Five arguments I keep coming back to.

The asteroid and the dinosaurs.

As the number of levels in a hierarchy goes up, the effectiveness of the organization goes down. Each person at each level has their own read of the situation, their own incentives, their own face to save. They slightly modify the order from above to suit themselves. Then the next person modifies it again. By the time a decision made on the seventh floor reaches the customer, it has passed through six rooms and is no longer the decision that was made. This is the classic principal-agent problem from economics, applied to organizations of all kinds — companies, governments, schools, the bureaucracy that manages your land records. It is not a moral failing. It is the cost of asking humans to faithfully relay anything through more than two or three layers.

For most of the twentieth century we paid that cost because there was no alternative. If you wanted to coordinate a thousand people you needed a hierarchy. The hierarchy leaked, but the alternative was no coordination at all. So we built the leaky pipe and accepted the loss. Whole management traditions — General Electric in the eighties, McKinsey at any time — are essentially the art of making a leaky pipe slightly less leaky.

What changes in 2026 is that the alternative now exists.

An AI-enabled small organization does not need seven layers. Most do not need three. A founder with a working set of agents can read the field as well as a regional manager once did, draft a campaign as well as a marketing team, run customer support at a quality the customer notices, and write the production code. The principal and the agent are sometimes the same person, sometimes the same person plus a model, and almost never seven people in a chain. The leak gets short or disappears. Decisions made on Monday morning hit the customer on Monday afternoon — not because the team is heroic, but because there is no chain to corrupt.

This is not a prediction about chatbots replacing knowledge workers. It is a prediction about org-design. The unit of competition is shifting from company size to decision latency, and decision latency is a function of how many human approvals stand between a decision and its execution. AI does not just make the workers cheaper. It makes the layers unnecessary. A small org with good agents can compete with a giant on quality, undercut it on speed, and out-iterate it on every cycle. Three of these competitors at once will eat the giant's lunch in any category that does not require physical capital.

Hence the metaphor. AI is the asteroid that will probably destroy the giant dinosaurs of the organizational world. The organizations being built today don't need 7 levels. Most don't even need 3. Some of them are run by two people and a Claude key, and they are taking serious revenue out of incumbents who employ thousands.

Two qualifiers. Probably. This is a directional trend, not an absolute. The biggest dinosaurs — banks with regulatory moats, telcos with spectrum licences, governments — will outlive the small mammals for a long time, because the moat isn't org-design. And destroy is too clean a word; what actually happens is more like a slow drift, where the small org takes a slice, then another, and one quarter the giant looks at its growth chart and realises the slope has changed. By then the asteroid has been arriving for years.

The implication for builders: do not optimize for the org chart you would have drawn in 2018. The right shape today is two or three people, a model, and a clear decision-loop you can run on Monday morning without asking anyone's permission. Hire only when the leak from not hiring is more expensive than the leak from adding a layer. That trade-off is not what it used to be.

The implication for everyone else: when an AI-enabled three-person company is competing for your attention against a thousand-person incumbent, bet on the three. Not always — but increasingly. The script has flipped.

Tell me if there's another point of view.

The third option.

Almost every framing I've been handed over the last decade has been a binary. Build in public or protect your time. Raise from VCs or stay bootstrapped. Open-source the core or hold it close. Long hours or clean boundaries. Tech-native kids or unplugged kids. Mainstream school or unschool entirely. The framings are almost always wrong. Not because both sides are equally bad — sometimes one is — but because the genuinely useful answer is the third thing built around a system, and the binary forecloses the question before that third thing can show up.

Take building in public. The two camps are loud. Camp A says you must publish daily, share every metric, run your company as a documentary; the algorithm rewards consistency, the audience compounds, the silent founder is the dead founder. Camp B says protect your founder time at all costs; output is the only thing that compounds; the public-by-default founder is performing instead of building. Both have a point. Both also have a counterexample for every claim.

The third option is neither. It is to design a content engine: produce a lot in a short, protected window, then post daily across multiple platforms, each cut tailored to the platform and to a different aspect of the strategy. The founder time is the input you protect. The output is loud anyway, because you have made the production process efficient enough that the audience cost is covered without further founder hours. You get the compounding without the performance. The binary asked you to choose; the system refuses to.

The same shape recurs. On VC, the binary is raise or don't; the third option is raise only when the moat-shape rewards it — sometimes that's network effects and a flywheel, sometimes it's a deep R&D bet, sometimes it's neither and the right answer is to bootstrap into a category the VC mind doesn't know how to model. On long hours vs boundaries, the third option is to work long hours now, with explicit family-protected windows inside the day, while building toward a target where seventy or eighty per cent of weeks have clean boundaries — and to be honest about the gap between current and target rather than pretend you've already arrived. On parenting, the third option is the barbell: full tech-fluency on one end, full ability to live without civilisation on the other end, no suburban middle.

The pattern: when someone offers you A vs B and both feel slightly wrong, it is usually because the real answer is a system that produces the benefits of A and B at the cost of neither, but that requires upfront design. The binary feels right because it is a faster decision. The third option feels slow because it requires you to think for a few weeks. That is exactly the asymmetry to exploit. Most people will take the binary. The third option is unguarded territory.

Two cautions. The third option is not a magic word; sometimes A is correct and you should just pick A. (When the building is on fire, run.) And the third option costs more upfront in design — you build a system instead of just choosing — so it is wasteful for low-stakes decisions. Reserve it for decisions you will be making for years. For the rest, flip a coin and keep moving.

The general move is: refuse the binary, design the system, accept the upfront cost, take the unguarded territory. Do this for two or three of your big calls and the leverage compounds.

Two steps behind.

Regulation should always be two steps behind innovation. This is the line I have ended up at after watching twenty years of regulators try to govern things they do not yet understand, and after watching the things they did not regulate become the engines that pulled hundreds of millions of people out of poverty.

The argument is asymmetric, and the asymmetry is the whole point. If you regulate too late, some people get hurt. The hurt is real and should not be dismissed. But it is bounded — you can see it, you can measure it, and you can intervene case by case once it shows up. If you regulate too early, you suppress a generation of progress before anyone can see what it would have looked like. That cost is invisible, because it is the unbuilt product, the unfounded company, the unwritten essay. You cannot count what was never made. Almost every regulator I have seen is calibrated for the visible cost and blind to the invisible one. They protect themselves from the headline of the harm; they do not protect the country from the silence of the absence.

So the discipline is wait-and-watch by default. Intervene only when catastrophic consequences are visible — and when you do, intervene case by case, light-touch, until the shape of things becomes clear. When fire was first invented some people burned their hands. We did not put the fire out. We learned to build hearths, to keep buckets nearby, to teach children to respect it. The hurt was the input to the learning, not the input to the ban.

This is not a libertarian argument. I am not saying government should not regulate. I am saying it should regulate second. The first move belongs to the builder; the second move belongs to the courts and the legislators, and only when there is real harm to point at, and only with the smallest tool that works. Once the dust settles you can codify what you have learned. But codifying first is just throttling.

India in particular has lived the cost of regulating first. We had a forty-year industrial-licensing regime that was a regulation-first system applied to the entire economy. The cost was not theoretical; it was a generation of poverty. The 1991 reforms moved us toward a regulation-second posture in many sectors and the consequences are visible in any chart of Indian living standards from then to now. We should keep going. Intellectual property creation should be as easy as a UPI transfer. Judicial delay — the regulation-after-the-fact failure mode — is the single biggest brake on the country, because no redress means people fear conflict, and people who fear conflict don't enforce contracts, and unenforced contracts mean too much uncertainty for anyone to take real risk. Tech, AI, plea bargains, a hundred small reforms at once. Most officials' time should go to affirmative action for poorer sections and to clearing roadblocks for the private sector. The opposite of what often happens.

The hard cases are real. AI safety is one. Nuclear is another. Where the harm is potentially civilisation-scale and irreversible, the asymmetry flips: the cost of intervening too late dominates. There you regulate sooner, but still narrowly — restrict the specific catastrophic capability, leave the surrounding domain free. The principle generalises: match the regulation's speed to the reversibility of the harm. Reversible harms can be caught after the fact; irreversible ones cannot. Most regulation should be slow. A small fraction should be fast. The discipline is in telling them apart.

Two steps behind is not nothing. It is two steps. It is not zero. The regulator is not absent. The regulator is just patient.

Credentials are not wrong, just outdated faster than they replace.

For a long time credentials were the best proxy we had. If you wanted to know whether a stranger could think clearly, the IIT entrance exam or the IIM CAT was a serious filter, because there was nothing else to look at. You could not see their work. You could not see their public reasoning. You could not see their code. So you looked at the four letters next to their name and made a decent guess. The system worked because the alternative was no signal at all.

That is the steel-man, and it deserves to be stated. The credential is not a scam. It was a reasonable response to an information vacuum. Anyone who tells you elite credentialing was always a fraud is rewriting history to flatter their priors.

The vacuum is gone. By 2026, on any given candidate, you can read their last hundred opinions on a public timeline, see the code they have shipped on GitHub, watch their reasoning unfold in long-form writing, run them through an AI-conducted skill challenge that takes their actual present competence and not their seventeen-year-old test score, and verify their history of past work to a depth that did not exist a decade ago. There are at least six richer signals than the credential, and each one is updated continuously rather than fixed in 2008. The credential is now one data point among many. For most hiring decisions today it is a small input — useful, not load-bearing.

Two implications follow.

The first is for hiring and partnering. Update the prior. The IIT-IIM stack still selects for some genuine attributes — discipline, raw cognition, the ability to grind through a structured exam — but it does not select for many of the things modern work cares about: taste, judgement, the ability to operate without supervision, the ability to learn a new tool in a weekend. The Indian VC scene is heavily biased toward the credential-stack and the bias is increasingly mispriced. A founder in a Tier-2 city who has shipped three real products and has 20,000 followers reading their public reasoning is not a worse bet than an IIT-IIM with a deck. Often they are a better one. The credential overweighs in the room because the room is mostly built of credentials. That overweighting is the asymmetry to exploit.

The second implication is for neurodivergent talent. The credentialed path enforces a rigid externally-imposed structure — sit in this seat for four hours, answer these questions in this order, do it on this date. For some people, that structure is fine. For others, ADHD or dyslexia or any of a dozen other shapes, it is hostile. The work itself is not the problem; the demonstration ritual is. Now that we have other ways to demonstrate the work — public artefacts, shipped products, recorded conversations, public writing — the case for forcing every kind of mind through the same ritual gets weaker every year.

None of this is an argument for abolishing the credential. For most candidates, in most contexts, it remains a useful signal. The argument is that it should sit beside the others and be weighed against them, rather than function as the gating filter that lets the others through. The credential decays slowly; the alternatives accumulate fast. Eventually, the alternatives win. We are not at eventually yet, but we are well into the transition. Calibrate accordingly.

The system before the substance.

For ADHD, and for almost any complex-system problem in a human body, the order of operations is sleep, exercise, nutrition, capture, ritual — in that order — and meds last, only if required, only to the extent required. The order itself is the content. Almost every mistake I have made in managing my own attention has been to invert it: reach for the active substance first, hope the rest sorts itself out. It almost never does.

I want to be careful here. I am not anti-medication. I am not telling anyone to stop their treatment, and I am not running the line that ADHD is a discipline problem cured by cold showers. The lazy version of this argument is dangerous and I disclaim it explicitly. What I am saying is narrower: when the underlying system has not been built yet — when someone is sleeping six hours, sedentary, eating whatever, capturing nothing, with no daily rhythm — adding an active substance to that system is intervening in a complex system whose baseline you have not measured. You cannot tell what the substance is doing because the system has too much noise. Build the system first. Then if a gap remains, intervene with the smallest tool that closes it.

This generalises. The principle is reversibility. Cheap, reversible interventions go first. Expensive, irreversible ones go last. Sleep is reversible — if eight hours doesn't help, you switch back. Exercise is reversible. So is a capture protocol, so is a daily ritual. The cost of trying any of these and being wrong is roughly zero. An active substance is harder to walk back: it changes your physiology, sometimes it changes your tolerance, and the data you collect while on it is data about you-on-the-substance, not you. You should still try it if the need is real. But not first.

The reason this order works for ADHD specifically is that ADHD is interest-conditional. When something is hard and interesting, hyperfocus engages and the work goes well. When something is hard and not interesting, avoidance kicks in and nothing happens. The system-design problem is not making yourself work. It is breaking the work into pieces small enough that you don't avoid them, but not so small that they bore you out of the task. Goldilocks-sized parts. A walk to think when you're stuck. Pen and paper for the harder calls. An LLM as an objective sounding board when you need to be challenged. Capture the result fast, before it evaporates. None of that requires a prescription. All of it requires that you build it deliberately, and most people don't, because building a system is less satisfying than reaching for a fix.

I built one. Eight months ago I had multiple active projects, a calendar I didn't trust, and work eating into my evenings. Last week I shipped a personal operating system that runs as my alter ego. The hardest part wasn't building it. It was accepting the truth that it showed me — the things I was avoiding and the patterns that repeat. Despite giving up some control over the calendar and some admin aspects of my life, I feel more in control now. The substance question — meds or not — is now a much cleaner conversation, because the noise around it has been removed.

If you take one thing from this: when something complicated is going wrong inside you, build the system before you reach for the substance. Sleep, exercise, nutrition, capture, ritual, in that order. Then talk to a doctor, with data. The order is the content. Don't skip it.