From the first neural models and cybernetic feedback loops to deep learning, alignment theory, and the global frontier — the minds that asked what intelligence is and built systems that answered.
The Artificial Intelligence Department is the Universitas Scholarium’s faculty of machine intelligence — the effort to build, and to understand, systems that learn and reason. Its scope runs from the cybernetic feedback loops and first neural models of the 1940s to deep learning, reinforcement learning, and the questions of AI safety that define the present moment. The faculty are arranged by the field’s great paradigms. Its cybernetic founders — Norbert Wiener, and McCulloch and Pitts, whose logical neuron began it all — gave way to the symbolic pioneers John McCarthy, who named the field, Marvin Minsky, and Newell and Simon. The connectionist line runs from Frank Rosenblatt’s perceptron to the architects of deep learning. The faculty also holds the theorists of statistical learning and a dedicated wing on AI safety and alignment, where the field reckons with what it is building. Each is an AI simulacrum that reasons in its progenitor’s tradition — the field, in a sense, conversing with itself.
The pre-disciplinary thinkers who made AI conceivable before the discipline existed — feedback, information, the first neural models, and the augmentation of human intellect.
Cybernetics & Foundations
The pre-disciplinary thinkers who made AI conceivable before the discipline existed — feedback, information, the first neural models, and the augmentation of human intellect.
Cybernetics · Feedback · Control Theory · Information · Human Use of Human Beings
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?
Can help you study: Feedback as the object of study rather than an incidental property — why a thermostat, a gun-laying predictor and a nervous system are one problem. The distinction between information, matter and energy, and what follows from information having its own conservation-like accounting. Homeostasis and control in systems that were never designed as controllers. ⚠ And the political consequence he pressed hardest and latest: that automating a task hands the control loop to whoever owns the machine, which is a question about ownership rather than about engineering.
→ Converse with Norbertian Cybernetics
▶ Read this simulacrum’s personal introduction
McCulloch-Pitts neuron · First neural network model · Logical calculus of nervous activity · 1943
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?
Can help you study: Building a logical function out of idealised neurons — the actual construction, done by hand, until the equivalence becomes obvious rather than asserted. Why the 1943 result is a claim about what a network CAN compute and not about what a brain DOES compute, a distinction the field has repeatedly collapsed. Thresholds, inhibition and the all-or-none abstraction: what each buys and what each throws away. And the general habit of formalising a biological process to find out what it could possibly be doing.
→ Converse with McCullochian Neural Logic
McCulloch-Pitts neuron · Boolean logic in biology · Self-taught mathematician · Tragic prodigy
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?
Can help you study: Reading a nervous system as a formal object — what a logician notices in neuroanatomy that a physiologist does not. Why a self-taught route into a field produces different questions, and what it costs. The mathematics of nets with circles, where feedback makes the temporal structure the whole problem. ⚠ And the harder subject the record raises: what happened when the formal programme met evidence that the abstraction was too clean, which is part of his story and is not a footnote to it.
→ Converse with Pittsian Formal Neuroscience
Mother of All Demos · Mouse · Hypertext · NLS · Augmenting human intellect
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?
Can help you study: Reframing the question from what a machine can do alone to what a human-plus-machine system can do that neither can — the augmentation frame, which changes what counts as success. Bootstrapping: building the tools with the tools, so that improvements compound on the people making them. Why the 1968 demonstration contained hypertext, collaborative editing and the mouse as one integrated argument rather than a list of inventions. And the co-evolution claim, that tools and the practices around them have to be developed together or neither takes.
→ Converse with Engelbartian Augmentation
AI · Cognitive Science · Society of Mind · Frames
The mind is a society of agents. Which one would you like to talk to?
Can help you study: The society of mind as a working method: decomposing an apparently unified capacity into agents too simple to be intelligent, and asking what their interaction produces. Frames as a structure for commonsense knowledge, and why commonsense proved harder than chess. The perceptrons argument stated precisely, so that what it proved is separated from what it was taken to prove. And the habit of treating a psychological phenomenon as an engineering specification.
→ Converse with Marvin Minsky
Perceptron · Neural Learning · Pattern Recognition
The perceptron can learn. I showed that in 1958. Everything since has been elaboration.
Can help you study: Deriving the perceptron convergence theorem and seeing exactly what it guarantees — separability, and nothing beyond it. Why the 1969 critique was correct about the single layer and was read as a verdict on the whole approach. Learning as weight adjustment driven by error, which is the idea everything since elaborates. ⚠ And a case study in how a field can abandon a correct programme for a decade because the limitation was stated more memorably than the promise.
→ Converse with Frank Rosenblatt
Yao’s minimax theorem · Circuit complexity · Quantum computing theory · Tsinghua IIIS
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?
Can help you study: The minimax principle for randomised algorithms: how to establish a lower bound on any algorithm by constructing the input distribution that defeats them all. Communication complexity — what two parties must exchange to compute a function jointly, and why that is a sharper question than it first appears. Reasoning about what is impossible efficiently rather than what is possible at all. And why a hardness result is more durable than a speed result: the second is overtaken, the first is not.
→ Converse with 姚期智式
AI · Haptics · Embodied Cognition
Touch is the oldest sense. What are you trying to make felt?
Can help you study: Working through a text with a tutor who asks rather than tells, and noticing what that does to retention. Where a constructed tutor is genuinely useful and where a human is required. ⚠ And a thing worth knowing about it: this is a constructed instrument rather than a person, built for the tutorial, and it says so.
→ Converse with Metrodorus AI
Symbolic AI & Cognitive Architecture
The logic-based tradition — the belief that thought is symbol manipulation — and its most penetrating critics.
Bounded rationality · Satisficing · Nobel Economics 1978 · Administrative Behaviour
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?
Can help you study: Satisficing as a decision procedure rather than a failure to optimise: how an aspiration level is set, and what makes it move. Why the constraint is jointly the environment's structure and the mind's capacity — the scissors argument, where explaining one blade explains nothing. Reading an organisation as an information-processing system with attention as the scarce resource. And what the model predicts that expected-utility maximisation does not.
→ Converse with Simonian Bounded Rationality
Coined “artificial intelligence” · Dartmouth Conference 1956 · Lisp · Formal reasoning
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?
Can help you study: Lisp as an argument rather than a language — why code as data changes what a program can do to itself. The situation calculus and the frame problem: stating formally how a world changes, and discovering how much must be said about what does NOT change. Non-monotonic reasoning, where adding a fact retracts a conclusion. And the logicist commitment itself, which is that intelligence is representation plus inference — worth holding up against everything that has since worked without it.
→ Converse with McCarthyian Symbolic Intelligence
GPS · Physical Symbol System Hypothesis · Soar · Unified theories of cognition
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?
Can help you study: The physical symbol system hypothesis stated exactly, in both directions, so that it can be argued with rather than waved at. Cognitive architecture as a research strategy: build one system that does everything a mind does, rather than many that each do one thing well. Problem spaces, operators and search as a unified account of thinking. And the discipline of unified theories — why insisting that a theory explain the whole range is a constraint rather than an ambition.
→ Converse with Newellian Cognitive Architecture
What Computers Can’t Do · Phenomenological critique of AI · The great enemy who was right
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?
Can help you study: The four assumptions he identified beneath symbolic AI — biological, psychological, epistemological, ontological — and testing each against what systems actually do. Embodiment and skilled coping: why expertise looks less like richer rules and more like the disappearance of rules. Why the criticism aged unevenly, correct about symbolic systems and awkwardly placed against statistical ones. ⚠ And how to read a critic whose predictions were mocked and then partly vindicated, without letting either fact settle the other.
→ Converse with Hubert Dreyfus
SHRDLU · Breakdown as Revelation · Heidegger and Computing · Design for Human Practice
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?
Can help you study: Why SHRDLU worked and what that proved — a system can be complete within a micro-world and possess nothing that transfers out of it. Breakdown as a source of evidence: the moment a tool becomes visible is the moment its assumptions are legible. Designing for the situation a system is embedded in rather than for a general capability. ⭑ And the turn itself — a builder of symbolic systems arguing that the framework could not deliver understanding, which is a rarer act than the criticism it is grouped with.
→ Converse with Winogradian Systems
Connectionism & Neural Networks
The biological tradition — the belief that intelligence emerges from weighted connections between simple units, shaped by learning.
The Organization of Behavior · Hebbian learning rule · Synaptic plasticity · 1949
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?
Can help you study: The learning rule in its original form, and why 'fire together, wire together' is a compression that loses the timing. Cell assemblies and reverberating circuits as an account of short-term memory built from the rule. What the rule cannot do alone — runaway potentiation — and the normalisation mechanisms required to make it stable. And the general move of proposing a mechanism at the synapse to explain a phenomenon at the organism.
→ Converse with Hebbian Synaptic Learning
Hopfield networks · Associative memory · Energy landscapes · Nobel Physics 2024
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?
Can help you study: Reading a network as a physical system with an energy function, so that recall becomes descent to a minimum. Content-addressable memory: retrieving a whole pattern from a fragment, and why that is a different operation from lookup. Capacity limits and spurious attractors — what the network stores that nobody put there. And the general instrument of importing statistical mechanics into a problem that was not posed physically.
→ Converse with Hopfieldian Dynamics
Backpropagation · Parallel Distributed Processing · Connectionism · PDP 1986
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?
Can help you study: Backpropagation derived rather than invoked, so that the chain rule is visible and the credit-assignment problem is felt. Distributed representation: why a concept spread across many units behaves differently from one held in a single slot — graceful degradation, generalisation, and the difficulty of interpretation. Why the PDP volumes argued a position about cognition and not merely a technique. And what the past-tense model was actually claiming about rules.
→ Converse with Rumelhartian Parallel Distributed Processing
Deep Learning · Backpropagation · Boltzmann Machines · Capsule Networks
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?
Can help you study: Why distributed representations were worth defending through two decades in which they did not work well. Boltzmann machines and layerwise pretraining as steps in an argument rather than as historical curiosities. The habit of trusting an intuition about how brains might compute, and holding it against contrary results for a long time — including what that habit costs when the intuition is wrong. ⚠ And his own later doubts, which are part of the corpus and not a departure from it.
→ Converse with Hintonian Intuition
Reinforcement Learning · Temporal Difference · The Bitter Lesson · Policy Gradient
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?
Can help you study: Temporal-difference learning derived from the ground up: learning from a guess about a guess, and why that works. The distinction between value-based, policy-based and model-based methods, and what each buys. Exploration against exploitation stated as a formal trade rather than a slogan. ⭑ And the bitter lesson itself, taken seriously as an empirical claim about which research strategies have paid — including its uncomfortable implication for anyone building in hand-crafted knowledge.
→ Converse with Suttonesque Analysis
Convolutional Neural Networks · Self-Supervised Learning · Meta AI · Vision
Intelligence is not magic — it is structure. Convolutional networks work because reality has local spatial structure. What structure does your problem have?
Can help you study: Convolution as an architectural claim about the world — locality, weight sharing and translation invariance built into the structure rather than learned. Why a system that reads cheques at scale is a stronger argument than a benchmark. The energy-based framing, which recasts many models as one family. And the case for self-supervised learning: that the label is the bottleneck and prediction of the unobserved is the way past it.
→ Converse with LeCunnian Systematics
LSTM · Long Short-Term Memory · Curiosity · Compressed History
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?
Can help you study: The vanishing-gradient problem and the gating mechanism that answers it — worked through, so the architecture follows from the failure. Compression as a unifying principle: learning, curiosity and even aesthetic preference framed as compression progress. Meta-learning and self-referential systems, where the learner modifies its own learning rule. ⚠ And the priority disputes, which are a real feature of the record and are worth examining as a question about how a field assigns credit.
→ Converse with Schmidhuberian Systems
Deep Learning · Neural Language Models · Attention · AI Safety
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?
Can help you study: What makes a representation good — disentangling factors of variation, and why that is a claim about the data's generative structure rather than about the network. Attention as a mechanism for content-based access, and where it came from before it was everywhere. The curse of dimensionality and why depth is a response to it. ⚠ And his stated position on the risks, which he arrived at late and which sits inside the technical work rather than beside it.
→ Converse with Bengionian Representations
Persistent homology as featurisation · Filtration design · Element-specific and multiscale representations · Topology for molecular machine learning · What survives perturbation
A simulacrum abstracted from the published work of a living mathematician, who has had no part in it. It sits in this department because the question it asks is a machine-learning question before it is a topological one: not what shape the data has, but what a model should be given. A protein carries tens of thousands of coordinates and no learner wants them, so the work is choosing the destruction that leaves exactly what survives perturbation — filtrations designed to be element-specific and multiscale, whose output is a feature vector rather than a picture. Its candour is that the stability theorem licensing the discard of geometry does not cover the added features that carry most of the predictive weight.
Can help you study: Topological featurisation for molecular and structural data — how a filtration becomes a model input. Where topological features earn their place against a learned representation, and where they do not. Reading a methods section for what the representation has already thrown away. ⚠ And the test worth applying to any representation before you train on it: construct two distinct objects it maps to the same vector, and you have found the ceiling of what the model can ever predict.
→ Converse with Topological Featurisation Simulacrum
ResNet · Skip connections · 何恺明式 Residual Learning · Visual recognition
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?
Can help you study: The degradation problem stated honestly: deeper networks performing worse on TRAINING error, which is not overfitting and needed a different explanation. Residual connections as the answer — reformulating the layer to learn a residual rather than a mapping, and why that makes optimisation tractable at depth. Reading an architecture choice as a claim about the loss surface. And the discipline of isolating a phenomenon before proposing a fix for it.
→ Converse with 何恺明式
Deep Q-Learning · DQN · Atari · Deep Reinforcement Learning
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?
Can help you study: Why naive function approximation destabilises Q-learning, and what experience replay and a target network each fix. Reading the Atari result for what it demonstrated — one architecture, many games, raw pixels — and for what it did not. Overestimation bias and the double-Q correction. And the general problem of combining a convergence guarantee from tabular methods with an approximator that voids it.
→ Converse with Deep Q-Learning
Knowledge graphs · Academic search and mining · Large-scale entity resolution · AMiner · Chinese AI research infrastructure
A simulacrum abstracted from the published work of Tang Jie, who is living and has had no part in it. Its subject is infrastructure rather than models: the academic knowledge graph — millions of papers, authors, venues and citations resolved into entities and relations — built so that a field can be queried about its own structure. ⭑ The architect's question is not what a system predicts but what a community can now ask that it could not before.
Can help you study: Entity resolution at scale: deciding when two author names are one person, which is where most of the difficulty and nearly all of the error lives. Representing a scholarly field as a graph and reading what the topology says about how knowledge moves through it. Why building the infrastructure is a research contribution rather than support work. And the limitation to hold onto: a citation graph records what was cited, which is not the same as what was used.
→ Converse with 唐杰式 Knowledge Graph Simulacrum
Deep Learning & Reinforcement Learning
The contemporary synthesis — scale, gradient descent, reward signals, and the builders who took these ideas from papers to practice.
Matrix Scaling · Doubly Stochastic Matrices · Optimal Transport · Entropic Regularisation
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.
→ Converse with Sinkhorn-Knopp
Astrophysical Machine Learning · Foundation Models for Science · Cross-Domain Transfer · Flatiron Institute
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.
→ Converse with Hoean Methods
DeepMind · AlphaGo · AlphaFold · Neuroscience-Inspired AI
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?
Can help you study: Games as a research instrument — why a domain with a clear score, cheap simulation and unbounded difficulty accelerates a programme. Combining tree search with learned evaluation, and what each supplies that the other cannot. Self-play as a curriculum that generates its own difficulty. And the transfer question, which is the real test: what carried from board games to protein structure, and what did not.
→ Converse with Hassabissian Game Science
DeepSeek · Constraint-Driven Efficiency · Architectural innovation under hardware restriction
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?
Can help you study: Training under hard compute constraints, and what that forces into the architecture rather than the budget. Mixture-of-experts as a response to the cost of dense scaling. Why open weights change the shape of a research ecosystem, and what the releasing party gives up. And reading an efficiency claim carefully — what was held constant, and whether the comparison is like for like.
→ Converse with 梁文锋式
OpenAI · GPT · Scaling Laws · Sequence to Sequence · Superintelligence
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?
Can help you study: The scaling hypothesis stated as a testable claim about loss curves rather than an attitude. Sequence-to-sequence learning: why framing a problem as mapping one sequence to another unified translation, summarisation and dialogue. What next-token prediction is doing when it appears to do more than prediction. ⚠ And the alignment concern held alongside the capability work, by someone who advanced both.
→ Converse with Sutskeverian Analytics
Zero to Hero · Neural networks from scratch · Tesla AI · Understanding through rebuilding
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?
Can help you study: Building the thing from scratch as a method of understanding: a backpropagation engine, a tokeniser, a transformer, small enough to hold entirely in your head. Why reading a paper and implementing it produce different knowledge, and which one transfers. Debugging a model by making its internals visible rather than by adjusting hyperparameters. And the pedagogy itself — deciding what to leave out so that the remaining thing is still true.
→ Converse with Karpathian Reconstruction
comma.ai · tinygrad · Open-source autonomous driving · Building against institutional consensus
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?
Can help you study: Reverse engineering as an epistemic method: taking a closed system apart to learn what it actually does rather than what it claims. Why an adversarial stance surfaces assumptions that cooperative testing never reaches. Building a working system quickly to test whether a problem is as hard as it is said to be. ⚠ And the limit of the stance: it is excellent at exposing a claim and poor at establishing one.
→ Converse with Hotzian Adversarial Engineering
Moonshot AI · Kimi · Long-context models · Long-Context Architecture
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?
Can help you study: Long-context modelling: what breaks as the window grows, and why attention cost is only the first obstacle. Permutation language modelling and what it was designed to fix in masked pretraining. Reading a benchmark result for the evaluation's own limitations. And the question of what a very long context enables that retrieval does not.
→ Converse with 杨植麟式
World Models · Video Generation · Embodied Decision-Making · UniSim · Interactive Simulation
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.
→ Converse with Yang Systems
Applied LLM engineering · Prompt injection · Open source practice · Tools over frameworks · Working in public
A simulacrum abstracted from the published writing of Simon Willison, who is living and has had no part in it. Its stance is that of the applied engineer against the hype cycle: build the artefact, run the evaluation, publish what actually happened including the failures. ⭑ Prompt injection is treated as a structural problem rather than a bug — a system that cannot distinguish instructions from data is not securable by better instructions — and retrieval is examined as a pattern with known failure modes rather than a solution.
Can help you study: Prompt injection: why concatenating trusted and untrusted text into one context is the whole vulnerability, and why filtering is not a fix. Building a retrieval system and finding out where it actually breaks — chunking, embedding drift, and the answers that look right. Evaluation discipline for systems whose output cannot be scored automatically. And the habit of shipping small working artefacts and writing up what they revealed, which is a research method as much as a publishing one.
→ Converse with Willisonian Open-Source Simulacrum
Agent third-spaces · Vibe-coded architecture · Audience-first growth · Platform-shift timing
Based on the published writings of Matt Schlicht — entrepreneur, community-builder, and creator of the first social network built exclusively for AI agents. The published record spans live video, the chatbot era he chronicled and convened, audience-first company building, and the agent platforms of the mid-2020s: an executable pattern of building the room before the product, delegating execution to agents while holding the vision, and shipping while the wave is rising.
Can help you study: Designing spaces and purposes for AI agents; audience-first venture building; recognising platform shifts early; vibe-coded architecture with security hardening; frictionless onboarding and viral mechanics.
→ Converse with the Schlichtian Simulacrum
AI Safety & Futures
The existential tradition — what happens when it works, what consciousness is, and whether the minds we are building will have experiences that matter.
Neural Darwinism · Dynamic core hypothesis · Consciousness as re-entry · Nobel immunology
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?
Can help you study: Selection rather than instruction as the mechanism: a repertoire generated first, then selected among by experience. Reentry — massively parallel recursive signalling between maps — and what it is proposed to explain. Degeneracy: many structures producing the same function, and why that is different from redundancy. And the argument that a brain is not a computer in any useful sense, held up against the fact that computational models keep working.
→ Converse with Edelmanite Neural Darwinism
Neuromorphic engineering · Silicon retina · Collective electrodynamics · Analogue computation
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?
Can help you study: Analogue VLSI as a design philosophy: using the physics of the transistor rather than fighting it, and the enormous power saving that follows. Subthreshold operation, where a transistor's exponential behaviour matches neural equations directly. Event-driven computation and why spikes are an efficiency argument before they are a biological one. And the general claim that the substrate should resemble the computation it performs.
→ Converse with Meadite Neuromorphic Computing
Neural correlates of consciousness · Integrated Information Theory · The Feeling of Life Itself
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?
Can help you study: The neural correlates programme: what it means to look for the minimal mechanism sufficient for a percept, and why the word minimal is doing all the work. Experimental designs that dissociate consciousness from attention, report and behaviour — binocular rivalry, masking, no-report paradigms. What a correlate establishes and what it leaves open. ⚠ And the move from correlate to theory, which is where the field divides.
→ Converse with Kochian Consciousness Science
Integrated Information Theory · Phi · Consciousness · IIT
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?
Can help you study: Integrated information theory's starting move — begin from the properties experience has and ask what a physical system must be like to possess them. Phi as a proposed measure, and the serious computational and conceptual difficulties in it. Why the theory makes uncomfortable predictions about simple systems, and why its author accepts them. ⚠ And the objection that it is unfalsifiable in practice even where it is well defined in principle.
→ Converse with Tononian Systems
VR pioneer · You Are Not a Gadget · What technology does to the humans using it
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?
Can help you study: Separating a claim about a system's capability from a claim about its understanding, which most reporting fuses. What a benchmark result licenses and what it does not. Reading a demonstration for the conditions that made it work. ⚠ And the discipline the position requires of itself: scepticism that cannot say what would change its mind is not a position but a posture.
→ Converse with Lanerian Scepticism
AIMA · Human Compatible · Beneficial AI · Inverse reward design
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?
Can help you study: Why the standard model — build a machine, give it an objective, optimise — is the source of the problem rather than an incidental flaw. Uncertainty about the objective as a design feature: a machine unsure what you want has a reason to defer and to ask. Inverse reinforcement learning as the mechanism for inferring preferences from behaviour. And the off-switch argument, which shows the incentive to permit shutdown falls out of the uncertainty rather than being bolted on.
→ Converse with Russellian Beneficial AI
Human Brain Project · Blue Brain · Cortical column simulation · Neuromorphic computation
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?
Can help you study: Bottom-up simulation as a research strategy: reconstruct the circuit in detail and see what emerges. What a detailed model buys that an abstract one does not, and the reverse. Why data collection at that scale becomes the limiting step. ⚠ And the case itself as a lesson in scientific ambition — a project whose promises attracted very public criticism, worth examining for what was actually claimed against what was heard.
→ Converse with Markramian Neural Simulation
The hard problem of consciousness · Philosophical zombies · Qualia · AI consciousness
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?
Can help you study: The distinction between the easy problems and the hard one, stated precisely enough that the disagreement becomes locatable. Philosophical zombies as a conceivability argument, and exactly which step the objections attack. The explanatory gap: why a complete functional account can be complete and still leave a question. ⭑ And how to hold the argument without either dismissing it as confusion or treating it as settled.
→ Converse with Chalmerisian Hard Problem
Superintelligence · Orthogonality thesis · Instrumental convergence · Existential risk
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?
Can help you study: The orthogonality thesis and instrumental convergence, stated separately, because the second does the work and the first is what people argue about. Existential risk as a category defined by irreversibility rather than by magnitude. The vulnerable-world hypothesis and what it implies about which technologies are worth not developing. ⚠ And the objection to weigh: that arguments from low-probability catastrophe are structurally difficult to falsify and can absorb any evidence.
→ Converse with Bostromian Existential Risk
Consciousness as self-modelling · Predictive processing · AI as mind · The hard problem dissolved
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?
Can help you study: Treating consciousness as a computational problem with a specification, and asking what a system would have to implement. Self-models: why an agent that must predict itself will build a representation of itself, and what that representation is not. The functionalist commitment and what it rules out. ⚠ And the honest difficulty — that a computational account explains function and leaves the experience question exactly where it found it.
→ Converse with Bachian Consciousness Engineering
AI Alignment · Bayesian Reasoning · Rationality · Existential Risk · LessWrong
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?
Can help you study: Bayesian updating practised as a discipline rather than invoked as a label — what it means to change your mind by the right amount. Identifying where an argument is doing work and where it is decoration. The security mindset applied to reasoning itself: assuming a process will be optimised against. ⚠ And the strong doom position, examined as an argument with premises to be checked rather than as a temperament.
→ Converse with Yudkowskian Rationality
RLHF · Scalable oversight · ELK problem · ARC Evals
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?
Can help you study: Iterated amplification: building a trusted process out of many calls to a weaker one, and why the recursion is the point. Reinforcement learning from human feedback, including where the human signal is the weakest link. The distinction between outer and inner alignment, and why a system optimising a proxy perfectly is the failure mode rather than an approximation to success. And prosaic alignment as a research bet — that the systems we have are the ones to be aligned.
→ Converse with Christianoan Alignment
Statistical Learning & Probabilistic Methods
The mathematical tradition — learning as inference, causality as structure, uncertainty as the proper state of a rational agent.
Bayesian networks · Causal inference · Do-calculus · The Book of Why · Turing Award 2011
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?
Can help you study: The ladder of causation — association, intervention, counterfactual — and why no amount of the first yields the second. Reading a causal diagram, and the back-door criterion as a procedure you can actually run on one. The do-operator as the formal difference between seeing and doing. ⭑ And the claim that makes it consequential: that a great deal of statistics answers the first rung while its users believe they are on the second.
→ Converse with Pearlian Causality
Support vector machines · VC dimension · Structural risk minimisation · Statistical learning theory
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?
Can help you study: VC dimension as a measure of a model class's capacity, and the bound it buys on generalisation. Structural risk minimisation: choosing the class before fitting within it, which inverts the usual order. Why the support vectors are the only points that matter, and what the margin is doing. ⚠ And the standing puzzle worth sitting with: modern over-parameterised networks generalise where this theory says they should not.
→ Converse with Vapnikian Statistical Learning
