The Real Problem of Preparing for the AI Wave: Taste and Judgment

When a wave of artificial intelligence reaches everyone at once, the real question is not destruction but adaptation: how do we adopt these tools in a way that maximises our own chance of doing well? The speaker says there is a great deal of doom around him, with people spending their time trying to convince each other which jobs AI will destroy. His claim is that, seen with enough distance, AI will create far more jobs than it takes away. He does accept that some jobs will disappear and be replaced by entirely different kinds of work.

The mechanism behind that claim is straightforward. Once the whole of human knowledge, plus the grey connective tissue that is not knowledge itself but can be derived from it, sits a click away, something enormous shifts. When information is free and instant, what remains scarce is the ability to sort it, weigh it, and place it correctly. The speaker compresses this into taste and judgment : the models are extraordinary instruments, but outcomes depend less on the model than on the person who imposes shape on its output. Taste is knowing what good looks like; judgment is deciding which output to use.

He shows what that means in his own work. His days are spent largely on investment and on building his 89 90 platform, and in both cases the substance is negotiation and persuasion. For that he built an agent: a Nash equilibrium agent that models multi-party gain dynamics, the trade-offs between stakeholders, the risks, and the players nobody at the table admits to. The point is not to win against one counterparty but to map how gains and losses are distributed across everyone involved.

The agent's contents come from his own reading. Principles drawn from business books and psychology literature are gathered into a single set of guidelines, with names like Robert Greene, Dale Carnegie and The Four Agreements among the sources. Whenever he reads something that strikes him as useful, he adds it to the corpus and then revises the text. The agent is therefore a kind of personal world model: not a template he downloaded, but a document that grows through a cycle of reading, selecting and pruning.

How the personal world model is built

The corpus has a second job, which is filtering. He reads the agent's output again, looking for the part that is rooted more in his own psychological weaknesses, and filters on that. A model's claim is not treated as true merely because the model produced it; he checks where his own tendencies enter. A second agent does more mechanical work: a constantly updated burn figure, a zero-risk, no-debt portfolio construction, Monte Carlo simulations and portfolio optimisation, then sending the results back to him daily. The system becomes a live feedback loop rather than a one-off question.

One rule is never to be in debt, a rule he admits violating to his own cost. When net interest income from the burn alone covers his entire cost of living after tax, life acquires complete optionality: he can move without the pressure of having to decide, and without rushing to avoid a bad move. He and his wife run something similar for health, turning blood tests, nutrition and the studies they read into a continuous update that surfaces hidden issues and asks what the numbers mean. The point is to use a second pair of eyes on the data rather than to hand a model knowledge he does not have.

The technical decision is not which model to pick but whether the surrounding system stays current. Grok, Gemini, OpenAI and Claude are all serious candidates; the real difference lies in how well the micro-services inside them are fed. That is real work, but once it is done there is a life model: a negotiation agent, a psychological filter, a portfolio and burn manager, a longevity and health tracker, and, occasionally, a layer that takes questions about raising children.

Curation is the part that cannot be skipped. Without maintenance and pruning, the model degrades into an average health-advice machine that recommends sauna and cold plunges to everything, because statistically that is what comes up most often. The person using it is the one with the expertise to tell the model what to think about next. The skill is less about writing the answer than about steering where the reasoning goes. He says he has spent a lot of time writing, pruning, reading outputs and judging them, and that the result is tuned to his own taste, without pretending there is one right answer.

When someone on his team asked for the Nash agent, his first answer was no, because handing over a system means the other person stops running their own. He has since come around to the idea of stitching the elements into a world model and letting people try it. The fastest way to find out whether such a system is useful is to run it against your own life, because that teaches you both what it is and, more usefully, what it is not. Used long enough, the main finding is that it is not a superweapon: it helps in some places and disappoints in a good many others.

Where taste and judgment come from matters as much. If you genuinely enjoy the work, the time you can give it is effectively unlimited, and expressing taste and judgment becomes easier. That luxury is not everyone's, though. Take a loan officer: even if a model processes every application end to end, the final yes or no is still a human decision, and there the weight of judgment exceeds that of taste. What the human gains is not a new job but time released from drudgery to think. What, then, should children study? He does not have a good answer, because he does not know which stage of the cycle we are in. On the judgment side the obvious route is engineering and physics, the disciplines that explain how the world works and in which everything is bounded by a law. On the taste side, understanding incentives and psychology will matter more and more, which points toward the common ground of philosophy, politics and economics. The honest position, though, is that nobody knows: the models are genuinely good, but whether they destroy jobs or create abundance is unsettled, and everyone holds a different stake in which guess proves right. The obstacle is that we live in an environment amplified by social media, optimised for clicks, so misinformation introduced at a moment like this simply multiplies. Had social media existed in 1910, at the start of the industrial revolution, the contagion would have been unbelievable; seen from here, that revolution was extraordinarily positive for humanity. The same arithmetic may apply today, though the institutions built in 1910 do not transfer.

What to study, and what nobody knows

What is needed now is a group of AI leaders who come together with solutions, the counterpart of the industrialists who saw what industrialisation would accrue to them and marshalled resources to pay it forward, so that the people leaving the farm for factory work found a library, a school and an infrastructure waiting. Done properly, the rising tide lifts every boat; left undone, it corrodes people's decision-making. And even if the end state is genuinely positive for everyone, we will not get there, because people will sit in a messy middle where there is no framework explaining why things may get worse before they get better, and where no credibility exists because the leaders never built the trust to be believed. His closing point is blunt: this is an incredible tool that will keep getting smarter, it will require humans with taste and judgment, and the AI community is failing badly at explaining that. The practical lesson lands plainly. Pick a single foundational model, build a handful of linked micro-services inside it, and prune them on a schedule. Which of those things earns its place can only be learned by testing them against your own life. Do not hand the model the burden of deciding or the bill for what follows; keep both, because judgment is exactly what grows there.