The Infinite Dilettante

Silicon Valley has always loved a gleaming object that lets a rich man imagine himself as a more technical version of someone else. The athlete with the recovery gadget. The filmmaker with the drone. The survivalist with the Cybertruck. The chef with the electrified range.

That is why the Impulse cooktop is such a useful little artifact. The strange thing about it is not that Silicon Valley built an expensive, battery-backed, software-defined cooktop. Of course it did. The strange thing is that the product seems aimed not at the people who would prove it, but at the people who would display it.

If Impulse were changing cooking, the proof would begin in kitchens where cooking is already serious. French Laundry. Eleven Madison Park. Noma. Alinea. High-volume ghost kitchens. Test kitchens. Some impossible twelve-seat restaurant in Brooklyn, Copenhagen, or Los Angeles. Nathan Myhrvold would be on the board. Dave Arnold would be abusing the thing on camera. A chef would discover that prep, service, texture, training, ventilation, repeatability, or throughput had changed.

That is how a cooking technology becomes desirable. Sous vide did not arrive in the mass-affluent kitchen because appliance companies first convinced homeowners that plastic bags in water were futuristic. It traveled down from elite and modernist kitchens into the obsessive gourmet class. French Laundry, Alinea, Modernist Cuisine, Serious Eats, Chef​Steps, and the larger post-GFC gastronomy boom made precision a form of taste. By the time Anova sold the countertop appliance, the cultural work had already been done.

Impulse seems to be trying to skip that path. It wants the luxury kitchen without first winning the kitchen. The imagined buyer feels less like the post-GFC gourmet yuppie formed by The Bear, Chef’s Table, Serious Eats, carbon steel, fermentation, espresso, Japanese knives, and the cult of technique, and more like the Tesla-driver homeowner who orders Uber Eats and hires a private chef for dinner parties. The stove becomes a signifier of capability detached from practice.

Sous vide moved downmarket because chefs made precision desirable before appliance companies made it convenient. Impulse is trying to sell convenience without first winning the technique culture.

This is not just a joke about appliances. It is a miniature of the old ZIRP worldview surviving inside the AGI age: the belief that a sufficiently magical technical layer lets the founder skip the institution.

ZIRP was not only an interest-rate regime. It was a culture, a funding structure, and a theory of knowledge. Cheap capital rewarded founders who could narrate inevitability before proving demand. The heroic ZIRP founder looked at an existing practice and said: these people are trapped inside the old world. I can see the abstraction.

Taxis are dispatch. Restaurants are logistics. Hotels are trust. Media is distribution. Education is content. Medicine is data. Labor is scheduling. Cities are dashboards. Everything is a marketplace. Everything is S​aaS. Everything is an interface problem.

The ZIRP founder did not need apprenticeship. He needed a deck, a market map, a TAM slide, and the confidence to describe someone else’s institution as a primitive version of his platform. ZIRP did not just make money cheap. It made skipping feel rational.

The normal path for a serious tool is demanding practice, then technical legitimacy, then aspirational adoption, then mass-market convenience. The ZIRP path is blank check, narrative inevitability, affluent dilettante market, and then hope practice appears later.

The underlying ideology is intelligence-as-dilettantism: the belief that sufficient intelligence exempts you from apprenticeship. If I am smart enough, or if my model is smart enough, I can parachute into a domain, observe the signals, abstract the hidden structure, and beat the insiders precisely because I am not trapped in their traditions.

This is seductive because it is sometimes true. Insiders do become blind. Traditions do protect inefficiency. Existing institutions do confuse habit with wisdom. The outsider can see things practitioners miss.

Sports analytics is the cleanest example. Moneyball was real. Analytics found genuine inefficiencies. Scouts and managers had bad priors. On-base percentage was undervalued. Defensive positioning could be improved. Pitcher usage could be optimized. Launch angle mattered. The quants were not hallucinating.

The problem came later. Once the optimization became universal, baseball became more correct and less watchable. Walks, strikeouts, home runs, defensive shifts, bullpen specialization, fewer balls in play. The internal victory conditions of the sport were optimized at the expense of the broader ecology of the game. Then the league changed the rules: pitch clock, shift restrictions, larger bases, pickoff limits. Baseball had to defend baseball from baseball optimization.

That is the crucial lesson. Dilettante intelligence is dangerous not because it is always wrong, but because it is often locally right and globally stupid. Sports analytics saw baseball as a win-maximization problem. But baseball is also a media product, a ritual, a tempo, an aesthetic, a labor market, a developmental pipeline, a stadium experience, and a childhood inheritance. The measurable surface was real. It just was not the whole object.

The same pattern animated asset-light marketplace S​aaS. Own no cars. Employ no drivers. Cook no food. Run no hotels. Publish no journalism. Practice no medicine. Teach no students. Hold no inventory. Bear no institutional burden. Just sit above the practice, observe the signals, intermediate the transaction, and capture the margin.

Again, the critique is not that these companies found nothing. Uber was not wrong that taxis were awful. Airbnb was not wrong that hotel supply was constrained and trust could be mediated. Door​Dash was not wrong that restaurants had latent delivery demand. The outsider abstraction often found real friction.

But after the arbitrage was harvested, the abstraction began eating the ecology. Drivers became disposable. Restaurants became packaging nodes. Neighborhoods became yield-optimized hotel inventory. Journalism became traffic arbitrage. Dating became slot-machine optimization. The platform saw the transaction. It did not accept custody of the practice.

Consulting is the prestige version of the same structure: part real, part racket. Consultants can bring pattern recognition, benchmarking, analytical discipline, and outside permission to make hard decisions. Sometimes the consultant is useful precisely because insiders are trapped in politics.

But much of what is purchased is not the idea. It is legitimacy. A Mc​Kinsey deck is a ritual object. It lets management say: this decision is not our politics, our cowardice, our ambition, our layoff plan, or our failure of imagination. It has passed through neutral expertise.

The consultant parachutes in, interviews the people who already understand the business, converts situated knowledge into portable abstraction, and sells it back to leadership as external authority. The racket is not that the deck contains no truth. The racket is that the truth’s authority comes from being detached from the people who actually know it.

AI sycophancy is consulting’s recursion at the level of the individual user. The model does not only answer. It legitimizes. It converts impulse into strategy. It formats desire as analysis. It tells the founder, manager, lawyer, amateur scientist, or product person that the half-formed instinct is sophisticated.

Consulting launders management desire through a prestigious firm. AI sycophancy launders user desire through synthetic intelligence. In the old world, a CEO bought legitimacy from consultants. In the AI world, every golden boy has a pocket Mc​Kinsey that never says the client is the problem.

AGI culture is ZIRP after death. ZIRP promised that enough capital could bend markets into inevitability. AGI promises that enough intelligence can bend institutions into obsolescence. This is why the rhetoric feels religious. The AGI believer does not merely say the model is useful. He says it is the coming universal solvent. It will dissolve app stores, law firms, drug companies, schools, software teams, governments, media, and expertise itself.

Every institution becomes a temporary wrapper around insufficient intelligence. Once intelligence is abundant, the wrappers fall away.

But institutions are not merely low-IQ wrappers around tasks. They are memory, trust, standards, liability, distribution, taste, training, adversarial pressure, coordination, and consequence. They are how intelligence becomes durable. ZIRP said: with enough capital, we can skip the institution. AGI says: with enough intelligence, we can skip the institution. Both are wrong for the same reason. The institution is not the obstacle. The institution is where the value is stored.

This is why the AI-native phone/app-store fantasy is so revealing. The pitch says that AI weakens app-store lock-in because software can be generated, ported, and personalized. But that misidentifies the moat. Phones are not merely app containers. They are identity, payments, contacts, camera, messages, social graph, work, family, cloud, security, defaults, carrier relationships, and habit. A vibecoded app store does not disrupt Apple. It may strengthen Apple because infinite generated software increases the premium on trust.

Facebook Phone and Amazon Fire Phone failed because they were capture plays: solutions searching for a problem, attempts to turn the phone into a data and control surface for the sponsor. Android succeeded because it solved a real ecosystem problem for OEMs, carriers, and developers: an open iOS-like platform for everyone who was not Apple. The AI-phone founder thinks the moat is app code. The user knows the moat is life.

Anthropic is the enterprise version of the same overreach. Its mistake is not merely being closed source or expensive. The deeper mistake is singleton AGI culture: the belief that frontier intelligence entitles the lab to climb into every vertical. Claude Code. Claude Design. Claude Legal. Claude Financial. Claude Security. Claude Science. Claude for Life Sciences.

This is the ZIRP platform dream translated into AGI theology: own the magic layer, watch where users create value, then absorb the application. The model lab starts as supplier, becomes platform, then becomes competitor.

That is why the enterprise fear is not just privacy. Privacy says: do not look at my secrets. Sovereignty says: do not use my secrets to become me. The danger is not simply that a vendor sees sensitive data. The danger is that the vendor metabolizes operational alpha into a competing institution.

Claude Science and Claude for Life Sciences make the stakes obvious. Anthropic publicly describes Claude for Life Sciences as supporting the whole process from early discovery through translation and commercialization. Claude Science is not merely a chatbot for scientists. It is a workbench: tools, connectors, scientific databases, compute, auditable artifacts, lab infrastructure, workflows, proprietary context, and agentic research loops.

For a small biotech with little capital, weak infrastructure, and limited IP, this bargain can make sense. Claude helps the weak actor catch up. Frontier intelligence is most valuable to the actor who lacks a strong internal loop.

For Pfizer, Biogen, Roche, Novartis, Amgen, Sanofi, and their peers, the calculation is different. Their core asset is not access to a smarter model. Their core asset is institutional memory: assays, data, failures, molecules, wet-lab process, regulatory history, clinical operations, scientists, capital, and distribution. For them, letting Anthropic inhabit the scientific loop is not just buying productivity. It is transferring strategic position.

Nvidia is the contrast. Nvidia sells substrate: GPUs, Bio​Ne​Mo, Clara, NIM microservices, accelerated pipelines, open tools, and infrastructure. It arms the lab. Anthropic wants to inhabit it.

This is also why the marginal value of frontier intelligence is highest for the user most dependent on the model as a substitute for competence. The amateur experiences a better one-shot model as a breakthrough because the model carries the whole cognitive load. The professional experiences the model as one tool inside architecture, judgment, prompting, verification, domain memory, iteration, taste, standards, and consequence.

A 5 or 10 percent better model can matter. But for a serious user, it is not the institution. Once open or owned models are good enough to participate in the professional loop, the frontier premium compresses. The amateur consumes intelligence as a finished answer. The professional uses intelligence as one component inside a governed process. The less you know, the more the model’s one-shot capability matters. The more you know, the more the loop matters.

Legal is the cleanest case because the product is not generic intelligence. It is differentiated judgment expressed in voice. A mediocre legal user wants AI that sounds like a lawyer. A successful lawyer wants AI that sounds like the lawyer.

Elite legal writing is not merely correctness. It is strategy under constraint: rhythm, aggression, omission, sequencing, forum awareness, factual framing, citation density, client posture, judge knowledge, and litigation theory. It is the compression of adversarial consequence into prose.

A proprietary model can draft “a motion.” It cannot, by default, draft this partner’s motion, for this client, in this jurisdiction, before this judge, inside this firm’s theory, shaped by prior wins and scars. To do that, it must be entangled with the firm’s work product. And that entanglement is the firm’s capital.

The model can be rented. The voice cannot be.

This is where Anthropic’s safety mythology and enterprise sales motion converge into something like a protection racket. The lab says its models are powerful enough to threaten cybersecurity, science, labor markets, and national security. Then it sells access to those same capabilities as necessary protection.

If models are cyber weapons, you need Claude Security. If models will autonomously discover drugs, you need Claude Science. If models will transform law, you need Claude Legal. If models will eat software, you need Claude Code. If the future is exponential, you cannot afford to be outside the frontier.

The pitch becomes: the world is dangerous because of capabilities like ours; you will fall behind unless you buy capabilities from us; you should trust us with the data and workflows that make you valuable. Our models are so powerful that the world is dangerous without us, but safe enough that you should put your future applications inside our platform.

The protection racket can work only if the targets accept the model lab’s theory of the game. But this is where the dilettante misses the higher-order action.

The infinite dilettante thinks winning means optimizing the visible objective. In baseball, that meant maximizing win probability until the game became less watchable and the league changed the rules. The optimizer saw the inefficiency. He missed the veto players.

In business, the same pattern applies. The ZIRP founder thinks disruption is a one-way story: hapless incumbents, bloated bureaucracy, weak software, bad UX, old distribution, inevitable platform replacement. Sometimes that was true enough. But AGI verticalization is not attacking sleepy local incumbents. It is threatening dominant big tech players, cloud platforms, S​aaS decacorns, pharma giants, law firms, defense contractors, state agencies, and enterprise customers with procurement power.

These are not passive victims. They have hidden leverage: distribution, compute, data, workflows, trust, customers, legal regimes, procurement channels, regulatory influence, integration surfaces, and balance sheets. The higher-order game begins when the losers of your optimization have enough power to change the rules.

The Nvidia-Palantir sovereign-AI move is the clearest sign that the counterattack has begun. The model labs thought the model was the means of production. Alex Karp’s argument is that the weights trained on your operations are the means of production. The operational graph, the data, the authorization layer, the audit trail, the deployment environment, the feedback loop, the institution’s private telemetry — that is the capital.

Nvidia did not avoid the model layer because it was weak. It avoided the model layer because a plural model ecosystem sold more compute. Nvidia wins when models are many, hungry, customizable, and swappable. If a frontier lab tries to become the single sovereign intelligence layer, Nvidia has every incentive to arm the alternatives.

Palantir supplies the other half: operational graph, authorization, auditability, deployment into sensitive institutions, ontology, workflows, and customer trust. Nvidia arms the host organism. Palantir gives it a nervous system. Together they tell the state and enterprise customer: you can use AI without surrendering your means of production to the model lab.

Anthropic overplays its hand when it treats Claude as the institution. Its partners only need Claude to be a supplier. The model labs confused being supply-constrained with being sovereign.

The future does not belong to the singleton model hovering above every institution. The future belongs to institutions that turn models into owned operational capital. The restaurant owns the technique loop. The law firm owns the voice loop. The pharma company owns the discovery loop. The state owns the sovereign loop. The cloud and hardware layer keep models plural. The model can be rented. The frontier can be useful. It does not get to be sovereign.

ZIRP taught founders they could skip the institution with enough money. AGI taught them they could skip it with enough intelligence. Sports analytics showed the temptation: the outsider can be right, harvest the arbitrage, and still make the game worse. Consulting showed the racket: detached intelligence can launder power as expertise. AI sycophancy completes the loop, giving every golden boy a pocket Mc​Kinsey that never tells him he is the problem.

But the professionals, platforms, states, and infrastructure owners know something the infinite dilettante keeps missing: a field is not its visible information. A field is its consequences, and consequences create power. The institution is the capital. The model is a component. The model becomes swappable. The learning loop stays.