After the Transformer

The last essay ended with a question: what would a genuinely different architecture need to do?

Before answering, let us deal with the counter-argument already forming. Because if you have been paying attention to AI over the last two years, the premise may seem already settled. AI solved 80-year-old mathematics conjectures. AI found 31 new planets in NASA data human astronomers had been staring at for years. AI is writing music. AI is generating drug candidates. The counterexample factory is running at full capacity.

So let us take each seriously. Because what the evidence actually shows, when you look at it carefully rather than at the headlines, is not a refutation. It is a precise illustration of the thesis.

When a Library Search Is Called Discovery

In early 2026, a large language model helped crack two long-open Erdős conjectures. Here is what Melanie Matchett Wood, a Harvard mathematician asked by OpenAI to review the result, actually said: "Maybe people should be spending more time playing devil's advocate." The AI's solution was a straightforward approach that no human had ever attempted, despite the fact that the tools had already existed for years. The humans were blocked not by inability but by the belief that the conjecture was probably true. The AI, which has no beliefs about the world, searched more directions without prejudice. Scientific American was careful: "Genuinely new, groundbreaking ideas remain beyond the reach of current LLMs, instead leaving the machines to mine the literature for rare gems where humans missed a relatively simple approach."

The IMO 2025 Problem 6 is the counterweight. Solved correctly by six of approximately 600 human contestants. Every AI system entered scored zero. Not approximately zero. Zero. The problems requiring genuinely new conceptual machinery remain entirely untouched.

A 2025 benchmark study from HKUST confirmed the structural fact: same operation, same architecture, 10 percent gain in mathematics, 0.14 percent in physics. Richard Feynman: "Mathematics is not real, but it feels real. Physics is real, but it does not feel real." The gap between those two sentences is exactly the gap the benchmark is measuring. Mathematics is a formal language where pattern and substance are the same thing. Physics is reasoning about the world through mathematics. Those are different activities, and the difference turns out to matter enormously.

When Pattern Recognition Is Called Discovery

In 2025 and 2026, AI systems including RAVEN at the University of Warwick confirmed over 100 new exoplanets in NASA TESS data. The project lead described the method precisely: "We trained machine learning models to identify patterns in the data." The data had already been analysed by human astronomers. The AI found signals by applying pattern recognition at a scale no human team could manage. Real and valuable. Not the same as Adams and Le Verrier independently predicting Neptune in 1845 from anomalies in Uranus's orbit, months before anyone looked for it. One is hypothesis formation from nothing but anomaly and mathematical intuition. The other is classification of known patterns in large data. Current AI does the second extraordinarily well. It has not demonstrated the capacity to form the hypothesis that made the data worth collecting in the first place.

When Statistical Assembly Is Called Composition

Carnegie Mellon published research in January 2026 finding that AI-assisted music was judged by listeners as significantly less creative than human compositions. A study in the Journal of the Acoustical Society of America played identical recordings to two groups, labelling one as AI-made and the other as human-composed. The "AI" versions were rated as significantly less emotionally moving despite being acoustically identical. A 2025 study measuring physiological responses concluded that "while AI can mimic human creativity, its strong performance in creative tasks is likely driven by non-creative mechanisms rather than genuine creative thinking," and found that AI has meaningful difficulty identifying which of its own outputs are actually the most creative.

We do not just hear music. We hear the mind behind it. When Miles Davis played a note, it was inseparable from the life that produced it. The composer Arvo Pärt said of his music: "Behind all this is an experience which I cannot yet explain." The experience is the thing. The note is the last step of a process whose earlier steps happen somewhere that is not linguistic, not computational, not predictive.

Where the Money Is Going, and What It Is Not Solving

Companies like Lila Sciences, backed by Flagship Pioneering with around $550 million, Periodic Labs, backed by Andreessen Horowitz with around $300 million, and FutureHouse, backed by Eric Schmidt, are building something worth placing precisely. Their architecture: AI proposes candidate molecules or experimental approaches, robotic laboratories run the experiments, physical reality verifies. This is the right design for hypothesis testing where the hypothesis space is known and the bottleneck is experimental throughput. Drug candidate screening. Materials property prediction. Protein structure verification. These companies will change the speed of science in those domains materially.

But notice what all of them require as input: a hypothesis. Someone or something has to decide which molecule is worth testing, which experimental perturbation is conceptually meaningful. The robotic lab tests a hypothesis. It does not generate one. The 20-year delay between Shannon's information theory and the genetic code was not a throughput problem. Shannon's insight was not blocked by the speed of experiments. It was blocked by the absence of anyone who saw that the two things had the same formal structure. Lila and Periodic Labs solve throughput. The recognition gap is untouched.

What Serious Researchers Are Actually Building

Neurosymbolic AI (Gary Marcus, Scallop at CMU, DSPy at Stanford): keep the neural front-end for perception, add a symbolic back-end for auditable reasoning with guarantees. Marcus demonstrated that pure neural systems fail simple block-stacking puzzles that 1970s classical planners solved trivially, and that performance collapses further when block names are obfuscated, proving the apparent reasoning was pattern recall dressed as deliberation. Where it runs out: if the extraction from language to symbolic structure is noisy, formal reasoning propagates the noise with official-looking authority.

Verifier-centric systems (AlphaProof, LEAP at DeepMind): put a general model inside a harness that checks outputs against something external and non-negotiable. The 2026 LEAP paper showed a stock language model with no additional training, inside an agentic harness with a Lean formal verifier, beat a purpose-trained specialist. Solve rate on formal mathematics went from under 10 percent to 70 percent. The harness outperformed the specialisation. Where it runs out: Lean can check a proof step. It cannot check whether a physics intuition is physically meaningful, or whether a cross-domain structural insight is real.

World models (LeCun, JEPA, Large Concept Models at Meta FAIR): stop predicting tokens, predict in abstract representation space. Meta's Large Concept Model (December 2024) encodes each sentence into a concept vector spanning 200 languages, reasons in concept space, decodes to words only at output. The same idea in any language maps to the same point. Where it runs out: building a concept space where genuinely different disciplinary frameworks converge requires training signal asserting they should. That signal does not exist at scale.

Causal representation learning (Schölkopf at Max Planck, Bareinboim at Columbia): beneath the surface variability of observations across different domains, latent causal variables remain invariant. Find what stays constant when the environment changes and you have found something real. The identifiability theorems show that under specified conditions, latent structure can be provably recovered from data. Where it runs out: those conditions are hard to satisfy in scientific corpora not collected as controlled experiments.

Hypothesis generation via GFlowNets (Bengio at Mila, LawZero): imitation is the wrong objective. A model trained to produce the most probable continuation of human text learns what humans have said, not what is true. GFlowNets sample hypotheses with probability proportional to evidential support rather than training frequency, maintaining honest distributions over competing explanations. Where it runs out: scoring hypotheses requires a reward signal, and the ultimate reward signal in science is an experiment.

Structure mapping (Gentner and Forbus, the analogy research lineage): analogy is not similarity. In her landmark 1983 paper, Dedre Gentner at Northwestern established that genuine analogy is a structure-preserving correspondence based on shared relational form, explicitly ignoring surface features. The systematicity principle scores correspondences by whether they preserve connected systems of relationships over isolated facts. Where it runs out: it requires structured relational representations as input, which means the neural-to-symbolic extraction problem is always the bottleneck.

Every approach hits the same wall from a different direction. To go beyond surface pattern-matching, knowledge must be represented in a form where structure is explicit and correspondence is checkable. Representing scientific knowledge that way, reliably, from the messy natural language in which it lives, is the hard problem at the centre of all of these research programs. The wall is not failure. It is a precise identification of what the next step actually requires.

The Golden Age for a Different Kind of Mind

Intellect has become a commodity. Jensen Huang said it plainly, and he is right in a way that goes deeper than most commentary has acknowledged. The analytical, information-processing, pattern-matching capabilities that formal education spent three centuries training are now available at superhuman scale and near-zero cost to anyone with an internet connection. The advantages of knowing more, synthesising faster, and calculating with fewer errors are gone. Not diminished. Gone.

The people writing that "follow your passion is bad advice" and that the rational move is to augment yourself aggressively with AI tools are not wrong in the short term. But they are missing the longer arc. The analytical capacity being augmented is the capacity being commoditised. Becoming very good at using a commodity tool makes you efficient. It does not make you irreplaceable.

Rex Jung's research shows that highly creative individuals display reduced prefrontal cortex activity during creative work: the executive control region has to quiet for creative intelligence to operate, not intensify. Marcus Raichle's work established that the default mode network, active during mind-wandering, is not idle noise but a metabolically expensive integrative process that analytical attention actively suppresses. Csikszentmihalyi's flow research found that peak creative performance is characterised by the suspension of self-referential analytical thought. The things AI is automating are exactly the things that suppress this capacity.

The golden age belongs to minds that have always been hardest to institutionalise: the first-principles thinker who cannot stop questioning what everyone takes as given, the scientist whose best work arrives from a felt sense that the current framework is wrong, the artist close enough to their own inner experience to channel something that exceeds what they can articulate. These people have always existed at the margins of institutions designed to reward analytical throughput. That marginalisation was always a mistake. In the age of AI, it becomes a civilisational one.

The Gap Nobody Has Framed as a Problem

Consider what actually happened in the discoveries that changed the shape of knowledge. Shannon did not run more experiments. He noticed that a communication system and a biological system had the same formal structure, and twenty years of delay evaporated in a single recognition. Boltzmann's entropy principle sat fully formed for a century before Friston saw that it described neural function. Lotka and Volterra wrote a differential equation for predator-prey dynamics in ecology, and the same equation described market oscillations in economics, and nobody working in economics noticed for decades because the vocabularies did not overlap.

In every case, the bottleneck was not the verification. Once Shannon saw the connection, the mathematics followed. Once Friston formalised the free energy principle, the predictions were testable. The gap was in the recognition itself: the moment before the formalism exists, before the hypothesis is stated clearly enough to be tested, when a mind that has gone deep enough into one domain suddenly perceives that another domain has been solving the same problem from the other side.

That recognition arrives not to the most analytically capable people in a field but to the people who have gone deep enough that they have moved past its vocabulary. Who have sat with the fundamental structure of a problem long enough to perceive it rather than describe it. Who can hold two bodies of knowledge simultaneously, not as two sets of propositions, but as two felt shapes, and notice when the shapes are the same.

Kekulé did not calculate benzene's ring structure. He saw it in a dream. Ramanujan received theorems whole from somewhere he could not name. Poincaré had the solution arrive the moment he stepped onto a bus in Coutances, having thought about something else entirely. These are not romantic stories about genius. They are data about a process that operates below the threshold of deliberate verbal reasoning and produces outputs the deliberate mind then transcribes and verifies.

The scientific institutions we have built are extraordinarily good at the second half of discovery: verify, publish, fund, replicate. The first half, the pre-verbal, cross-domain, structurally-oriented recognition that produces the hypothesis worth testing, has no institutional home, no curriculum, and no tool. In the age of AI, when verification, experiment, synthesis, and data analysis are all being automated at speed, the first half becomes the exposed bottleneck. Not because it was not always the bottleneck, but because everything around it is now being removed, and the gap is finally visible precisely because it is the one thing left standing.

Why I Am Attempting to Build Consilience

This is the gap I ran into directly, which is not the same as reading about it. I began building a research tool for scientists working across disciplines. The premise was retrieval: if a physicist's framework and a contemplative tradition's account of the same phenomenon were pointing at the same underlying structure, they should surface together. I built the retrieval system. It worked by similarity, by the nearness of meaning in an embedding space. And it worked, in the way similarity always works, which is to say partially, and in a way that made its own limitation impossible to ignore.

Similarity is not the relation that scientific discovery is made of. Shannon's information theory and the genetic code are not similar. They are structurally identical under a mapping, and the vocabulary walls between them are genuine and almost total. A similarity-based system finds what is linguistically near. It misses what is structurally the same but lexically alien. This is not a retrieval tuning problem. It is the wrong relation, computed at very large scale.

The deeper I went into the architecture question, the more the frontier pointed in one direction: not larger models or more training data, but a different kind of representation where the structure of knowledge is explicit and correspondences between structures are detectable and, where possible, verifiable. This is the direction all the research programs surveyed above are pointing toward. Each hits the same wall from a different angle.

I am still researching what the right foundational design is. The verification side has partial answers that will only improve. AlphaProof and LEAP handle formal mathematics. The empirical sciences have data to check against. Increasingly capable reasoning systems will extend those tools further. None of that touches the earlier moment: the recognition that a hypothesis is worth forming, the connection across traditions that no existing vocabulary names, the structural correspondence a researcher perceives before they can prove it.

The goal of Consilience is to build infrastructure for that moment. Not to generate hypotheses algorithmically, but to give the researcher who already perceives something, whose recognition outruns their vocabulary, the means to find what else in the corpus of human knowledge points at the same structure. To serve the researchers whose most important insights arrive in a form no existing tool is built to receive: the ones who know something before they can say it, who are working in the territory that all these architectural approaches are circling, and none have yet reached.

The foundational architecture is still an open question. That honesty is the most accurate thing I can say about where the work stands. The gap is real. The right approach to it is not yet fully clear. But that it exists, that it matters, and that almost nobody has framed it as a problem worth solving: those things I am certain of.