Sparks and Parrots

The last time I was in San Francisco was November 2024, a week after the election. I stumbled out of SFO into a second morning; the flight from London had swallowed the day, placed its finger on the clock and held it stuck. In the Uber to the hotel I chatted to the driver, a young Yemeni-American who had just quit his job as a trucker. He talked about the struggle of gas prices and groceries, and told me about sleeping in the parking lot of his gym because driving back to LA, where he lived, took too long. As he spoke we took a turn off the freeway and the intangible was added to the tangible. A billboard flew past: "Welcome to AI Country".

That night, in one of those half sleeps that was more exhaustion than tiredness, I dreamt about the City Lights book store on San Francisco's Columbus Avenue. I remember somewhere in that dream looking up at its roof and seeing a macaw, its feathers red and gold, reciting Howl by Allen Ginsberg.

Somehow my mind had scrambled two lines: the sparks of AGI and the stochastic parrot.

In The Map and the Landscape I described the problem of intelligence: how using the word blurs our capacity to disentangle what AI technology does from what we think it is. Reduced to an intelligence scale people coalesce around its poles. Either the models show 'sparks of AGI' (artificial general intelligence) on the straight line to 'superintelligence', or they are 'stochastic parrots', predictive systems dissimulating us by simply predicting the next token.

Each side claims victories when new models demonstrate unexpected capacities or when they fail in spectacular or seemingly stupid ways.

I saw the best minds…

In this essay I'd like to depart from either of these framings. The work and the experiments done inside Distributional Surface Theory are not to deny or minimise the power of this technology, in fact the opposite. They are trying to define more precisely what it is as a technology and what it is not as a mind; in a sense to remove the question of intelligence entirely.

If the DST evidence holds up, we can understand it in the context of the first essay's central metaphor: the map and the landscape.

The metaphor says: probabilistic AI systems navigate the surface of a map, built from the training data, sequentially; humans also navigate probabilistically and sequentially, but they possess a different kind of cognitive capacity which comes from the experience of what generated the training data: this is an ability to hold different aspects of this landscape in superposition, synthesising ideas and concepts simultaneously to generate novel combinations.

As we've seen in DST the metaphor is capturing something about the differences in 'cognition' between humans and AI systems. But can we now make this metaphor more concrete and use it predictively?

Yes: by describing each of its halves, the map navigated sequentially, and the landscape navigated by superposition, and breaking these two down in receptive and active experiences:

What differentiates these two pairs is how they operate in time.

The process by which we hold multiple things simultaneously.

An experience includes much more than just the information that transmits it. In the first essay I described this as 'decompression', the way we restore the three-dimensional landscape of experience from the flat surface of data. When humans generate ideas, this often involves an act of superposition: holding different concepts together to synthesise something new. Ideas can emerge 'complete' and then can be explored or explained.

The process of moving through information in sequence.

Parsing is attention to a sequence of information. Moving your eyes across this sentence, word by word is the act of parsing. Production is the sequencing of information out, words placed one after another for example, or code written line by line.

These two substrates, in humans, operate in a kind of interference, phasing between one and the other; they feed and mingle. Parsing builds experience, which in turn is what enables generation, sequencing is how we take that generated idea and make it legible for another person to parse, and then experience, and so on.

These two modalities are defined by different states of time: superposition is dimensional, a thick present; sequence is flat and linear. If we want to communicate an idea, its shape, it must be collapsed down into linear time.

Though these might just seem like shorthands they point to something more. In the triptych's final essay, Beethoven and the Bitter Lesson, we'll see how studies showing quantum cognition effects in humans are not reproducible with LLMs. This is not to say though that what characterises substrate-1 is literally a 'quantum' process.

With these two substrates defined more clearly, and the evidence inside DST itself, we can characterise better what modern AI systems are doing.

Current AI systems are a technology navigating a compressed statistical surface distributionally. They parse and produce information without access to the original structure that created the training data. Because the statistical relationships in the data were generated from the structure, the sequenced output is sometimes indistinguishable from the superposed process which created it.

DST has been about taking this definition and trying to stress test it; and, mapped against the pairs it suggests something simple. Current AI systems have extraordinary capabilities within, and are bound by, the bottom half of these pairs, the second substrate.

The training process, a prompt, any information you throw at a model can be parsed at a super-human speed. AI systems are also able to produce much more quickly and efficiently (at least in some cases) the sequenced outputs. Setting aside the question of quality and 'understanding', we can make this vivid: an LLM could read the whole of Shakespeare in roughly 5-10 minutes, and write a 10,000 essay about it in about 2. It takes a human about 60 hours to read, and let's say 30 hours to write. A speed differential on the order of five hundred to a thousand times.

But the failure modes that DST predicts are precisely those that are solved by the capacity for superposition: novel rules that contain in them structural similarities to patterns already learned, where the superimposing those structures would enable solution; an understanding of something which is structurally impossible; a temporal understanding of 'what will happen', the ability to look ahead. Indeed, I would say for a 'world model' to exist, something more comparable to human cognition, the interference between these two patterns - superposition and sequence - is a prerequisite.

So, instead of worrying about sparks and parrots, perhaps we can reframe: we are creating a technology that hyper-scales the parsing and production of information; a strand of human cognition, not intelligence.

That hyper-scailing will, of course, change how both pairs operate.

The idea that to 'solve' problems, or to change the world, AI systems have to be 'intelligent' seems to me reductive. Why should it matter? Any sufficiently powerful technology reshapes society, independent of whether it has a 'world model' or internal states. In our fixation to decide whether the mind is a computer, or something different, we miss this.

If we think of AI as the hyper-scailing of substrate-2, piggybacking on all the statistical patterns locked into the vast amount of information we have already produced, we can set aside the intelligence frame and look more clearly.

First, we should celebrate this as one of, if not the most, extraordinary technical achievements in history. The total parsing of all the available data generated from human knowledge, as well as the production of more data out of the statistical regularities present in the data, will enable us to solve a lot of problems.

Indeed, we have no idea how many of our open questions have answers already in what we've formulated, discoverable just from running a deep enough computational process.

The reframe is not to worry so much whether these predictive technologies are conscious or intelligent, but instead if, trained on sufficient data, they may be capable of solving the problems we set up for them with a speed and efficiency we cannot match.

This distinction is, I think, the important one.

Before we talk about separating the kind of problems, and work, that AI systems will accelerate, let's start by making visual the substrate distinctions between human and machine. Below you can drag the slider across and see how time spent alters ability across our 2x2 substrate grid.

The premise of DST is that we do not yet have evidence that current AI systems migrate to substrate-1 simply by developing more powerful substrate-2 capacities. But what we do have is evidence that a sufficiently powerful AI system can sequentially navigate statistically related information produced by substrate-1 much more efficiently than humans, as long as it has a goal.

A substrate-2 system, to resolve a problem, needs an articulated criteria against which candidate solutions can be scored. In the first essay we met AlphaGo, the DeepMind system that beat one of the world's top Go players. Systems like AlphaGo have an explicit articulated criteria (in its case a value function that estimates the probability of winning from any board state). Different AI systems use different articulated criteria, but they all have them.

So we can now use these three variables to help us formulate how complex a problem is for an AI system, and what class it can fall into:

Together, these three sort problems into the four categories below defined by: how much substrate-1 work the problem requires, where on the spectrum its criterion sits, and how much substrate-2 demand it generates.

Problems where the conceptual, substrate-1 work has already been done by humans, and the problem has a clearly defined criteria. In science these are problems bottlenecked by substrate-2 work, the raw sequential processing needed to reach a goal. Using AI systems to run through category 1 problems is the obvious space for scientific acceleration.

AlphaFold is the canonical example of AI helping us to solve an intractable category 1 problem. The fundamental conceptual breakthrough, the substrate-1 work, was done by Anfinsen when he discovered that the shape of a protein was determined by its amino acid sequence. From this came the articulated criteria. Fifty years of crystallography work produced the data. But only a hyper-scailed substrate-2 system, AphaFold, could have made 200,000,000 protein structure predictions in months.

Problems where enough substrate-1 work has already been done to make the direction of search articulable, but not enough for the problem to be handed over completely.

A human still has to remain inside the process: setting goals, judging partial results, changing direction. The AI system supplies substrate-2 acceleration: generating candidates, testing variants, writing code, searching data and surfacing patterns at a speed no human could match.

In 2025 Ernest Ryu worked with GPT-5 on a long-standing problem in optimisation theory. The goal was articulable: prove why Nesterov's accelerated gradient method remained stable. But the problem was too complex to be handed off to an arbitrary verifier. GPT-5 generated approaches, many of them wrong; Ryu applied the evaluative criteria, judging a good direction from a bad one, adjusting and testing.

It describes solving category 2 problems beautifully:

"[The] turning point emerged when the model proposed a way of restructuring the equations that govern the NAG method. The suggestion was not correct as written, but Ryu recognized the structural feature that seemed meaningful, developed it more rigorously on his own, and began asking the model targeted questions to test the potential proof's viability. That line of inquiry ultimately became the backbone of the final proof he wrote out himself.[1]"

Problems where the components of the answer already exist, but not in the same place. These problems sit outside an obviously defined frame; but superposing two or more ideas, conception or domains together can solve them. Once they are solved that underlying connection may seem obvious, but before there is no clear line between one domain and another.

CRISPR is a category 3 problem. In 2012 Jennifer Doudna and Emmanuelle Charpentier showed that a bacterial immune mechanism could be re-engineered into programmable DNA-cutting machinery. The substrate 1 move, the act of superposition, was realising that a system bacteria used to defend themselves against viruses could become a general tool for editing genomes.

This human judgement that one disparate thing can resolve another desperate thing still seems elusive to AI systems. The strongest examples of progress right now are in categories 1 and 2.

Problems where the work done by substrate-1, the superposition of ideas or concepts, creates the criteria by which the work is judged. The articulable criteria does not exist before the work itself, the work makes it.

General relativity is such a case. A theory in which gravity was not a force at all, but the bending of space and time itself around objects with mass, was a reorganisation of the search space of physics itself. The major breakthroughs in human knowledge sometimes have this shape. In their creation they alter knowledge itself. This is what separates category 4 from 3.

This feels like the truest and furthest frontier, what we associate in society with acts of genius.

For all of the examples above I have chosen scientific advancement as the test case, but this set of categories is applicable not only to science but to other domains, as well as white and blue collar work.

Breaking problems into these sets is an attempt to try to articulate something gestured at and described by AI labs, founders and people working with the tools. What's missing is a kind of specific vocabulary. The moment we're living in is one of uncertainty not just because the advent of this technology provokes it, but because what that means for the future is not settled.

Dario Amodei described the uncertainty of whether models could move up to category 3 or 4 work in a conversation with Dwarkesh Patel:

"My one little bit of fundamental uncertainty, even on long timescales, is about tasks that aren't verifiable: planning a mission to Mars; doing some fundamental scientific discovery like CRISPR; writing a novel. It's hard to verify those tasks…The world in which we don't get there is the world in which we do all the verifiable things. Many of them generalize, but we don't fully get there. We don't fully color in the other side of the box."

Or as the framework might describe it, the scaling of substrate-2 may not lead to capacities in substrate-1.

But it's important not to be trapped in an absolutist definition: the hubris of people who have doubted the limits of computation systems has been badly bruised over the past century. Hassabis, who has described himself a 'Turing's champion', has produced work at DeepMind that should give anyone pause before committing to a position of 'impossibility' for AI systems on category 3 and 4 problems.

If a sufficiently sophisticated AI system could test and generate and endless set of potential criteria, and that criteria generation was available inside the training data, then category 3 problems should be in reach. And Hassabis has posited a thought experiment about category 4 problems, arguing that a test for AGI might be to drop a system into 1911 and see if it could discover general relativity.

So the aggressive case for the computational maximalism says: perhaps the resolution of category 3 problems is just an illusion. By priming the search effectively enough, we can run through so many possibilities that we will uncover the relationships that we thought an act of superposition required. I myself would not bet against this. If we've learned anything in the past 4 years it's been that our own suppositions about the richness of the statistical regularities 'encoded' have been way off. It's not impossible in principle that we will find a second, new frontier.

But this is precisely why it is a frontier and not a conclusion. The evidence for these category 3 problems falling is thin; we can't adjudicate right now but we'll probably find out soon. Category 4 problems are a further frontier still and another essay, 'The Undiscovered Country', tries to broaden the frame to include not just science but the arts, and show how wide the frame of computational maximalism has to stretch to hold it.

But we can think now about how the shape of knowledge itself will change. As Terrence Tao put it: "I do believe that hybrid human plus AIs will dominate mathematics for a [long time]"

With this hyper-scailed substrate-2 capacity we will probably run through a certain set of problems at enormous speed. There may be problems we've previously thought intractable that will be, like AlphaFold, completed in a matter of months. But if AI systems don't have a clear path to mimicking substrate-1 work this will be bottlenecked by the process of ideation, evaluation and criteria definition. Termed loosely, creative work.

And in the case where this happens perhaps the problems that we will be faced with will slowly migrate up the abstraction curve. Conceptual thinking, the core of substrate-1 work, will become more and more valuable and the cost of the mechanical work of parsing and sequencing ideas dropping essentially to zero.

Two things follow from this. First, as we break down the tractable problems we may discover that the class of complex questions grows. There are simply more category 3 and 4 problems that we knew about, because the tractable ones were 'masking' more fundamental things we had not yet reckoned with. Second, the whole system of education and academia that we have constructed is set up to train and enable the laborious and necessary substrate-2 work that complements substrate-1. In this case we may have to rethink how both of these are structured.

So it is not that knowledge is complete, in fact it may be that knowledge becomes less complete as the problems that we have remaining are harder.

This move, the gradual shifting of 'value' into deeper forms of abstract complexity and thinking has a parallel, strangely, in the industrial revolution and may help us to think about how we can map these categories onto the real changes in work and the world that we're about to experience.

As a parallel we could see the industrial revolution as a process of abstracting work out of the body and into the mind. When the physical act of labour can be automated it follows that there must then be 'cognitive management' of that mechanical automation and whatever labour is left over.

But efficiency dictates that even this 'cognitive management' must itself be managed. We think of large swathes of white collar work as, to some degree, attempts to tighten and verify something that seems unmechanical into a mechanism: intelligence.

Sometimes when people discuss the impact of LLMs on the job market they react as if this technology is sui generis, as if it does not have a market logic behind it. I disagree with this.

From the mechanisation of sequential manual processes in the industrial revolution, through to Taylorism and the management culture of the 20th century, we have been dividing work until it reaches the smallest unit of reliable, repeatable, process. In the language of this essay, this is the deliberate flattening of many jobs onto the surface of substrate-2, stripping out more and more of substrate-1 work. We have been deliberately 'mechanising' intellectual labour by making it target based, goal based, and systemically verifiable.

So LLMs are not a technology which has emerged by accident. They are the logical end point of this project. A tool that navigates probabilistic paths across the surface of language faster than humans, for an economy engineered into that shape. The jobs at risk are not just 'low skill', and people will not lose them because AI is suddenly 'more intelligent' than they are. They will lose them because the economy has already engineered a system which makes them automatable.

So now the categories which we applied to science can be recast in terms of work itself.

Category 1 work, where the criteria are explicit, the task is dense is substrate-2 traversal, is an area already affected by the arrival of AI. Data entry, routine customer service help, basic coding, drafting of standard documents against templates. These tasks have been flattened onto substrate-2 before LLMs or any other variant of this technology already arrived. As these systems improve and the criteria becomes tighter more and more of this will disappear.

Category 2 jobs are compressing. I see this in my own company and in others. The human stays in the loop providing iterative direction and feedback, the substrate-1 work. But AI systems are automating the time heavy aspects of these jobs. An experienced developer can automate parts of their coding so they can produce much more code: but still need to guide taste, resolve complex conceptual problems and audit the final output. This work was the apprenticeship for substrate-1. Judgement was built from years of substrate-2 work that may not exist for the next generation.

Category 3 work holds. Cross-domain superposition, the ability to see that one thing answers another, has not been flattened because it could not be flattened. It was always the work that resisted forms of criteria based mechanisation. Strategic thinking, integrative judgement, conceptual synthesis: these survive because no management regime ever found a way to systematise them; at least not on purpose.

Category 4 work is rare in any economy and remains so. A small number of roles produce the conceptual frameworks that organise everyone else's work. These are not at risk; they are also not where most people work.

The pattern that emerges is not "low-skill versus high-skill." It is: how much of this job was already mechanised by the time the technology arrived? Jobs that had been progressively flattened are at acute risk. Jobs that retain substrate-1 work are robust, but the pipeline that produced capable people may be breaking.

A bigger worry for the white collar economy is actually those jobs that appear to require creative thinking and intuition, that we think are category 3, but in reality can be accomplished just as well with clear evaluative criteria for an AI system and enough sequential processing. I don't think we've fully come to terms with how much of the economy lives in this blurred 2.5 category.

The areas of traditional 'white collar' employment that seem most at risk, coding, law, finance, are a blend of substrate-1 and substrate-2 work. But as employees progress up the corporate hierarchy the ratios reverse. The higher up an employee goes the more of their time is taken up with substrate-1 work: strategic and conceptual thinking, cross-domain superposition, judgement under uncertainty.

When the standard discourse says AI will take certain jobs away, it misses the point. The economy required substrate-2 work at scale, and the people employed to do it, through years of immersion, accumulated the domain depth to do substrate-1 work. In most cases this was never designed; we left it to chance and assumed it would happen through exposure.

This assumption is about to break, and what will be hollowed is not just whole categories of work, but the institutional pathways through which people developed the substrate-1 capacities that made them economically valuable.

We do not have a replacement for this process because we have assumed it would continue indefinitely. Our education systems divide knowledge across subject areas, and then relentlessly specialise as students approach adulthood. Almost all our standardised testing measures the capacity to produce probabilistic answers. This is not a failure of the system, it is the system producing people who are ready to move into the flat map of substrate-2 work. That some people improve their capacity to think dimensionally, across domains, has been a function of talent, luck and time.

The challenge posed by the coming changes of AI is this: are we willing to deliberately train the substrate of thinking we have previously left to chance and privilege? If we have, at the extreme frontier of intellectual work in science and mathematics, directional evidence for where this technology is going to take us, how are we going to apply this to a whole society?

Understanding how we help people to build the capacity for conceptual thinking, what is characteristic of substrate-1, and how we inject this into society fast enough to offset the damage done by automation, seems a massive challenge.

But there is precedent for our capacity to make great educational change. For most of human history, literacy was very rare. In 1540 only a tiny number of people could read. Then, over more than four centuries, societies decided that reading was too important to leave to specialists. In Europe, the Reformation demanded that ordinary people read Scripture. Locke argued that everyone had equal capacity for learning. Prussia built the first modern public education system, and the model spread across Europe and North America. By the mid-twentieth century the global literacy rate had passed 40%; today it stands at 87%.

Whatever the process to 're-skill' ourselves up the ladder of abstraction, it's something we need to address right now. We have to believe that developing substrate-1 capacity is teachable and that people deserve to learn it; that it is the responsibility of society to do this.

And this is where a more precise frame about what is happening becomes, I think, important. The obsession with intelligence has stopped us from seeing the shape of work that we made, and the new technology to automate it. Comparisons with the industrial revolution are floated, but more as 'description of the scale of change'. The comparison may be even more direct that people realise.

Our need to make narrative meaning stops us dealing with the present: the futurist questions blind us. The problem of intelligence has been intensified by the language of the AI labs themselves. Discussion of superintelligence may be raising real risks, but it also pulls public attention towards 'speculative minds' and away from the more immediate transformation already happening: the automation of substrate-2 work. Whether AI develops consciousness seems to me an open question, whether it will alter the world seems much more settled.

It can be simultaneously true that AI systems could be as dangerous as nuclear weapons, and that they do not need to be 'intelligent' to be that dangerous; that they can upend and reshape the economy without being 'smarter than us'.

We cannot let the speculative future absorb our responsibility to understand the present.

A few days after I arrived in California I decided to take a Waymo. It was the first time I'd been in a 'self-driving vehicle'. The doors opened, I pressed a panel confirming I was ready to leave, and the wheel turned the car, ghost-smooth, back into the road. The feeling of a miracle, of living Clark's third law: that any sufficiently advanced technology is indistinguishable from magic.

The Waymo stopped at a light heading south towards wherever I was having a meeting. Absorbed in the jet-lag wash of the newness, watching a car drive itself, I hadn't been looking out of the window. Then through the windscreen I noticed someone on one side of the crosswalk, unmoving as the pedestrian signal turned from red to green. A man in tatters, bent over himself; the top half of his body so collapsed that it looked like it had been hung off the bottom of his spine. I learned later people called this the 'fentanyl fold'. The future is here, I thought inside the Waymo, but it is not equally distributed.

When we describe the impending decimation of jobs, the reshaping of the economy, we do so from the fear of loss. We ask why we can't have the economic utopia of the post-war generation, why we have to live through the reconstitution of white collar work. But whether it be in London, San Francisco, or any other major city in the world, for many a decimation already happened, well before the coming of AI.

The question is not whether we can preserve the world as it was, we cannot. The question is whether we can look clearly at this new technology, separate the science of it from the mythology surrounding it, and take responsibility for the society it will help to build.

AI does not have to be conscious to transform the world; and it does not need to be smarter than us to reorder knowledge, work and power.

To make a plan we have first to see the thing itself: a technology build for a world we already made.