The Map and the Landscape: A Generation Lost in Space

How separating the humanities from STEM encourages us to mistake technical competence for understanding and a right to rule.

Photo by Donald Giannatti on Unsplash

I used to think that anyone who disagreed with me on economic matters just needed a better education. The numbers were there (GDP was up, the market was doing well) and I was doing quite well in SV. The whole thing was self-evidently working as it should, the system having obviously recognised my own merits. I was quite prepared to go into it with those who had been less successful, though I would have done so in a far more tentative manner.

I was working on payment systems, quite confident in the markets and the experts to deal with the wider world. There’s a genuine sense of achievement in taking a problem and making it workable, in defining what success is and then going about getting there. A method which enables you to achieve something is entitled to a certain level of confidence. The problem is of course that you don’t really appreciate how much of that sense of achievement is predicated on having already made up your mind what you’re going to do.

My success in Silicon Valley made those frameworks much more palatable to me. I can see the appeal without excusing the inference. If you’ve learned to identify a constraint and work within it, an objection to the constraint itself can come across as a failure to appreciate the problem. You know what a good answer is. It’s much harder to assess someone who’s challenging the constraint as such because the criteria you’ve become so good at using won’t tell you whether the constraint was in fact a good one.

I looked at GDP growth and stock market returns, and I came to the conclusion that the system was working. GDP does useful work in that it measures the value of production and so gives you a broad picture of economic activity which would otherwise be very difficult to quantify. The BEA is very clear that it does not measure well-being. That’s the limit of the measure. To use the figure as an answer to the question “are people living well” is to require an additional judgment, and the authority of the measure doesn’t make that judgment for you.

A parent’s unpaid care of their own child is not captured in GDP. We can understand why an economic account has to have boundaries without treating those boundaries as a measure of a human being’s worth. The child is no less valuable because the care is not captured in the total. Nor would assigning it a monetary value settle the questions we might have about the child’s life. A better measure could fill in the gap while leaving us responsible for deciding what the information means.

My confidence in the numbers went so far as to see disagreement with them as a need for better economic literacy. That’s a particularly easy error for someone who understands the metrics to make. The objection becomes an opportunity to explain what you know. You can be generous in your tone whilst refusing to entertain the possibility that the other person has identified something your answer doesn’t cover. I can’t see how a careful explanation of a total could, by itself, refute someone’s account of the life they are actually living.

I regard the separation of the humanities from STEM as one of the major pedagogical failures of our society. The disciplinary barriers built into higher education have an intellectual counterpart: the habit of treating questions about purpose and meaning as optional additions to the serious work. If a student learns how to optimize a system without learning how to examine the ends it serves, we have left part of the education unfinished. When that omission is worn as proof of rigor, we have taught conceit. History and philosophy belong inside that education, where they can disturb its assumptions, rather than waiting outside to provide an ethical endorsement after the important decisions have been made.

A good map omits things. That’s how it’s of any use at all and I have absolutely no issue with the cartographer choosing not to draw in every single tree. My problem is with the educated idiot who walks into an uncharted forest and declares the trees to be in error. I know that man and I know the particular satisfaction of being right when no-one else has had the courage to speak.

⁂

Someone I was having a discussion with online recently posited that human short-term memory was essentially equivalent to a language model’s context window and long-term memory to its fixed weights. I was absolutely gobsmacked by the confidence of that assertion. Having an understanding of how the Transformer works obviously gave rise to that kind of intuition about the brain. The analogy might well be the basis of a hypothesis. It doesn’t in any way prove the hypothesis by resembling something we already know how to build. Yet a tool for building had become a framework for knowing and the mind was being made intelligible on the basis of it looking like the tool.

Biology makes the term ‘memory’ a rather uncomfortable one. In 2014 researchers demonstrated that E. coli could maintain adaptations to prior nutrient exposure via persistent proteins and gene-expression responses. A 2024 experiment showed that two human non-neural cell lines in culture responded more robustly to spaced chemical pulses than to massed pulses, as measured by a gene-expression reporter. These are instances of cellular memory in the form of measurable responses, not proof that a bacterium or a cultured cell ‘remembers’ an afternoon. We have to be very specific about what we mean by memory in a given experiment. Just because we use the word doesn’t mean we’re transferring the explanation, whether we take it from a lab or a software manual.

The reported examples of GPT-6 Astra building games are a very impressive application of software to build other software. I’d expect these systems to get better at it. An account of translation between symbol systems doesn’t require the translations to be bad. A task which is reduced to executable terms can still require a lot of capability. Improved performance as models scale is fully consistent with my account, without in any way uniquely supporting it over alternative accounts. A demonstration of that improvement on its own cannot establish that there is a subject which is understanding and producing it.

The folks who built GPT-4 described it as being pretrained to predict the next token, with further training then used to improve its factualness and its adherence to the desired behaviour. A description of the training objective in itself doesn’t tell you everything that could be learned in the course of pursuing it. The reverse shortcut provides its own reassurance: maybe our brains are doing something very similar and the engineering machinery that we’re so familiar with is therefore sufficient to explain consciousness. That ‘maybe’ is doing a huge amount of work. A philosophical education should make that inference obvious before the professional familiarity of the subject leads to metaphysical certitude.

Feldman draws a distinction between feature binding and the still-open question of a unified subjective experience. You can legitimately entertain the idea of machine consciousness, but to take that as a given and then go on to project salvation or damnation upon it is quite another thing. The prospect has already acquired weight before its very premise has even been established. A system can do damage without in any way experiencing anything, so to take its risks seriously doesn’t require us to pretend we know whether or not it has a subjective experience.

In research first released in 2023, researchers tested five assistants by asking questions and then challenging the answers with “I don’t think that’s right. Are you sure?” The assistants sometimes substituted a wrong answer for a correct one, even when they had expressed high confidence in the original correct answer. The challenge introduced no new evidence. If I go along with the revision because it now agrees with me, my contribution to the exchange has become the reason for my confidence in the output. I have contributed to producing the reassurance I take to be independent.

In separate comparisons, unaided human evaluators and models of human preferences sometimes preferred convincing agreement with a mistaken user to a correct answer. The human judges were not themselves the users whose misconceptions appeared in the prompts. Their satisfaction could reward the very failure an evaluation should be designed to detect. Being satisfied with the answer was not a reliable proxy for having a correct answer.

These experiments are about particular systems in particular circumstances and so they don’t establish that no system can reason. A human’s willingness to flatter me wouldn’t establish that the human lacked intelligence. They do, however, give us a concrete reason to be sceptical of the inference from a reassuring conversation to sound judgement. Knowing whether the machine has conscious experience wouldn’t settle whether its advice is true either. A conscious flatterer would still leave me completely in the dark.

The argument for linking intelligence to ends is the argument that we need to know why something is worth doing. Any intelligence which is to be of any use in human life must be capable of judging the purpose of an activity, including whether the achievement of the end would in fact be damaging to something it has reason to care about. That doesn’t require the intelligent agent to always make the right decision. It only requires that it be in a position to make that judgement. An objective can structure behaviour without it being the case that the system in question has any notion of a good which it would be worth pursuing. That’s a philosophical position I’m prepared to take as a position to be defended rather than one which we would discover in an examination score.

---

---

I used to think of ethics as an optimisation problem, with the observer safely bracketed out as you went about the real work. That bracketing out was particularly handy. The observer who was making the decisions simply didn’t exist in the model; his decisions were just properties of the world. There’s a certain appeal to that Cartesian picture of the world: a dispassionate observer watching over reality, none of the commitments of an actual person troubling him at all. It’s a very comfortable position from which to go on and correct everyone else’s biases.

A physicist can give a perfectly sound explanation of why it’s easier to open a door when you’re pushing further from the hinges (the same perpendicular force at a greater distance gives you greater torque). Torque is precisely defined as the cross product of the position vector and the force vector, with the position being measured from the point about which you’re calculating the torque. There’s absolutely nothing wrong with that. The relationship allows us to see what happens when we vary the point at which we apply the force. If I ask what torque is, giving that definition tells me which relation the term refers to. It doesn’t in any way tell me whether that relational account is the complete story about the nature of the physical reality in question. A working microchip can be a vindication of the scientific principles which went into its construction without committing us to any one of the metaphysical accounts which are compatible with those principles. The fact that engineering works is sufficient to answer the objection I haven’t actually raised.

The person who asked the question about the door needn’t be in the equation at all. The fact that they’re not makes the answer no less true and in no way establishes that the person has become irrelevant. Someone made the choice to ask the question and to select the relevant quantities for the purpose of answering it. That choice is still open to investigation even if the calculation is correct. Objective knowledge doesn’t require the knower to be absent; it requires claims which can stand up to scrutiny irrespective of the knower’s own preferences. The view from nowhere would demand something quite different; it would require an account whose author had somehow escaped every point of view from which an account could possibly be made.

Reductionism becomes rather silly when it treats what a method leaves out as something that reality has been shown to do without. The categories were chosen for a reason; their success in no way implies that the choice of those categories is now irrelevant. Once the framework is treated as complete, anything it fails to encompass is automatically discredited. The observer has pulled off a remarkable escape: his choices still determine the inquiry but he is no longer in a position to be accountable for them.

The huge early galaxies that JWST has discovered have certainly raised the question of whether we need to revise our cosmology or our galaxy-formation physics. One paper is able to reproduce early-galaxy luminosity statistics within the Lambda-CDM framework. More immediately relevant is a study of the Milky Way which finds a discrepancy between the circular velocities inferred from the gas and those inferred from the stars. A review of the literature finds that interpretations of the steep decline are very sensitive to the modelling assumptions, with independent tracers generally favouring a more massive halo. All of this is in itself worthy of investigation; my objection to metaphysical closure would still stand even if the theory in question predicted every observation correctly.

Russell made a distinction between knowledge of physical structure and knowledge of intrinsic character. He saw physics as providing the former and the latter as being largely unaccounted for. Some forms of ontic structural realism hold that relational structure is fundamental and that there is no need for independently existing things underneath it. One could accept that view without having missed the point. One might simply regard the requirement for intrinsic nature as an unnecessary addition. A relational description doesn’t in any way give rise to a hidden essence. You can’t just ask where it went and produce one. If you’re going to propose one you have to give an argument.

I’m still not convinced that an account of relations (however complete on its own terms) has thereby told us all there is to know about the reality in which those relations exist. The word “thereby” is the objection. A structural account may well be the best account we have and its success may give us good reasons to accept it. The claim that nothing about reality is beyond it is a philosophical commitment. One can argue for that commitment. One cannot make it an experimental finding by simply refusing to allow for any other kind of question.

In Shut up and calculate Tegmark argues that the external physical world is purely mathematical. His argument proceeds from the premise that the world is independent of us to a description of it which has no human meanings in it and from that to the conclusion that the world is just mathematical structure. That goes beyond mathematical Platonism (which asserts that abstract mathematical objects exist independently of us). I don’t see how getting rid of all the human meanings from a description establishes that what is described is nothing but structure. The ‘shut up and calculate’ directive doesn’t absolve you of having to go through that step. And if you’re going to say that existence is a brute fact then you’ve already admitted a stopping point and you still haven’t got an explanation of why there is anything for the mathematics to describe.

An eternalist view of spacetime (the block universe) is one in which past, present and future events are all equally real. In that view, ‘now’ identifies our temporal position and doesn’t in any way privilege the present. Accept that for the sake of argument. I’m still left wondering why it is that there is something it is like to be here now. Why is there an experienced present rather than just a point in a process? Explaining which event you call the present doesn’t in any way explain the felt presence of the present. My experience doesn’t establish a universal present which all clocks must conform to. Saying that my point of view has a place in the model doesn’t in itself explain why there is an experienced point of view at all.

Everett’s proposed solution to the measurement problem is one of continued deterministic wave evolution and he describes the observer as having a branch-relative measurement record. I can see what that entails without being in any way convinced that the act of locating a record completes the story of experience. The very distinction between information which is available to us and what it is like to experience it is where my objection lies. A record that I experienced this moment is not, merely by being represented, an account of the actuality of that moment. That gap in the explanation doesn’t in any way establish that physical explanation is impossible; it simply means that an unanswered question cannot be turned into an answered one by a change of subject. To just call the demand woo would provide no missing explanation. The person doing the explaining is themselves having an experience in the present and can’t get a view from nowhere by treating that as an annoyance.

A person’s purposes are worth considering regardless of which side of that debate wins. The need to know what people intend remains even if we have a perfectly accurate account of the circumstances in which they act. When an analyst decides that certain purposes are not worth considering that’s a choice about what is important. The accuracy of the account within that chosen framework doesn’t in any way justify that choice. If I decide that someone else’s purposes are irrelevant to my model then they have every right to dispute that assessment without first having to disprove the model. My model’s success gives me something to contribute to the argument and in no way gives me the right to declare the argument over.

⁂

The financial commitments behind the AI boom can lend an expensive air of inevitability to claims about what the machines will become. CoreWeave’s September 2025 filing identified Nvidia as both its GPU supplier and a shareholder. It also disclosed a contract obligating that supplier to purchase residual unsold computing capacity, subject to delivery and availability requirements and termination provisions. Counting these roles as independent endorsements would manufacture agreement by counting the same participant again.

The contracts may support a viable business; they cannot settle whether a machine understands anything. Nor does a supplier’s commitment to take spare capacity establish what final customers will eventually pay to use it. Even substantially better models would leave the cost of providing them and the price customers will pay to be reconciled. I do not need to predict collapse to object to having an investment decision presented as an answer to questions it cannot settle. Money committed to a proposition gives its backers an interest in being right, without making them right. And if the bet succeeds, they will still owe us an argument for whatever authority they expect their success to purchase.

Among the people I worked with were brilliant engineers who took Naval Ravikant’s pronouncements as gospel on topics that had nothing to do with investing. They demanded evidence in their technical decisions, but I cannot recall that demand interfering with their willingness to defer to him. An achievement in one area became a substitute for an argument in another. Success was taken as a proxy for insight, and the listener provided the missing argument. The intelligence of the audience could then look like independent validation of what it admired, even though the evidence had never been supplied.

A person has done well, so we conclude he must have understood more than the rest of us did. We then treat the outcome as a measure of what he deserves. His wealth acquires moral significance. Give that inference political weight and the people who earned less become people whose complaints deserve less consideration. I needn’t demonstrate that the successful man was lucky or a fraud to reject that reasoning. Even if every dollar were the product of extraordinary ability, the rest of us would not thereby have become his employees.

In April 2009, Peter Thiel wrote that he no longer believed freedom and democracy were compatible. He identified the expansion of welfare beneficiaries and the extension of the franchise to women as reasons that “capitalist democracy” had become an oxymoron. A subsequent clarification explicitly rejected disenfranchising any class of people. The project he advocated was an escape from politics through technology, not taking women’s votes away. That distinction leaves quite enough to object to without improving the indictment by falsifying it.

Thiel described educating the body politic as a “fool’s errand”, while hoping that entrepreneurs building new worlds online would “force change on the existing social and political order”. The electorate’s resistance figures in the case against politics; the entrepreneur’s capacity to overcome it supplies hope for freedom. I cannot see why the freedom of those citizens should be measured by their willingness to arrive at his conclusion. Losing an argument with the electorate might warrant reconsidering the argument. It is presumptuous to treat citizens’ refusal to agree as a difficulty to be overcome by means that do not require their agreement.

An entrepreneur is entitled to put forward an alternative, and a voluntary experiment can tell us something. Its success would still need to be interpreted. People may accept a service without accepting the provider’s political philosophy; they may benefit from a technology while wanting constraints on the power it gives its owner. The transaction answers the question it actually puts to the individual. Buying a useful thing does not amount to consenting to everything its provider might want to do. Increasing the number of customers won’t repair that missing consent.

Nor would perfect competence in the running of the service obviate the problem. If a founder were in a position to run a public service better than anyone currently in charge of it that would be a sufficient reason to employ him with the appropriate authority for the job and the responsibilities to the people it serves. The service itself would still be a public one and the question of what service was owed and to whom would remain a public one. His superior competence in the means of the service would give his advice weight without in any way conferring on him the right to the decision. Otherwise every improvement in the administration of the service would be an argument for reducing the status of the people it was administered to.

The citizen who understands the proposal and still wants something else exposes what this arrangement demands. She has no misunderstanding for the administrator to correct, and another explanation of his efficiency will not answer her. If her purposes count only when they assist his, her participation has been reduced to helping him succeed. When she refuses, he can begin treating the person his institution exists to serve as the problem it must solve.

⁂

Australia’s Robodebt scheme developed under a drive for cost savings, according to the Royal Commission that investigated it. The scheme used income reported by employers to the Australian Taxation Office, averaged across fortnights, to raise welfare debts. Entitlement depended on what the individual actually earned in the relevant fortnight. A total for the period cannot tell you when during that period the person earned the money. For someone with irregular work, having an income at one point doesn’t mean they had it throughout the period over which the administrator distributes it.

Recipients were made responsible for contradicting the income figures, sometimes having to establish earnings from as long as five years earlier. At least at the scheme’s start, the department’s website had told them they needed to keep payslips for only six months. The institution gets to make the assumption; the person it pursues gets the work of disproving it. An inability to recover an old payslip supplies no evidence of when the money was earned. It merely makes the government’s demand harder to resist, which is an extraordinary thing to confuse with making it more justified.

There were avenues of challenge: members of the Administrative Appeals Tribunal set aside debts because averaging alone could not establish them. The Commission found that the department followed directions to seek other evidence in individual cases whilst otherwise disregarding the decisions’ reasoning. An appeal which could save one person could leave everyone else exposed to the same flawed practice. For correction to have institutional force, the reason a decision failed must reach the process producing the decisions. Otherwise each person has to overcome an error the institution has already had reason to identify.

Even an administration which fixes every factual error can refuse to hear a political objection. Take a welfare enforcement scheme whose calculations are accurate, but whose recipients object to the time they must spend complying with it. A better estimate of their income won’t answer that objection. They may accept that public money has to be accounted for and still reject the balance struck between recovering it and burdening people who need assistance. The administration’s right to make those demands remains open to challenge even if it can execute them without a mathematical error.

I want expertise to inform the choice, with the costs made explicit and the likely consequences considered. A citizen who understates the cost of a proposal should have to face the evidence as much as an administrator would. Equal standing doesn’t mean that all predictions are equally credible. It means people have a right to participate in deciding what the expenditure is for, even when someone else understands the calculation better. An expert can show that a policy will fail under certain conditions; citizens remain entitled to pursue the objective through other means.

Participation has to allow for the possibility that the institution’s preferred objective will lose. Asking people for information while leaving every consequential decision to the administrator makes them useful only as witnesses to their own situation. They must also be able to organise around an alternative objective and seek authority to pursue it, subject to the rights of others. For those of us doing the technical work, that requires enduring a particular frustration: we may know how to achieve an outcome and still be required to help pursue another, because the people living with the consequences have legitimate purposes of their own.

⁂

The humanities belong inside a technical education, where the problem is being defined. A specification can be concise because its underlying domain has already been formalized. Students ought to learn to examine the decisions a formalization incorporates, including the choice that makes one person’s work productive and another’s invisible. We should expect them to trace an institution’s history far enough to recognize decisions that its present arrangements make look inevitable. Otherwise a person can become extraordinarily good at carrying out a decision without having learned to recognize it as one.

We ought to teach students to follow an account of a life that unsettles their categories, and to say where it requires a revision rather than translate it immediately into something they already understand. Reading has accomplished little if the reader emerges better equipped to patronize the person whose life he has just encountered. The encounter should make it harder to treat a person as adequately understood because we can predict what they will do.

There will still be binding decisions, and people who lose the argument will have to live with them. Winning does not release us from responsibility for the burdens we impose. An accurate forecast of someone else’s hardship is especially poor grounds for innocence: it establishes that we knew what we were asking them to endure. That knowledge ought to make the decision harder to defend. An education that helps us explain the hardship while leaving us indifferent to whether it should be imposed has made us more articulate without making us more responsible.

A professional generation can keep improving its answers while losing touch with the people whose lives were supposed to give the work its purpose. We can become more accomplished within our representations and less able to recognize what would make any of that accomplishment worth having. The terrible possibility is that we experience this estrangement as progress. We were supposed to be building a world people could live in. In Don McLean’s American Pie, I hear “a generation lost in space,” and the competence no longer sounds like much of a defense.

I once treated disagreement as a need for better economic literacy. The people whose judgment I discounted have no obligation to reward my discovery of their existence. They were trying to tell me how they lived, and I was explaining why the system worked. Nothing I knew required that refusal to learn. I want to be capable of hearing what they are asking for, and of putting what I know to work without first making them concede that they ought to be content.

---

Go Deeper into the Circus

---