The Twenty-Watt Machine
Short watts and wafers. Shorter judgement.
"My experience is what I agree to attend to."
- William James
"Attention is the rarest and purest form of generosity."
- Simone Weil
The constraint on artificial intelligence is no longer ideas. It is not talent, and it is no longer really capital. Gavin Baker puts it in four words: we are short watts and wafers. Every frontier lab is bottlenecked by the same two physical inputs: advanced silicon and the electricity to run it.
We have spent five essays mapping this buildout, from the memory wall to the power stack, and they converge on a single conclusion: the binding constraint on machine intelligence is energy. Whoever solves deliverability, not generation, deliverability, owns the decade.
All of which makes it strange that the most energy-efficient inference engine ever built is sitting in your chair, and nobody is spending a shilling to learn how it works.
The Neglected Machine
The human brain draws roughly twenty watts, about that of a dim bulb. On that budget it performs feats the gigawatt campuses are still chasing: one-shot learning, causal inference from a handful of examples, judgement under genuine uncertainty and it does so for eighty years without a scheduled maintenance window.
Now that is a false comparison, the twenty watts is inference cost only. The brain arrives pre-trained, four billion years of evolutionary gradient descent amortised into the hardware, while the model’s megawatts must cover the training run from scratch.
So while the comparison is false, the more important question is why we treat the two machines so differently. We are spending three hundred billion dollars a year optimising one of them, every watt audited, every layer profiled, entire conferences devoted to squeezing loss from the training curve and essentially nothing learning to operate the other. The average portfolio manager knows more about the thermal design of an H100 than about the machinery producing his own decisions. He can tell you the chip’s failure modes. He cannot tell you his.
This essay is about the neglected machine: what it was built to do, why that differs from what we ask of it, and the one capability it possesses that no data centre yet does.
Ends Over Means
Longtime readers will notice I am breaking my own rule. I once argued, at length, that your brain is not a computer, and that metaphor obscures more than it reveals. I stand by that. But for the next few thousand words lets treat it as a machine anyway, because the machine framing exposes something the humanist framing hides: this machine has a capability no other machine possesses.
The capability is metacognition: cognition about cognition.
The machine can take its own processing as an input. It can watch itself decide, notice the watching, and revise the decider. Psychologists sometimes describe this loosely as having a theory of mind about your own mind, but the phrase is imprecise: theory of mind is the modelling of other minds, and what I mean here is stricter — metarepresentation, the recursive trick of holding your own thought as an object of thought. A model of the world is table stakes; every nervous system has one.
A model of your own modelling, editable at runtime, is the rare thing.
The obvious objection is that the data centres are catching up. Frontier models now critique their own chains of reasoning, grade their own outputs, revise their own drafts. Fair — but look at what the recursion covers.
Machines are acquiring metacognition over means: they can ask whether the reasoning was sound. They cannot yet ask whether the question was worth answering, because their objectives arrive from outside, welded in place by whoever wrote the reward function.
Metacognition over ends — the capacity to audit not just your reasoning but your goals — remains, for now, a twenty-watt monopoly.
If your brain can inspect and revise its own objectives, why do so few of us do it? The answer is that the brain ships with factory settings — and the factory was not optimising for what you think it was.
The Governor
In the 1990s the physiologist Tim Noakes proposed that the exhausted feeling that stops a runner is not the muscles failing. It is a governor in the brain, shutting the body down while a large reserve remains untouched, a safety margin, enforced by producing the sensation of collapse well before actual collapse. The limit you feel is not the limit you have. It is the limit the machine has decided to show you.
Take that mechanism out of the legs and put it in the mind, and you have the shape of the human operating system. Our hardware was tuned by an optimiser running a single objective for four billion years: survive long enough to copy the code. Survival is a conservative goal. It rewards flinching from the ambiguous shape in the grass, holding the position that kept you alive yesterday, treating a loss as more urgent than an equivalent gain. These were the correct settings for a creature that could die once. They are close to the worst possible settings for one trying to compound judgement over a career.
The asymmetry is easily measurable. Kahneman’s arithmetic — a loss weighed at roughly twice the gain — is not a bias in the sense of a mistake. It is the governor’s signature, the survival optimiser still pricing risk in a world where the stakes are money rather than blood. And underneath the arithmetic runs the chemistry.
John Coates, who traded before he studied the endocrinology of traders, showed that a run of losses raises cortisol, that sustained cortisol shortens the horizon and narrows the appetite for risk, and that a body marinating in stress hormones will refuse the very trade the mind knows is correct.
The governor runs below the level at which you can observe it. You do not experience factory settings as settings. You experience them as reality. Governors run at every scale. An individual has one. So does a generation of allocators. Their defining property is invisibility: a limit that announced itself as a limit could be argued with, and this one does not.
Which is exactly why the twenty-watt machine’s rarest faculty matters. A governor you cannot see is a governor you cannot override.
Metacognition And Flow
Ask people who study elite performers what excellence looks like from the inside, and they tend to name two things. One is relentless self-observation: the constant stepping-back, the audit of one’s own technique, the refusal to let any part of the craft go unexamined. The other is flow — total immersion, the self dissolving into the task, the performer so absorbed there is no one left to observe anything.
Gio Valiante, who has spent his career inside the heads of professional athletes and investors, names both as hallmarks of the best. He does not tell you how they coexist, and at first glance they cannot. One is the machine watching itself. The other is the machine forgetting it has a self at all. Metacognition and flow appear to be opposites.
They are, but the resolution is simple.
Metacognition is training mode. Flow is inference mode.
You do not retrain the weights during the performance. You retrain them in the workshop, deliberately, self-consciously, watching yourself fail and adjusting, so that on the stage you can run them without interference. The self-observation and the self-forgetting are not in contradiction because they never happen at once. They happen in different rooms.
The surgeon rehearsing the approach the night before is in the workshop; the surgeon three hours into the operation, hands moving faster than deliberation, is on the stage. The investor building the position, stress-testing the thesis, arguing with himself, is in the workshop. The investor holding it through a drawdown, when every input screams sell and the trained conviction holds anyway, is on the stage — and the conviction he is running was compiled earlier, in a quieter room, by a version of himself who did the work while it was still possible to think clearly.
A master’s real skill is neither mode. It is knowing which room he is standing in and refusing to run the wrong one. That is a metacognitive act about metacognition itself: the judgement of when to stop judging. It is also, not coincidentally, the hardest thing to teach, because it cannot be reduced to a rule. Which is why what comes next is not a set of rules but a set of protocols — the closest thing the twenty-watt machine has to a user manual.
The Three Protocols
Valiante states that the variance within an individual — the gap between a person’s best self and worst self — is larger than the variance between individuals. Your good days and your bad days differ from each other by more than you differ from your rivals.
This is both devastating to the way most professionals think about edge but also probably the truth when we really watch ourselves.. The edge you are chasing is not out there, in some skill your competitor has and you lack. Most of it is the distance between how you decide when the machine is well-run and how you decide when it is not. Close that gap and you have found more alpha than any new model will give you.
So the user manual is not about becoming someone better. It is about running the machine you have closer to the top of its own range, more of the time.
The first protocol is confidence. Confidence is self-efficacy — the specific, evidenced belief that you can execute a specific task. It is not self-esteem, which is a global feeling about your worth and largely useless under fire. Self-efficacy is domain-bound and earned.
Picture a portfolio manager deep in a drawdown. The governor is fully engaged, cortisol up, horizon short, every screen confirming the worst read. The instinct is to swing for recovery, to make the losses back in one decisive stroke. This is the machine, running its factory settings, walking toward the cliff. The protocol is the opposite. You do not rebuild self-efficacy with one heroic trade; you rebuild it the way it was built in the first place — from evidence, in small stacked wins, deliberately sized down until the machine has proof it can execute again. Baker, who lost heavily in a pharmaceutical blow-up early in his career and rebuilt afterward with the help of a few touchstone books and a hard exercise habit, compressed the whole discipline into a rule about knowing yourself: panic early or double down late. Both can work. What ruins people is not knowing, in advance and in cold blood, which kind of machine they are.
The second protocol changes the environment instead of the mind. Valiante calls it situated cognition: Do not try to will yourself into better behaviour, because the governor will win. Change the system the machine sits in. The trader who cannot stop checking the position moves the screen. The investor who panics on red days stops pricing the book daily. You are more programmable through your inputs than through your intentions, the machine will obey an environment long after it has stopped obeying a resolution. Metacognition’s most reliable output is not self-control. It is the decision to stop relying on self-control and re-engineer the room.
The third is about which engine you are running at all. There is a mastery motivation: the pull toward the craft itself, and an ego motivation: the need to be seen winning. They can produce the same resume and they behave completely differently under pressure. Ego distorts risk-taking because it is secretly playing a different game than the one on the screen: not is this a good trade but what will this say about me. Valiante’s line for the wreckage this causes is that we spend our adult lives undoing the debris of childhood. The machine can be run on either fuel. Only one of them lets you see the position clearly when it matters.
Billion Dollar PDFs
Jeremy Giffon, speaking on Invest Like the Best, describes a shift in what software sells. The old model sold strings: code written once and copied at zero marginal cost, the highest-margin business ever devised, because the second customer costs nothing to serve. The new model sells compute. The product is regenerated for every query, the cost recurs every time, and margins compress downward toward the physical layer, toward the watts and wafers we started with. What is true of software is becoming true of analysis. When a thesis can be generated on demand, the thesis is no longer scarce.
The marginal cost of a competent-sounding opinion is falling toward zero, and price, eventually, follows marginal cost.
And where does the abundant analysis go? Onto the timeline.
Giffon’s observation is that everyone now reads the same paper. The feed serves hundreds of millions of people roughly the same five hundred posts a day, and among the readers are nearly all of the people who matter to capital: the allocators, the founders, the policymakers, the journalists who used to be the paper. It is the global newspaper, except that you can watch the influential react to each article in real time, and their reaction becomes the next article. Giffon’s precise claim is that with passive flows absorbing most of the volume, the judgement that remains - the marginal, discretionary, human trade - is increasingly fed its narratives by an algorithm optimising for a single quality, and that quality is not accuracy. It is entertainment. The last human decision in the pricing chain is downstream of a machine choosing what that human finds most fun to believe.
Which is how you get what Giffon calls the billion-dollar PDF. Every so often someone crystallises a notion at exactly the right moment, in exactly the right register, confident, quotable, arriving when everyone is uncertain and looking for someone to set the story and billions of dollars of capital form around it. The document does not need to be right. It needs to be timely and sure of itself, and the capital follows it the way ten-year-olds play football: everyone chasing the ball, the ball being the story.
A billion-dollar PDF is an objective function, installed from outside, into minds that never audited the installation. The model’s goals are welded in by whoever wrote the reward function. The allocator’s goals, it turns out, can be welded in by whoever wrote the PDF and the weld holds until the next PDF comes along.
So run the compression on yourself. When analysis is abundant and narrative is algorithmically selected, what remains scarce?
Not watts. Hours.
Some of what remains is chemical. The machine feels no fear, and on the narrow question of sitting still through a drawdown that makes it our superior, not our inferior; it holds for free what costs us everything. But that is not where our edge was. The edge is that markets are still priced, at the margin, by other twenty-watt machines running the same governor, cutting the position at the wrong moment because the cortisol got a vote. Conviction held through a drawdown is scarce not because the machine lacks it, but because almost every human on the other side of the trade lacks it too. The bravery is an edge over them, never over the silicon.
Machines are now learning to optimise means. They will out-analyse us, and soon. What they cannot yet do, because their objectives are welded in from outside, is choose which game is worth playing at all. That decision is metacognition over ends, and it remains a twenty-watt monopoly.
The tempting version of this ending is: understand the metagame, and win it. Learn how the timeline prices securities and trade against the people it captures. But that is a trap wearing the costume of wisdom. Playing the timeline more skilfully is still being played by it — captivity with better statistics. The metagame is still a game someone else chose for you.
The rarest use of the machine is not out-competing at the game everyone is playing. It is the allocation decision that sits above every game: what to spend the twenty watts on, and whether a given board deserves them at all. This choice exists at every scale — which meeting, which trade, which argument, which feed — and almost no one makes it deliberately at any of them. Some readers will know the particular quiet that follows finally asking.
Most people never once ask whether the board in front of them is theirs to play.
What we covered:
Remember the governor: The brain wasn’t optimised for judgement, it was optimised for survival — and survival is a conservative objective. Loss aversion is the governor’s signature, cortisol is its enforcement mechanism, and its defining property is invisibility: factory settings aren’t experienced as settings, they’re experienced as reality.
Metacognition is the override: The machines are acquiring metacognition over means — they can ask whether the reasoning was sound. They cannot yet ask whether the question was worth answering, because their objectives are welded in from outside. Metacognition over ends remains a twenty-watt monopoly. Metacognition is training mode, flow is inference mode, and the master’s real skill is knowing which room he’s standing in.
As analysis is commoditised, the scarce input moves upstream — to attention, and to the choice of board: What remains scarce isn’t information, it’s undivided attention, the raw material from which trained perception is compiled and the only raw material it has. Which is why the timeline matters: the most heavily engineered artefact of the age, optimised for entertainment rather than accuracy, aimed precisely at that input, installing billion-dollar PDFs as objective functions into minds that never audited the installation. The rarest use of the machine is the allocation decision that sits above every game: what to spend the twenty watts on, and whether a given board deserves them at all.
Don’t share this because it was good. Share it if it changed what you’ll do on Monday.
Everything else is entertainment — which, as we’ve established, is what the algorithm was optimising for all along!