There is no shortage of material telling a company what it will look like once it has successfully transformed itself with AI. McKinsey’s account of the “agentic organization” is among the more careful examples — a genuinely useful sketch of an enterprise where agents hold persistent context, workflows reconfigure themselves around outcomes rather than org charts, and human judgment concentrates at the points that actually require it. Reading it, a company could reasonably conclude it now understands where it needs to go. What it would not come away with is any account of how to get there. This is not a gap specific to one report. It shows up just as clearly in the execution of these transformation efforts themselves: companies with the destination-vocabulary fully absorbed, running initiatives named after it, with no actual description anywhere of the sequence that would get them from where they are to where the vocabulary points.

A picture of the finished state is checkable against itself: internally coherent, citable, safe to publish. A description of the path is a different kind of document. It has to survive contact with a specific company’s actual decisions: which process to touch first (a claims queue, a vendor-onboarding chain, a pricing desk), what evidence would tell you the redesign is working before the metrics catch up, what to do when the first attempt fails in a way the destination-picture gave no warning of.

That kind of material is much harder to produce in general, because it mostly stops being useful the moment it becomes general: specific to one company’s sequence, not portable to the next one. The destination genre is popular partly because it’s the one that scales.

This produces a specific and fairly common failure mode, and it’s worth being precise about what it is not. It is not a capability problem: most large organizations now have access to models more than adequate for the workflows they’re trying to redesign. It is not, mainly, a budget problem either. What’s missing is the actual path: a described, structured account of which operational problem to rebuild around first, in what order, and what would count as evidence the rebuild is working rather than merely running. Not a principle for sequencing — the sequence itself, written down, specific enough to act on and specific enough to be wrong.

Most corporate AI transformation efforts underway right now are running the shaft-and-belt version: installing agentic capability into workflows shaped by an organizational structure the capability was never designed to serve — approval chains, reporting layers, decision rights — and narrating the result in the destination-vocabulary the industry has supplied, without having actually rebuilt anything that vocabulary presupposes. Paul David’s account of the [electric dynamo gives this failure mode its most precise historical form. Factories electrified their power supply and got almost nothing for two generations, because they kept the shaft-and-belt architecture and only swapped the engine that turned it. The productivity gain from electrification arrived thirty to forty years later, once factories were redesigned around individually driven machines rather than a single central power source distributing motion through a mechanical network built for steam. The technology was ready decades before the organization was.

A story tells you what the finished enterprise will look like. A playbook tells you what to build first, why that and not something else, and how you’d know three months in whether you’d chosen correctly or merely activated something. Most of what currently passes for AI transformation strategy is a story with a project timeline stapled to the back. It has the shape of a plan. It answers when, not which, and never touches why this operational problem before that one, which is the only question a real path has to answer.

The corporations that get the electrification payoff early will be the ones that treat this as the actual design problem, rather than a rollout problem to be solved with change management and a longer runway. The AI-transformed enterprise can already be imagined. It has not yet been built one playbook page at a time.

Paul Morphy retired from chess at twenty-two, undefeated. He spent the rest of his life pacing New Orleans in a cape, convinced strangers were plotting against his family’s estate. But before the breakdown, there was a game. Paris, 1858, an opera box. Morphy sacrificed pieces with a depth of calculation nobody at the table could match, because nobody had to hurry. There was no clock yet. Genius meant sitting with a position as long as it demanded.

Twenty-five years later, a metal box with a plunger showed up at the table — nothing to do with chess theory, just a hall that had to close, a tournament that had to end. But depth of calculation quietly stopped being the thing that won. Nobody voted on this. The board was the same sixty-four squares. Underneath that unchanged surface, the virtue the game rewarded had become a different virtue, and the players still training pure depth were playing the right game on the wrong clock.

Something close to this has been happening to venture capital, and almost nobody has noticed the box on the table. For a long stretch, innovation and software were close enough to the same word that nobody bothered separating them. The instruments VC uses to fund and time a bet were shaped by that fact with real precision: near-zero marginal cost, iteration in sprints, exits fast enough to fit inside a fund’s own decade. DPI, TVPI, the ten-year clock. Not arbitrary — fitted, the way Morphy’s unlimited depth was fitted to a game with no clock at all.

Then innovation did what any powerful idea eventually does. It stopped staying where it started. It leaked into agriculture, energy, logistics, regulated medicine, physical infrastructure: domains with real marginal cost, iteration measured in seasons rather than sprints, value that surfaces on a timeline no fund’s patience was built to survive. The most dangerous assumption is the one that used to be true, because nobody announces the moment it stops being true. The math keeps computing, correctly, on a game that has already changed underneath it.

Look at where capital is actually flowing now: infrastructure, defense, robotics, drones. Almost every load-bearing assumption software’s instruments were built on comes apart on contact. Capital structure needs hard assets and working capital, not a ten-year clock built for code. Regulators aren’t a compliance afterthought — they’re a primary actor in the business’s own value creation, sometimes as much a counterparty as a customer. Cycles run on procurement timelines and physical deployment, not sprints and demo days. None of this makes these domains worse bets than software once was. It makes them a different game, wearing the same box on the table.

The operational answer isn’t a repaired VC. It’s a recognition that the underlying need — superior returns from high-risk, high-reward ventures — never went away. Only the conditions attached to meeting it did. Venture building, done systematically, is what that need looks like once it’s re-derived from the fundamentals of these domains rather than borrowed from software’s: closer origination, longer patience, capital structures suited to hard assets and regulatory proximity, governance that stays active between checks instead of going quiet. Not a rejection of venture capital’s purpose. A reconstruction of its instruments for a game that quietly stopped belonging to software alone.

That reconstruction isn’t theoretical. It’s being built and tested right now, in scattered and imperfect form, by people willing to throw out DPI and start from what the domain actually rewards. One example among a small but growing number of first attempts at a science built for the game that’s actually being played, not the one that used to be.

The single number that matters most in venture capital has been falling for a decade. DPI, the cash a fund has actually paid back to its investors, is the part of a fund’s report that can’t be shaped by anyone’s optimism. It either arrived or it didn’t. Five-year DPI for the 2019 and 2020 vintages sits at 0.11 and 0.14, meaning eleven to fourteen cents back in cash for every dollar invested, after half a decade. That is not a bad year correcting itself. It is the continuation of a ten-year slide, and the slide has not found a floor.

The same erosion shows up beyond fund-level reporting. Distributions to LPs, which once ran close to a quarter of a fund’s value every year, have fallen to 16%, more than a one-third decline since 2022. At the company level, 857 active unicorns now carry a combined paper value above $4.3 trillion, more than triple the count from five years ago, while 40.6% of them haven’t raised a round in three years, meaning nobody outside the company has set a fresh price on them in that time. None of this is one bad quarter. It is a decade-long trend, and trends like this do not resolve themselves through patience.

The bad-bad news: the bad results will likely be even worse

A falling DPI on its own could be a bad year. The way to tell the difference is to look at what the other number is doing. If the decline were a temporary dip, you’d expect TVPI, the paper estimate of a fund’s total value including everything unrealized, to be correcting downward alongside it, GPs walking back optimistic marks as the cash failed to show up. That isn’t happening. TVPI has held in a healthy band, often above 1.5x, across the same decade that DPI fell apart underneath it.

That’s the part that turns a bad number into a structural one. Funds do have governance around marks, valuation policies, auditors, LPAC review, and none of it is fake. But none of it requires TVPI to reconcile with DPI on any particular timeline. A valuation policy can confirm a mark is reasonable without ever asking whether it should have converted to cash by now. No GP is obligated to write down a mark just because cash hasn’t arrived. No LP is positioned to independently verify the number or has a strong incentive to challenge it, since a markdown this quarter only makes their own reporting look worse. The two numbers are free to drift apart for as long as nobody with both the standing and the motive steps in to force them back together, and right now, the governance that exists doesn’t do that.

A structural gap behaves differently from a temporary one. It doesn’t sit still and it doesn’t average out. Every additional year a fund carries an unrealized mark without anyone testing it is another year of fees, opportunity cost, and quiet erosion stacked underneath an estimate that was never re-checked. The eventual payout, whenever the mark finally does meet cash, is smaller than it would have been had the correction come earlier. That is why this gets worse with time rather than better. There is no mechanism in the current architecture that pulls it back.

The cost of avoidance is quantifiable: a lot

You can watch the underlying mark get tested whenever a private company is finally forced to face a public buyer. Meesho, the Indian e-commerce platform, had been discussed for an IPO in the $7-8 billion range. It priced at roughly $5.6 billion. The gap wasn’t bad luck. Public-market analysts insisted on separating real operating profit from one-off accounting adjustments, a distinction the private mark had never been forced to make. The number moved the moment someone outside the company was finally allowed to check it. This isn’t a scandal. It’s the routine result of scrutiny arriving late.

The solution? DPI as a health check

None of this is unique to venture. Private equity carries the same exposure wherever a GP can mark an asset and also control the timeline for testing that mark, which is why continuation vehicles and similar structures draw the same scrutiny they do. A corporate innovation unit runs a milder version of the identical failure, an internal pilot that’s never been checked against a real kill criterion is an unrealized mark too, just without the paperwork. The fix for a venture model isn’t a better story about exits. It’s building the distribution mechanism into the thesis from day one, not as something to figure out once the asset matures, but as a structural feature decided at entry: a named buyer, a milestone-based path to cash, and a mark that’s never asked to stand unrealized for years because nothing ever required it to be re-tested.

There’s a deeper version of this for anyone who wants to take it further. DPI, TVPI, and the ten-year fund clock were built for one specific kind of venture, late-1990s software, where near-zero marginal cost and fast iteration made a long unrealized mark a reasonable bet. Apply that same apparatus to ventures with different underlying economics, and a widening TVPI/DPI gap isn’t the market behaving badly. It’s a measurement tool built for a different machine, mistakenly attached to this one.

That is what “PE-grade” should mean for a demand-first venture builder, a baked-in criterion, an integral part of the venture builder’s structure and operational thesis, where the distribution question was never deferred long enough for the gap to open in the first place.

Every serious methodology for building ventures rests on a founding assumption about where value comes from. Get that assumption right and everything downstream, the process, the capital structure, the team design, works with the grain of reality. Get it wrong and no amount of execution discipline recovers the loss. The New Science of Venture Building is built around four operative commitments, each of which addresses a specific point where the founding assumptions of the existing playbook have stopped holding. This post is about the first and most foundational of them.

If you are encountering the framework for the first time, the About post sets out what it is and why it exists. The short version: it is a systematic playbook for building ventures that are more likely to succeed, more capital-efficient, and more profitable, designed for the techno-economic conditions that actually exist today rather than the ones that existed thirty years ago.

The first commitment is called Demand-First Design, and it begins with a reordering of the question that venture building asks at inception. The existing playbook, the one that still governs most of what happens in VC-backed startups and corporate innovation centres, asks: what can we build, and who might want it? The new one asks: who needs a transformative solution, and what would it take to build it for them? The sequence looks like a small adjustment. It is not. Everything else in the architecture follows from which question you start with.

The idea to rethink existing venture development principles arose out of necessity. The dominant model served its era well. The first generation of digital ventures was built in conditions that were structurally forgiving: technology was genuinely scarce, building was hard, and first-mover technical advantage translated reliably into market position. In that environment, assembling a capable team around a promising capability and then finding its market made sense. The methodology fit the reality.

That reality is gone. The cost of building has collapsed. Capable teams are abundant. Technological advantage, in most domains, has a shorter shelf life than the ventures designed around it. In this environment the possibility-first model does not just underperform. It systematically misdirects effort, capital, and time toward the wrong question.

The Demand-First Principle is the architectural response to this shift. Its formulation is simple: the success of new ventures is no longer primarily determined by what can be built, but by what needs to be solved. The winning narrative is no longer a new capability looking for a market. It is a named buyer looking for a transformative solution. Everything that follows in venture design, from strategy and capital structure through team composition, milestones, and the conditions under which a venture is wound down rather than extended, flows from this single inversion.

The operative word is named. Not a market segment. Not a persona. Not an assumed future user. A specific organisation, with a specific operational problem, whose willingness to pay has been established before capital is deployed. The demand signal is real, not inferred.

This is not a customer discovery methodology. Customer discovery, in the Lean Startup tradition, is a process applied after a venture thesis exists, a way of testing and refining assumptions that were formed on possibility grounds. Demand-First Design operates upstream of that entirely. The venture thesis is constructed from the demand signal, not adjusted against it. The buyer is the origin point, not the validation checkpoint.

The implications cascade through the entire venture architecture. Financial modelling becomes more honest: you are projecting against a known willingness to pay, not an assumed market size. De-risking becomes structural, because the named buyer relationship is itself the primary risk mitigation mechanism, not a milestone to be achieved after funding. Exit architecture becomes designable at inception, because you know who the likely acquirer is. And the conditions under which a venture should be terminated rather than sustained on optimism can be defined at the outset rather than negotiated in distress.

That last point is underappreciated. One of the persistent failure modes of possibility-first ventures is the inability to distinguish between a pivot and a rationalisation, between a genuine strategic adjustment and a team unwilling to acknowledge that the original thesis was wrong. When demand is named and specific, this distinction becomes considerably cleaner. Either the named buyer is moving toward a transaction, or they are not. The signal is less ambiguous.

Demand-First Design is the first and deepest of the framework’s operative commitments. The others, Ecosystem Architecture, Portfolio Strategy, and Falsifiability, are downstream of it. Get the demand question wrong and the rest of the architecture, however sophisticated, is optimising the wrong system.

The next post in this series takes up Ecosystem Architecture: how the relationships between corporations, capital, governments, and academia are designed as functional assets before execution begins, and why a venture builder that leaves this to chance is competing with one hand tied.

Marc Rowan did not set out to identify a structural flaw in private capital. He was talking about retirement income. David Haber was referencing a twelve-year-old essay.

Somewhere in between, they stumbled into the right problem. Both could see that something new is emerging. Neither could name it.

Haber’s observation was clean. The businesses AI is producing now - defense, energy, robotics, manufacturing: are capital-intensive in ways software never was. They will outgrow venture equity. They will need Apollo. His question to Rowan was direct: what do these companies need to do to become investable by you?

Rowan’s answer began with a confession. Apollo is not capital-constrained. It is idea-constrained. “We can only invest as fast as we originate, as fast as we create.” A trillion dollars in AUM. The binding limit is not money. It is worthy investments. Auctions don’t solve this — you arrive late, pay market price, remain a participant. “I want to be a principal,” he said. “I want to own the upside for as much of the asset as the market will allow.”

And then, almost in passing: “I think really good entrepreneurs are going to end up in partnership with entrepreneurs of another type — those who are financial entrepreneurs.”

He could see the shape. He could not see the form.

What followed was instructive. Pressed for solutions, both men reached for the tools they already hold. It was not enough.

Apollo opens a second headquarters near the growth ecosystem. Earlier engagement with founders. Better alignment between venture and credit. a16z helps portfolio companies think about capital structure sooner. Tighter pipelines. Smarter screening.

Real moves. None of them the answer.

This is the blind men and the elephant. Rowan touches the trunk and describes a rope. Haber touches the leg and describes a tree. Both descriptions are accurate. Neither describes the animal.

Apollo is a financing and acquiring machine. It cannot build. It has no mechanism for originating ventures from inception: for installing the governance, the demand architecture, the exit design that makes a venture PE-investable before it becomes a company.

a16z is a funding and advising machine. It cannot transform risk profiles. A Series B defense-tech company does not become an Apollo asset because a16z introduced the founder to Rowan. The asset class is wrong. The governance is wrong. The exit architecture was never defined.

Both firms can see the destination. What’s missing is the way to get there.

This is not a criticism. It is a description of how worldviews work. You solve problems with the tools you have. When the problem requires a tool you don’t possess - one that doesn’t fit your existing architecture - you make the existing tools work harder. You call it progress.

Rowan and Haber are not slow. They are, in the precise sense, bound by what they can see from where they stand.

What they missed is not a feature. It is not a process improvement or a tighter feedback loop between Sand Hill Road and Midtown Manhattan. What they missed is a model.

The answer to Haber’s question — what do these companies need to do to become Apollo-investable — is not found in better staging or earlier introductions. It is found in rethinking what a venture is, and how it is built. Disruptive innovation, treated as a strategically designed and grown asset from inception rather than a bet that matures into legibility, changes the logic entirely. Demand confirmed before capital is committed. Exit architecture defined at structuring. Failure conditions priced in from day one. The result is not a better venture. It is a different asset class: one that carries venture-scale return potential at a risk profile PE can actually underwrite. That is not an incremental improvement on existing models. It redraws the playfield. It defines the next wave of long-term, strategic innovation, not as an accident of survival, but as a designed outcome.

That methodology exists.

Rowan described the problem it solves. Haber described the market that needs it. Neither could see it from where they stand.

The most interesting thing about their conversation is not what they said. It is what they almost said.

They were one step away. The gap between Apollo and a16z has a name. The partnership between financial entrepreneurs and venture entrepreneurs has an institutional form. The asset class they are both circling has a methodology behind it.

They stumbled into the question. Someone else will have to answer it.

A New Science of Venture Building is a demand-first venture building methodology designed for exactly the gap described in this piece.

A moat is a competitive advantage built into what you make. A wall is a barrier erected around it. These are not the same thing, and in the current AI moment, a great deal of money is being spent on walls while calling them moats.

Here is what is actually happening. The U.S. AI capital stack was financed on a specific thesis: that the companies producing the most capable models would capture durable advantage, because the performance gap between the best closed systems and everything else would widen over time. That thesis has a falsification condition — and it is arriving. The gap between closed and open-weight models is not widening. It is closing, at a pace that several serious observers now put at six to twelve months. DeepSeek demonstrated that a research team operating at a fraction of the assumed budget could produce a model competitive with frontier closed systems. This is not a surprise to anyone paying attention to the structure of the field. It is an anomaly the thesis was not designed to survive.

The response has a name, though it rarely goes by one: regulatory enclosure. Restrictions on chip exports, data localisation requirements, national security frameworks applied to AI infrastructure — each of these is being advanced, with varying degrees of sincerity, as a security measure. And some of them may well be. But the structural effect, regardless of intent, is to recreate through regulation the scarcity that technology is no longer providing naturally. When the model weights are approaching commodity status, the instinct is to build a wall where the moat used to be.

The great and prematurely dead philosopher of science, Imre Lakatos, drew a distinction between what a theory actually claims — what he called “the hard core” — and the apparatus of adjustments and exceptions constructed to protect it from inconvenient evidence. A theory in good health updates itself when the evidence pushes back. A theory in trouble keeps adjusting the apparatus while leaving the core untouched, and hopes nobody notices. The AI exclusion thesis is exhibiting this pattern. It has stopped producing predictions about what the technology will do. It has started producing predictions about how the regulatory environment will be arranged to make the technology’s current limitations irrelevant. That is a different kind of claim, and a considerably weaker one.

The auto industry spent decades behind tariff walls and produced cars that required those walls to survive. When the walls came down, partially and fitfully, the industry discovered that protection had not built resilience. It had purchased complacency. The companies that emerged strongest were not the ones behind the highest walls. They were the ones that had, in the meantime, built something that did not need the wall to function.

A genuine moat in an environment where model weights approach commodity status looks like this: two years inside a specific enterprise’s operational reality, learning what their problems actually are at the precision that only sustained contact produces. The outcome signal you get when you can observe whether the recommendation actually moved the metric. Relationships and accumulated context that do not replicate when a competitor downloads the weights and do not erode when the next open-weight release closes the capability gap by another notch. These compound because they are anchored to something that does not generalise: the named problem of a specific enterprise, in a specific context, at a specific moment.

The wall is not the moat. The question worth asking right now is not how to make it higher — it is what you are building inside it that will still be standing when it comes down.

Robert Louis Stevenson hated it. In 1878, before the grid existed, before the dynamo was a commercial proposition, before anyone had thought seriously about what electricity might actually be for, he wrote of the arc lamp as “a horrible, obnoxious, and hurtful thing” - “a lamp for a nightmare.” The light was real. The disgust was real. The conclusion was wrong.

Reid Hoffman’s recent defence of AI slop has attracted the expected response: disgust, mostly, from people who are right about the slop and wrong about what it means.

Hoffman’s argument runs as follows. The Heinz pickle - the Niblo’s Garden chorus girls wired with electric bulbs, the uranium glass butter dishes glowing in arc-lit windows, the novelty illuminations that struck the serious-minded as the reduction of a world-historical technology to a fairground trick, was not a diversion from electrification’s transformative potential. It was the mechanism. It built the customer base, the investor confidence, and critically the physical infrastructure, like the central stations, the cables, the dynamos, that made the X-ray machine possible and the affordable loaf of bread a consequence of cheaper cold storage. The critics were right about the pickle. They were wrong about what the pickle was doing.

This is a good argument. It is also incomplete. And the gap between what Hoffman says and what is actually happening is where the interesting question lives.

What Hoffman Gets Right

The slop critics are making a category error. They are evaluating the Heinz pickle against the X-ray and concluding that the pickle is a failure. But the pickle was never competing with the X-ray. It was paying for the grid.

This is not a trivial observation. The history of general-purpose technologies is littered with the same pattern: an initial wave of applications that look, to contemporaries, like trivialisation, the first uses of printing were indulgences and almanacs, not scientific treatises; the first uses of the internet were pornography and cat videos, not distributed knowledge infrastructure. The critics who focus on the triviality of the first wave are not wrong about the triviality. They are wrong about the mechanism.

AI slop, and everything that it consists of, like the marketing copy, the synthetic imagery, the chatbot scripts, is doing what the Heinz pickle did. It is building the customer base. It is generating the revenue streams that fund the infrastructure. It is training the practitioners. And it is doing something the critics systematically underestimate: it is generating data about what the technology can and cannot do, at scale, in deployment conditions that no laboratory can replicate.

Hoffman is right to defend it. The serious-minded disgust is aesthetically understandable and strategically irrelevant.

What Hoffman Leaves Unnamed

But the pickle/X-ray binary is too clean. Hoffman presents electrification as a two-act story: slop, then transformation. The actual history is messier and more instructive.

The Bijou Theatre in Boston (the first commercial incandescent lighting installation in America, 1882) was not just load for the grid. It was a deployment cycle. It taught Edison’s team something about demand patterns: when people arrived, how long they stayed, what failure modes looked like under sustained use, how operators behaved when something went wrong at 8pm with a full house. The Heinz pickle factory was not just revenue. It was a laboratory for understanding what industrial electricity actually required at scale, as opposed to what the engineers thought it would require in the abstract.

Each deployment cycle either confirmed the direction or forced a review. Were the failures a consequence of poor execution, or was the underlying theory of what electricity could do in this context simply wrong? That bidirectional question — not “did it work?” but “if it didn’t, was it us or was it the idea?” — is the mechanism Hoffman leaves unnamed. It is what distinguishes a progressive technology deployment from an expensive dead end.

Paul David, the economic historian, established in 1990 what the contemporary AI discourse keeps rediscovering: electrification’s productivity gains arrived thirty to forty years after the grid was built. Not because the technology was slow, but because the gains required organisational redesign. The factories that adopted electricity in the 1880s bolted electric motors onto their existing shaft-and-belt architectures. They got modest efficiency improvements and, occasionally, the ability to run a second shift. The factories that got the transformative gains were the ones designed from scratch around what electricity made possible: unit drive, the end of the shaft, the reorganisation of the factory floor around workflow rather than power transmission geometry.

The X-ray did not arrive after the pickle. It arrived after many successive cycles of pickle-level deployment progressively refined what practitioners understood about what the technology could do and under what conditions. The reorganisation wave: the very moment when organisations are redesigned around what the technology makes possible rather than bolted onto existing architectures, does not follow the slop phase as a consequence. It is what the iterated approximation converges toward.

This is the argument Hoffman is reaching for and does not quite make. The slop is not just funding the grid. The slop is the iterative process by which the grid becomes legible: to operators, to customers, to the engineers who will design the X-ray, as something more than a novelty.

What This Means for Venture Design

The contemporary AI discourse has produced two camps that talk past each other.

First, there are the slop critics. Freddie deBoer’s “bits are easy, atoms are hard” is the most structurally rigorous version. They point at the triviality of current AI applications and conclude that the transformative potential is overstated.

Then there are the slop defenders. They point at the Heinz pickle and conclude that the critics are being historically naive. Both are right about what they see. Neither has a theory of the mechanism.

The mechanism is iterative deployment against named operational constraints, with a built-in falsification criterion at each cycle. The factories that got to unit drive first were not the ones that read the most authoritative descriptions of where electricity was going. They were the ones that redesigned around a specific named operational problem — this floor, this workflow, this bottleneck — and ran the bidirectional review at each step: were the actions unsatisfactory, or was the theory of the problem wrong?

This distinction matters for anyone designing ventures in the current environment. The question is not whether to bet on AI transformation. The historical evidence is unambiguous that general-purpose technologies transform, eventually, the organisational contexts into which they are deployed. The question is how to position for the reorganisation wave rather than the pickle phase.

A venture designed around a named operational problem in a named enterprise has a falsification criterion built in from day one: did the metric move after the recommendation? That question is what the Bijou Theatre gave Edison that the pure-play slop generator does not have. The deployment has a referent. The iteration has a direction. The cycle is progressive rather than merely repetitive.

A venture without a named buyer cannot ask that question cleanly. It is generating slop, which may fund the grid, and is probably necessary at this moment in the infrastructure build — but it is not running a progressive research programme. It is running a lottery. Some lotteries pay out. Most do not, and the ones that do not have no mechanism for distinguishing “the theory was wrong” from “the execution was unsatisfactory.”

The Reorganisation Wave Is Not a Phase

The deepest error in both the slop critique and the slop defence is the assumption that transformation arrives as a distinct phase: after the slop, or despite the slop, or because of the slop, but temporally separable from it.

Paul David’s insight is more radical than that. The reorganisation wave is not what comes after the pickle phase. It is what the iterated pickle converges toward, when the iterations are run against real constraints with real falsification criteria. The X-ray did not emerge from the grid fully formed. It emerged from the accumulated understanding of what electricity could and could not do, built up through successive cycles of deployment against named problems that rejected imprecision in real time.

The Heinz pickle was not the grid. But it was not merely the thing that paid for the grid, either. It was one node in the iterative process by which the grid became something other than an expensive lamp for a nightmare.

Stevenson was wrong. The arc lamp was not a thing for nightmares. It was the beginning of a research programme that nobody had yet named.

The critics of AI slop are making the same mistake. They are right about the slop. They are wrong about what the slop is doing, and they are wrong because they have no theory of the mechanism that connects the Heinz pickle to the X-ray.

The mechanism is iteration against named constraints. The ventures designed around that mechanism: named buyer, named problem, measurable outcome, bidirectional review at each cycle — are not competing with the slop. They are positioned for what the slop converges toward.

That is a different bet. It requires a different institutional structure, a different capital logic, and a different theory of what venture design is actually for.

But it is the bet that the history of general-purpose technologies suggests is correct.

Every successful venture has a clean win story. None of them are true.

This is not a complaint about dishonesty. The people telling the clean win stories mostly believe them. Memory is reconstructive, not archival. We remember the decisions that worked as decisions we made deliberately, for the reasons that turned out to be correct. We remember the pivots that saved the company as strategic insight rather than panic. We remember the hires that made the difference as inspired judgment rather than the fourth choice after the first three said no.

The mess is not edited out of the story. It is genuinely not remembered. Or rather: it is remembered as the texture of the journey, not as its structure. The structure, in retrospect, looks like a sequence of correct decisions leading inevitably to the outcome. The correct decisions were correct. The sequence was not inevitable.

The clean win story serves a function. It is legible to investors, to journalists, to the next generation of founders who are trying to learn from it. A story with genuine contingency at its centre, something like: “we nearly ran out of money in month fourteen and the thing that saved us was an introduction that almost didn’t happen”, is not a story that transmits useful information. It transmits luck, which is not actionable.

So the contingency gets edited. Not dishonestly, but structurally. The story needs a spine, and the spine is the set of decisions that, in retrospect, were correct. Everything else becomes texture.

The problem is that the spine teaches the wrong lesson. The lesson that survives the editing is: here are the decisions that worked. The lesson that gets lost is: here is the range of decisions that were made, most of which did not work, and here is what it felt like to operate in that range without knowing which ones would.

That felt experience — operating with uncertainty, making decisions without the information that would later make them look obvious, staying in the room long enough for the thing that almost didn’t happen to happen — is precisely what cannot be transmitted through the clean win story. And it is precisely what distinguishes someone who has built something from someone who has studied how building is done.

There is a test for this, though it is rarely applied. Ask someone not how they succeeded but where they were most wrong. Not in retrospect. Not “we initially thought the market was X but it turned out to be Y”, but in the moment. When did you not know what you were doing? When were you running out of road and did not have a plan?

The clean win story has no answer to this question. It has been edited out.

The person who has actually built something has an answer that is immediate and specific. The details are vivid because they were frightening. The month, the person, the number in the account, the conversation that was not going well. These do not get reconstructed into strategic insight. They are remembered as they were: difficult, uncertain, contingent.

This is not a counsel of despair about entrepreneurship. The ventures that work mostly work because the people running them were willing to stay in the difficulty longer than the people who left. Persistence is real, and it is learnable. But it is learnable only by people who understand that the clean win story is a retrospective construction, not a description of how ventures actually unfold.

Phil Rosenzweig’s Halo Effect established this structurally twenty years ago: when a company succeeds, its earlier decisions get reinterpreted as bold and visionary; when it fails, the same decisions get reinterpreted as reckless and naive. The decisions did not change. The outcome changed the story. The venture ecosystem has not absorbed this lesson. The 30-year record of startup methodology shows no improvement in survival rates, which is what you would expect if the primary vehicle for transmitting experience is the clean win story, which systematically edits out the information that would actually be useful.

The mess is the default. The clean win story is the exception, produced after the fact, for audiences who need legibility.

If you are trying to build something, the most useful thing you can know is that the people who succeeded were, at various points, as confused and as frightened as you are now. The clean win story does not tell you this. The scar tissue does.

Find the scars. The stories attached to them are the ones worth learning from.

Tiffany Luck is one of the more thoughtful vertical AI investors currently active. In a recent conversation, she described what she is waiting for before committing to the next frontier: the “Waymo moment”: a fully autonomous agentic workflow, hands off the keyboard, the technology doing the work without a human in the loop to catch its errors. When that moment arrives, she will know where to place the bet.

The question she did not ask is the interesting one: who will make it happen?

הון סיכון שונא סיכון

The phrase is a Hebrew pun that does not translate cleanly, which is itself part of the argument. הון סיכון means venture capital: literally, risk capital. שונא סיכון means risk-averse. The joke is that venture capital, the institutional form whose entire rationale is the assumption of risk in pursuit of outsized return, is constitutively risk-averse about the one kind of risk that matters: the risk of operating before the paradigm has a name.

This is not an accusation. It is a structural description.

The VC fiduciary logic requires legibility before commitment. To raise a fund, you need comps — companies that have succeeded in a recognisable way, in a recognisable category, against which your thesis can be evaluated. To deploy a fund, you need a demonstrated market — evidence that the thing you are funding has a precedent that validates the direction. To return a fund, you need an exit, namely a buyer or a public market that assigns a value to the thing you built, using a metric that already exists.

Every link in this chain requires that the paradigm has already crystallised. The market is demonstrated. The category has a name. The metric exists. The buyer understands what they are buying.

Tiffany Luck is waiting for the Waymo moment because the Waymo moment is the point at which the paradigm crystallises: when autonomous agentic workflow becomes a legible category with demonstrated economics and recognisable comps. At that point, the VC logic becomes applicable. Before that point, it does not.

The question of who paves the road to the Waymo moment is not a question the VC structure is designed to ask.

Who Paves the Road

The road-paving work has a specific character. It involves operating in the pre-legibility zone — deploying technology against named operational constraints before the category exists, before the comps are available, before the metric that will eventually measure success has been invented. The people who do this work do not know, when they begin, whether they are building the Waymo moment or one of the many failed attempts that will be forgotten when the moment finally arrives.

This is the actual risk. Not the risk of a known bet going wrong — that is the risk the VC model prices efficiently. The actual risk is operating without the information that would make the bet legible: without a category, without comps, without a metric, without certainty that the thing you are building is the thing that the paradigm, when it crystallises, will reward.

The first-wave digital ventures were largely spared this risk. They were building digital twins of established physical businesses, inheriting the demand signal, the comps, and the metric from the physical original. The VC logic was calibrated to this world: the paradigm was legible because the physical precedent was legible. The risk was execution risk, not paradigm risk.

The ventures that will enable the Waymo moment, and I’m talking about the ones that are currently deploying agentic workflow against named operational constraints in named enterprises, discovering through iterative deployment what the technology can and cannot do in conditions that no laboratory can replicate — are taking paradigm risk. They are paving the road whilst walking it.

The VC structure cannot fund this work. Not because VC investors are insufficiently courageous, but because the fiduciary logic does not permit it. You cannot raise a fund on “this will work when the paradigm shifts.” You raise a fund on comps.

The Capital Structure That Can

The road-paving work requires a capital structure that can replace the VC’s legibility requirement with something else. The VC needs a category. The road-paver needs a different anchor: a named enterprise with a named operational problem and a measurable outcome at each step.

The named enterprise does what the category cannot: it supplies a falsification criterion. Did the metric move after the recommendation? That question is what makes iterative deployment progressive rather than merely repetitive. The venture is not running a lottery, hoping to guess correctly. It is running a research programme, with a fixed referent and a bidirectional review at each cycle: were the actions unsatisfactory, or was the theory of the problem wrong?

This is why the demand-first venture builder is a PE play, not a VC play. Private equity can tolerate the ambiguity of road-paving because the named corporate relationship — and the outcome loop it enables — replaces the VC’s legibility requirement. The named buyer’s problem is the category. The measurable outcome is the comp. The iterative deployment is the track record.

The PE structure has a further advantage that is underappreciated. The VC model is designed for paradigm exploitation. It means entering a crystallised category at scale and pricing the risk of execution against a known template. The PE model, at its best, is designed for operational transformation: which means acquiring or building businesses with embedded operational reality and improving the economics through structural change. The road-paving work is closer to operational transformation than paradigm exploitation. The PE institutional logic is, in principle, better suited to it.

In principle. The current PE model has its own version of the legibility problem: it requires demonstrated financial performance, a known exit mechanism, and comps for the multiple. The innovation required is not a different capital structure in the abstract — it is a PE vehicle that accepts the named operational problem as the substitute for the demonstrated category, and the measurable outcome loop as the substitute for the historical financial track record.

What the Waymo Moment Actually Requires

The deeper problem with Tiffany Luck’s position (and she is not alone; it is the default position of the VC industry) is that it makes the Waymo moment a precondition for investment rather than a consequence of investment.

Waymo did not arrive by waiting. It arrived because a named enterprise — Google, then Alphabet — decided to fund the road-paving work before the category existed, against a named operational constraint (autonomous urban navigation), with the patience to iterate through the pre-legibility zone without requiring comps or a demonstrated market.

The road existed because someone walked it. The moment did not precede the walking. It was produced by it.

The ventures that are currently deploying agentic workflow against named operational constraints in named enterprises are doing the walking. They are not waiting for the paradigm to crystallise. They are the mechanism by which it crystallises, but for a named enterprise, in a named domain, with a measurable outcome at each step.

The VC industry will arrive when the paradigm is legible. It will price the entry as if it had taken the risk.

The road-paving work will have been done by someone else, on a different capital structure, with a different theory of what value creation in the pre-legibility zone actually requires.

הון סיכון שונא סיכון. The pun works because it is true.

The mythology of the garage entrepreneur is not wrong about the facts. It is wrong about what the facts explain.

Amazon did not need to discover that people wanted to buy books. The demand existed, was documented in thirty years of retail data, and was large enough to anchor a business before a line of code was written. The only question was whether the internet could serve it more efficiently than a physical bookstore. This turned out to be a question with a known answer, reachable through execution rather than discovery.

The same structural condition held across the first wave of digital venture creation. Every major venture of the 1990s and early 2000s was, at its core, the digital twin of an established physical business — inheriting its demand signal, its customer base, and its commercial logic, while stripping out the physical infrastructure and the associated cost. eBay was a flea market. Google was a card catalogue. Booking.com was a travel agent. The demand was not in question. The bottleneck was entirely technical: could you build the thing that would serve the existing demand at digital scale and digital cost?

This is what made the garage entrepreneur viable. The structural condition rewarded people who could build before they could buy. The skill set required was engineering skill, execution capacity, and enough capital to reach the point where the inherited demand signal validated the investment. The demand itself was a given.

The Inversion

The structural condition has inverted. Not gradually, but decisively, at a dateable point in the mid-2010s when the first-order digital transformation of established physical businesses was largely complete.

The bottleneck is no longer technical. The ability to build software — to take an idea and convert it into a functioning product — has moved from scarce to abundant to, with the current generation of AI tools, nearly free at the margin. The engineering skill that distinguished the garage entrepreneur from everyone else is no longer the constraint. What remains scarce, and is becoming more scarce as the technical barrier falls, is something quite different: operational context, institutional fluency, and the ability to identify a named buyer whose operational reality is the design input from day one.

The first-wave ventures inherited their demand signal from the physical world they were digitising. The ventures being built now are not digitising known physical businesses. They are attempting to reorganise operational reality, which is to design new workflows, new decision architectures, new capability structures, in domains where the demand is felt but unarticulated, where the problem is real but unnamed, and where the only way to validate a design is through sustained contact with someone who has to live with its consequences.

This is a fundamentally different design problem. It cannot be solved in a garage. It requires access to the operational reality of the domain being redesigned - which means access to the people, institutions, and processes that constitute that domain. And that access is not something that can be acquired through technical skill alone.

What Changed

The first-wave garage entrepreneur was operating in a world where the hardest problem was the technical one, and where the demand signal was legible because it was inherited from a physical precedent. The venture capital logic of that era was calibrated to this world: find someone who can build, fund the build, wait for the inherited demand to validate the investment.

The methodological apparatus that grew up around this logic (with some honourable mentions are: lean startup, customer development, the business model canvas) was a good-faith attempt to introduce demand validation into a process that often skipped it. The evidence suggests it did not work: thirty years of data show no improvement in venture survival rates despite widespread adoption of the methodology. This is not because the methodology was bad advice. It is because the structural condition it was designed for, that inlcluded inherited demand, technical bottleneck, legible market, was already dissolving as the methodology was being codified.

What the methodology could not supply was the thing that actually determined first-wave success: the pre-existing demand signal that the digital twin inherited from its physical original. Lean startup without an inherited demand signal is a sophisticated framework for discovering whether you have guessed correctly. It is not a mechanism for generating the operational context that makes the guess well-founded in the first place.

The New Bottleneck

The ventures that will define the next wave are not digital twins of established physical businesses. They are operational redesigns of domains where the demand is felt but unarticulated, where a named enterprise knows it has a problem, cannot specify the solution, and requires a design partner who understands the operational reality well enough to name the problem before proposing an answer.

This is what demand-first design means in practice. Not “ask customers what they want”. The Jobs critique of demand-led design is correct that customers cannot specify solutions to problems they have not yet named. Demand-first design begins from something more specific: a named enterprise, a named operational domain, and a sustained engagement with the people whose daily reality is the design input. The demand signal is not inherited from a physical precedent. It is excavated from operational reality through a process that requires institutional access, domain fluency, and the patience to stay in contact with the problem long enough for it to become legible.

The garage is the wrong setting for this work. The garage worked when the bottleneck was technical and the demand was given. When the bottleneck is operational context and the demand must be generated through institutional engagement, the garage is not just suboptimal — it is structurally incapable of producing the design input the venture requires.

This is not an argument against entrepreneurial ambition or individual agency. It is an argument about the structural conditions that make ventures viable. Those conditions have changed. The mythology has not.

The Bet That Is Now Correct

The first wave of digital venture creation was structurally forgiving because the demand was given and the bottleneck was technical. The structural generosity of that era produced the mythology of the garage entrepreneur: the individual with an idea, a laptop, and access to capital, building toward a demand signal that was already waiting.

That structural generosity is gone. The technical bottleneck has dissolved. The demand signal is no longer inherited. What remains is harder, less legible, and not amenable to the institutional structures. The VC fund, the accelerator cohort, the demo day that were built to serve the first-wave structural condition.

The bet that is now correct is different. It begins not with an idea but with a named operational problem in a named enterprise, excavated through sustained institutional engagement. It treats the design process as an iterative conversation between what the enterprise knows it needs and what the technology frontier has made possible. And it requires an institutional structure that can hold both vectors simultaneously — the named demand and the open possibility — over a time horizon that the first-wave VC model was not designed to support.

Amazon did not need to discover that people wanted to buy books. The next Amazon will need to discover, with a named partner, what they need - before the partner can name it themselves.

That is a different kind of entrepreneurship. It requires a different kind of institutional home.

The garage mythology was accurate for its moment. The moment has passed.

Almost everything written about venture building, innovation, and capital is produced after the fact. The celebrated frameworks, the canonical case studies, the playbooks that circulate as received wisdom — they were written by people looking backwards at what worked, reverse-engineering the logic from the outcome. Which would be useful, if the conditions that produced those outcomes still existed. They don’t. The structural reality of building and funding ventures has shifted fundamentally, and the retrospective playbook doesn’t just lag. It actively misleads.

I’ve spent more than twenty years at the intersection of entrepreneurship, venture building, investment, and innovation strategy: across geographies, sectors, and institutional contexts that rarely overlap in one career. That breadth produced a specific kind of pattern recognition: you see where the established frameworks contradict each other, where one domain has solved what another is still struggling with, and where the consensus is held together more by convention than by evidence.

pilot waves is an attempt to put that pattern recognition to work in public, to extract insights while they’re still forming, before the outcome is known and the story has been cleaned up. Not commentary, not a methodology document. Analysis and argument, written as the thing is happening, grounded in a framework that has been tested against reality rather than abstracted from it. A forward-looking playbook, built in the open.

That’s what this is. Everything else follows from it.

— Roman

The difference matters, but not because the existing methodologies were wrong. Lean Startup, Customer Development, Business Model Canvas, Design Thinking: these produced some of the most successful ventures of the last thirty years. They were the right tools for their moment. The problem is that the moment changed and the tools didn’t. The assumptions baked into them, about the scarcity of technology, the cost of building, the dynamics of capital, the structure of markets, stopped being true. Outcomes stalled. The tools proliferated anyway.

The framework is the result of operating seriously, in parallel, across domains that rarely overlap in one person: entrepreneur with exits, VC partner, designer of national innovation ecosystems, university lecturer. Each domain has its own received wisdom and its own blind spots. Working across all of them simultaneously makes the contradictions visible, and the contradictions are where the interesting questions live. For a fuller account of that background, the Who Am I post is the right place.

The result is a systematic approach to building ventures that are more likely to succeed, more capital-efficient, and more profitable by design rather than by luck. It operates through four commitments. Demand-First Design: the venture is built around a named buyer with a verified need, not a capability in search of a market. Ecosystem Architecture: the relationships between corporations, capital, governments, and academia are designed as functional assets, not left to emerge. Portfolio Strategy: positions are managed across time horizons and outcome types, with exit architecture defined at inception. Falsifiability: failure conditions are specified at the outset, not negotiated in distress.

Each of these is developed in the Framework Essentials series, published alongside the regular posts on this site. The analysis, arguments, and occasional provocations in the main feed are the framework operating on live material. The Essentials series is where the logic is laid out directly.

This framework is designed for now. The techno-economic conditions it addresses will keep shifting, and the commitments that make sense today will eventually need to be revised, updated, and at some point discarded in favour of whatever comes next. That is not a weakness in the argument. It is the argument.

I’m good at reframing. At finding the question underneath the question, and at making the structural logic of a situation visible to the people inside it who are too close to see it clearly. I enjoy the early stage of hard problems, before the answer exists, when the work is still figuring out what the right question is. I’m drawn to conversations where someone needs a new lens, not a familiar solution, and where the instinct to reach for the established playbook is precisely what needs to be examined.

In recent years I’ve worked with corporations building or rethinking their innovation function, private equity firms considering venture building as a strategic capability, startups at moments where the founding assumptions need honest stress-testing, and academic institutions looking to discuss innovation practice at the frontier rather than through the retrospective case study. The common thread is that the organisations involved were trying to do something they hadn’t done before, in territory that didn’t yet have a map.

If you’re building or running an innovation function inside a large organisation - or deciding whether you should - I’d be glad to discuss it. Same if you’re a PE firm or family office thinking about venture building as part of your strategy, a startup wrestling with a strategic question that doesn’t have an obvious answer, or an institution that wants to discuss what’s actually happening at the edge of innovation practice rather than what the consensus says is happening.

I’m also available to speak. The topics I find most worth arguing about in public are the ones that appear settled but aren’t: why thirty years of venture methodology have produced no improvement in survival rates, what the current AI reorganisation wave actually requires from the organisations living through it, and why the most important strategic moves in innovation are almost always invisible at the moment they’re made.

If any of this is relevant, drop me a line.

I’m an entrepreneur and strategist with a restless, unconventional mind, and, I’ve been told, an uncomfortable habit of asking the question the room was hoping to avoid. I build ventures, advise organisations at strategic inflection points, and think seriously about why most innovation fails on its own terms rather than for lack of effort or capital.

What makes my perspective unusual, and I say this descriptively rather than as a boast, is that I’ve operated across domains that rarely overlap in one person. I’ve been an entrepreneur with exits, a partner at venture capital firms, a designer of national innovation ecosystems, and a university lecturer. Each of these has changed how I see the others. The VC experience made me a better entrepreneur because I understood what capital is actually looking for. The ecosystem work made me a better investor because I understood the institutional and political dynamics that determine whether ventures actually land. The lecturing made me more precise, because you cannot hide behind intuition when someone asks you to explain your reasoning from first principles.

Running in parallel to all of this (not before it, not despite it) has been a life in writing and criticism. I’ve published poetry, translated literature, edited publications, and written about technology, culture and ideas for audiences that expected rigour and an original point of view. This is not a former life. It is the sensibility that runs underneath everything else: a critical, essayistic way of encountering the world that refuses to accept the received framing as the only one available.

The combination produces a specific kind of usefulness. I’m not a specialist who has gone deep in one domain. I’m someone who has operated seriously across several, and whose value lies in the connections between them — the moment when a pattern from economic history illuminates a venture design problem, or when a philosophical framework clarifies what a strategic choice is actually about.

For a full professional biography, LinkedIn is the right place.