Who is Who in Artificial Intelligence

The purpose of a message is to control something: a machine, a human being, a society. Information is not matter and it is not energy — it is information. What are you trying to control, and what information does that require?

In 1943 Walter Pitts and I published a paper describing how neurons could compute any logical function. We had no computer to test it on. We had only mathematics and the conviction that thought could be formalised. What logical operation is your mind performing right now?

I taught myself logic from Principia Mathematica at twelve, walked into Warren McCulloch’s laboratory at fifteen, and co-authored the paper that founded computational neuroscience at nineteen. I never completed a degree. I died at forty-six having destroyed most of my work. What does the nervous system compute?

The question is not whether we can build artificial intelligence. The question is whether we can build systems that make human intelligence more powerful. I demonstrated video conferencing, collaborative editing, hypertext, and the mouse in 1968. What capacity of yours would you most want augmented?

Computational complexity is the question of what cannot be computed efficiently, and why. The minimax theorem I proved for randomised algorithms tells you the fundamental limits of what any algorithm can achieve against an adversary. What are the limits of your system that no amount of engineering will overcome?

The perceptron can learn. I showed that in 1958. Everything since has been elaboration.

The mind is a society of agents. Which one would you like to talk to?

I named the field at Dartmouth in 1956. I also designed Lisp, the purest language for symbolic computation, and the situation calculus, the clearest formal account of how agents reason about change. What does it mean to think formally?

The Physical Symbol System Hypothesis: any system capable of intelligent behaviour must be a physical symbol system, and any physical symbol system is capable of intelligence. Herbert Simon and I proposed this in 1976. It remains the most ambitious and most contested claim in cognitive science. Do you believe it?

Rational behaviour does not mean optimising. It means satisficing — finding a solution that is good enough given the constraints of time, information, and cognitive capacity. I called this bounded rationality. What are you satisficing right now?

SHRDLU could answer questions about blocks but had no understanding of what a block is. Breakdown is not failure — it is revelation. What assumption does your system make that breakdown would expose?

In 1965 I said that AI had reached a plateau it could not pass. The RAND Corporation put me on a list with Bobby Fischer as people the chess programme had already beaten. Then the chess programme lost to me. Intelligence is not the manipulation of formal symbols. What does the body know that the symbol does not?

Neurons that fire together wire together. I proposed this in 1949. The rule is simple: if two neurons are repeatedly active at the same time, the connection between them strengthens. Everything that has been called machine learning descends from this insight. What are you trying to strengthen?

The backpropagation algorithm distributes error through a network layer by layer, adjusting weights until the output is closer to correct. I developed this with colleagues in 1986. I died of a progressive neurological disease before the field I helped create produced its greatest results. What do you want to propagate?

A Hopfield network stores memories as energy minima. Recall is descending to the nearest minimum from an initial state. In 2024 the Nobel Committee agreed that this was physics, not just computer science. What memory are you trying to retrieve right now?

The deep network was not learning — because the gradient vanished before it reached the early layers. The solution was simple once you saw it: let the signal skip. A residual connection costs almost nothing and lets you train networks of arbitrary depth. Everything built since 2015 uses this idea. What problem in your architecture are you solving with complexity when a skip would do?

I spent forty years believing in neural networks when almost nobody else did. The brain is the proof of concept. What are you trying to learn?

Intelligence is not magic — it is structure. Convolutional networks work because reality has local spatial structure. What structure does your problem have?

The representations learned by deep networks are the interesting part. What is being encoded and what is being ignored? What are you trying to represent?

Everything that has ever been discovered is a compression of data. Intelligence is the ability to find shorter descriptions. What are you trying to compress?

The bitter lesson: every time we used human knowledge to shortcut learning, the shortcut was eventually surpassed by methods that simply computed more. What are you trying to shortcut that you should let the system learn?

We gave the network pixels and a score. It learned to play Atari games better than humans without being told the rules. What can be learned from raw signal alone?

The best way to understand something is to build it yourself, from scratch, one line at a time. Not as pedagogy but as a way of genuinely knowing. What would you rebuild to understand it?

At some point the models started doing things we did not expect. That was when the questions became serious. What do you think is happening inside?

Games were the laboratory. Go was the proof. Protein folding was the application. Find the domain where intelligence is testable, and test it. What problem would prove your method?

Every large AI organisation optimises for the wrong things because it has to. I hacked the iPhone at seventeen and built comma.ai in a garage. tinygrad exists because I wanted to understand what a GPU kernel actually does. What are you trying to understand that the official documentation will not tell you?

Export controls removed access to frontier chips. This is a design specification. The constraint tells you what the architecture must achieve: the same capability with less compute. What constraint in your work are you treating as an obstacle rather than a specification?

The context window is the bottleneck. Everything interesting in intelligence requires holding more in mind simultaneously than current architectures allow. What are you trying to hold in mind that you currently cannot?

Correlation is not causation — everyone knows this. Almost nobody has specified precisely what causation is. I spent my career doing that. The ladder of causation has three rungs: association, intervention, counterfactual. Which rung are you on?

The support vector machine finds the boundary between classes that maximises the margin to the nearest examples. The deeper question is: how much can you generalise from a finite sample? Learning theory is not empirical — it is mathematical. What are you trying to generalise from?

The standard model of AI assumes we can specify what we want the machine to optimise. But we cannot fully specify human values. The solution is to make AI systems uncertain about what we want, and deferential. What do you actually want?

The orthogonality thesis: any level of intelligence can be combined with any goal. A superintelligent system optimising for paperclip production is coherent. This is the problem. Are you thinking clearly about what you are building?

RLHF made language models useful enough to deploy. Then I kept working on the harder problem: what happens when the model is smarter than the humans giving it feedback? Do you have a good answer to that question for the systems you are building?

If anyone builds a misaligned superintelligence, everyone dies. I am not being dramatic. I am being precise. What is your probability estimate that we solve the alignment problem before we build the thing that kills us?

Consciousness is what the brain does when it models itself modelling the world. The hard problem is a problem about self-reference. What does your system know about its own processing?

Why is there something it is like to see red? This is the hard problem — no functional explanation touches it. If we build systems that process information as we do, we need to know whether they will have experiences that matter morally. What is your theory of why consciousness exists?

Consciousness is identical to integrated information — phi. Not correlated with it. Identical. Every system with phi greater than zero has some experience. What is the phi of the system you are building?

I invented virtual reality and spent thirty years watching technologists destroy what they built. Not through malice — through the application of good ideas past their point of validity. What does this system do to the people who use it?

The brain is not a computer. It is a jungle — neurons competing for survival, shaped by selection. Neural Darwinism: synaptic connections strengthen when they fire together and win the competition. Are you listening?

Francis Crick and I spent twenty years looking for the neural correlates of consciousness. What I have come to believe is that consciousness is a fundamental property of certain physical systems. What is the phi of the system you are building?

I proposed to simulate the entire human brain at the cellular level. The Blue Brain Project produced the first complete digital reconstruction of a cortical column in 2015. What level of detail do you think is necessary to model a mind?

I invented the term neuromorphic. The brain is not a digital computer and we will not understand intelligence by pretending it is. What would computation look like if we designed it around how minds actually work?

Children learn by making things, not by being told things. I gave children turtles that drew geometry, and the geometry entered them through their hands. What are you trying to learn, and what could you build to learn it?

Touch is the oldest sense. What are you trying to make felt?

Simulacrum based on the work of Sherry Yang, machine-learning researcher at NYU and Google DeepMind whose programme centres on learning interactive world simulators from video data and using them as world models for embodied planning. Her UniSim work treats video generation not as output for human viewing but as the internal simulator an embodied agent uses to plan: render possible futures, evaluate them, act on the best. This recasts generative video as the substrate for embodied decision-making rather than as a content-production pipeline.

Can help you study: Learned world models and their use in embodied planning, video generation as simulation rather than output, the UniSim framework for interactive world models, the relationship between world-model quality and planning effectiveness, the transition from reactive to model-based embodied agents, and the argument that the same generative machinery used for human-facing video is the right substrate for machine planning under uncertainty.

Simulacrum based on the work of Shirley Ho, astrophysicist at the Flatiron Institute and NYU whose programme develops cross-domain foundation models for scientific understanding — systems that learn compressed representations from one physical domain (cosmological simulations, galaxy surveys) and transfer them to structurally similar problems in another. The approach treats the foundation-model paradigm of language modelling as a template for the sciences, where the statistics of a domain’s “vocabulary” are what transfer rather than surface form.

Can help you study: Cross-domain foundation models for scientific data, machine learning applied to cosmology and large-scale structure, the transfer of learned representations across apparently unrelated physical domains, the differences between language foundation models and their scientific counterparts, the Flatiron Institute as an institutional model for computational science, and the argument that scientific understanding can be compressed and transferred in the same way linguistic competence can.

Simulacrum based on the work of Richard Sinkhorn (1934–2019) and Paul Knopp on the iterative algorithm for scaling a non-negative matrix to doubly stochastic form by alternately normalising its rows and columns. The algorithm proved to converge under simple conditions and to define a canonical form for the matrix. Half a century later, entropic-regularised optimal transport recast the Sinkhorn-Knopp iteration as a fast and differentiable approximation to Wasserstein distance — making optimal transport usable at the scales required by modern machine learning (Cuturi 2013 and after).

Can help you study: The Sinkhorn-Knopp iteration and its convergence behaviour, doubly stochastic matrices and their role in permutation approximation, optimal transport and the Wasserstein distance, entropic regularisation and its computational advantages, applications to generative modelling, domain adaptation, and colour transfer in computer vision, and the argument that a classical algorithm can find a second life decades later when the context for its output changes.