Related: sources · notes · metadata · Published Pieces

Deep Tech Before Lean Startup

Deep tech builds the substrate; lean startup tests the projections.

Lean startup is smart. Deep tech is respectable.

This distinction matters because people keep applying the wrong entrepreneurial grammar to the wrong kind of work. If the uncertainty is mostly market uncertainty, lean startup is a powerful discipline. Talk to users. Ship a small version. Measure behavior. Iterate. Avoid building monuments to your own taste. Let contact with the market correct your fantasy.

But not all companies begin with market uncertainty. Some begin with a capability frontier. Some begin with a technical thesis whose product form is not yet legible, because the substrate must exist before anyone can feel what it is for.

That is deep tech.

Lean startup is a sensor. Deep tech is an engine.

A sensor tells you what the environment is doing. An engine creates motion that did not previously exist. You need sensors. A vehicle without sensors crashes. But a vehicle made only of sensors goes nowhere.

The mistake is treating user feedback as the sovereign epistemology of company formation. Sometimes it is. Sometimes the customer already exists, the pain is legible, the workflow is known, and the market can tell you what to build. Sometimes the product category exists, but the incumbent version is bad, and a founder with taste can build a better object. And sometimes the relevant product cannot be discovered by asking users, because the technical possibility that would make the product desirable does not yet exist.

There are at least three primary entrepreneurial gradients.

Market-driven entrepreneurship begins with a buyer. Product-driven entrepreneurship begins with the object. Technology-driven entrepreneurship begins with a capability: a substrate emerging before the product has fully appeared.

All real companies are hybrids. Amazon began as market, product, and technology at once. Books were an ideal initial market for e-commerce because they had vast SKU variety, were easy to ship, and suited online cataloging. The product was an online bookstore. The technology was the internet. Later, the same technological and operational substrate produced AWS. That was not the original product. It emerged from the engine.

OpenAI was not lean startup in the ordinary sense. Anthropic was not lean startup. SpaceX was not lean startup. They may have used experiments, pilots, internal previews, and rapid product iteration, but the companies were founded on technical conviction before the market knew the final product. Large language models had to be scaled before ChatGPT could be recognized as the product. Rockets had to become reusable before Starlink could become obvious. The engine came first. The product emerged from it.

People often confuse late-stage product iteration with original company logic. After the substrate exists, a deep tech company can move with lean-startup speed. It can ship research previews, watch usage, discover workflows, iterate on interfaces, and rapidly productize capability. But that does not mean the company itself was market-driven. It means foundational capacity finally reached the point where product surfaces could be tested.

Choir is technology-originated and product-emerging.

The original thesis was not “build a better notes app,” “build a chatbot,” “build an agent platform,” “build a social network,” or “build an AI radio app.” The thesis was deeper: the world needs an ideal data engine. Web 2.0 built data flywheels by extracting the lower nature of users: outrage, envy, status anxiety, impulse, lust, tribalism, boredom, and shallow engagement. Those flywheels produced enormous training data, but the data was often socially degraded.

If a platform could instead extract the better angels of human nature — serious thought, correction, synthesis, memory, argument, teaching, taste, generosity, and disciplined disagreement — it could produce a better information ecosystem and better AI.

That is a technological thesis, but not only in the narrow engineering sense. It is social-scientific deep tech. The question is not merely how to scale servers or train a model. It is how to build a system where human expression becomes higher signal over time, public thought becomes durable, prior work gets cited automatically, contributions become assets, disagreement improves the artifact instead of just producing engagement, and the resulting discourse graph becomes a better substrate for intelligence.

The product form has been emerging from that substrate: the private automatic computer, the public automatic newspaper, and the audio layer of automatic radio.

I did not start with “automatic radio.” That would have produced something shallow: AI podcast generation, voice chat, NotebookLM with a better skin, synthetic hosts, or personalized audio slop. Automatic radio only became thinkable after the automatic computer and automatic newspaper existed as technical and conceptual substrate. Once there is an artifact graph, citation network, provenance system, vtext layer, background agents, and a private workspace, audio becomes more than generated speech. It becomes traversal.

This is why user feedback matters but cannot define the frontier.

A user can tell you they are confused, that onboarding failed, that the first minute does not land, or that the product feels invasive, boring, too abstract, too unfinished, or too hard to explain. That is valuable data. But users cannot always tell you what category the product belongs to before the category exists. They will map it to what they know: chatbot, feed, podcast, notes app, desktop, website, search engine, tutor, social network, agent builder.

The founder’s job in deep tech is not to obey users. It is to collide with them. Let the collision reveal friction, confusion, desire, and emotional contract. But do not surrender the ontology of the product to the current market’s vocabulary.

Lean startup says: do not build in a vacuum. Correct.

Deep tech says: do not let the vacuum of present demand define the possible. Also correct.

The synthesis is simple: deep tech builds the substrate; lean startup tests the projections. The substrate is the engine. The projections are the product surfaces users touch.

Sometimes the market is latent because the old order has not finished failing. The automatic newspaper does not yet feel necessary to most people because they have not fully experienced the AI information war. The automatic radio does not yet feel obvious because they have not hit the limits of voice chat, generated podcasts, and shallow personalized audio. The automatic computer does not yet feel obvious because most people have not run long agents for hours and felt the transcript collapse under real work.

But the incentives point forward.

Every institution will use AI to shape perception. Every public speaker will leave a machine-readable trail. Every serious knowledge worker will need systems that ingest, cite, compare, and preserve prior work. Every agentic workflow will need state outside chat. Every AI-mediated media system will need provenance.

This is the patience deep tech requires. Not slowness. Patience.

The goal is product liquidity: the ability to spin out new projections of the substrate rapidly — a new audio mode, vtext transform, public track-record view, work-radio workflow, appagent, citation surface. The engine makes product space cheap.

If you skip the engine and chase the nearest visible market, you may ship faster. But you will ship inside the old topology: another chatbot, summarizer, AI podcast generator, productivity wrapper, shopping assistant, or workflow automation tool. Useful perhaps. Profitable perhaps. But not foundational.

Technology-driven entrepreneurship is different because the product space is emergent. You build the machine, and the machine reveals what it can become.

First build the engine. Then test the projections.

That is deep tech before lean startup.