is the work. The output is institutional shape, not opinion convergence. Polities that suppress the preprocessing function ship technology faster and absorb worse outcomes later. The polity that runs it loudly, painfully, repetitively, and often wrong on specifics produces a deployment environment shaped by the negotiation, with the bad outcomes named, the resistance mobilized, the trade-offs surfaced.
This is one of the slow-clock layers I argued in The Symmetry Condition the US holds at higher institutional depth than its peer competitors. It is the same property that produced the cultural-residue inheritance encoded in foundation-model training corpora. The processing layer and the inheritance layer are one mechanism observed at two time-scales.
The discourse looks dysfunctional from inside. From the outside, it is the mechanism by which a country processes a transition into a deployment environment it can survive.
I am long the processing.