The Architecture of Intelligence

Objective

Survey definitions of AI, linking them with biology and academic tribes working on AI over past, present and future.

Seventy-five years of moving the goalposts — most of these names recur tonight

Acknowledgements

Leonardo Donando
Jonathan Florez Giraldo
Alex Kuznetsov
Delany Adom
Andrei Alain González Galeano
Pedro Domingos — for The Master Algorithm and the five tribes

With gratitude for their contributions.

Four Threads

Intelligence is Task Solving by learning
Minds run on prediction — animal, human, machine alike
“AI” is an Ambiguous Term — today, a marketing label for LLMs
Human intelligence is Sample Efficient Novel Task Solving — written first in our biology
Good prediction indicates learning

Outline

I Minds: consciousness & animals
II Human intelligence as machine learning
III “AI”: an ambiguous term
— coffee —
IV Generative AI & LLMs
V The five tribes
VI State of the art
VII What we want from AGI
VIII The future

I. Minds: Consciousness & Animals

Sentience vs. Sapience

Sentience — to feel, to experience (qualia)
Sapience — to judge, to act with wisdom
Today’s AI mimics sapience, shows no sentience
That asymmetry is the hard problem

Self-Awareness

Introspection — the self as distinct from the world
A thinker and a sculpted head: a mind watching itself think
Not the same as intelligence, nor consciousness
Perhaps a prerequisite for both

A person seated at a desk behind an openwork bronze sculpture of a head.

A mind watching itself think — and three meeting themselves: bear, infant, child pass the mirror test

A Ladder of Agency

Thermostat / bacterium — reactive feedback
Dog — bonding, mapping, goals
Octopus — distributed cognition, tool use
Dolphin — self-recognition, communication
Human — metacognition, culture

Primate: Our Closest Relative

Shared ancestry — chimps, bonobos ~6–7 Mya
Primate cognition: a rough draft of ours
The differences are as instructive as the likenesses

Bonobo

Octopus: Intelligence Reinvented

Last common ancestor ~550 Mya
Evolved completely independently of ours
A true natural experiment in cognition
Similar solutions → convergence is real

Messi: Football Intelligence

Growth-hormone deficiency as a child — physical limits may have driven compensatory perception
Reads play patterns before they unfold — anticipation over speed
Like chess intuition: pattern recognition, not raw calculation
A human specialization of the same predictive engine
Caveat: the hormone–cognition link is a popular theory, not proven

Getafe, 2007 — the pattern, read early

Emotion → Feeling → Sentiment

Emotion — raw input; a valence signal: good or bad, now?
Feeling — the conscious experience of that input
Sentiment — a compressed label from expressed feelings
Machine analogues: reward signals, then sentiment analysis

Friston: Prediction Under Uncertainty

The brain predicts the world, not just records it
Free Energy Principle: minimize surprise
Consciousness as ongoing active inference

\[\underset{\color{#f0a868}{\textbf{posterior}}}{\color{#f0a868}{P(\text{world} \mid \text{sense})}} \;\;\propto\;\; \underset{\color{#63b3a6}{\textbf{likelihood}}}{\color{#63b3a6}{P(\text{sense} \mid \text{world})}} \;\;\cdot\;\; \underset{\color{#7aa0e0}{\textbf{prior}}}{\color{#7aa0e0}{P(\text{world})}}\]

This closes Section I and is the keystone for the whole talk. Friston’s Free Energy Principle: the brain is not a passive recorder, it’s a prediction machine constantly minimizing surprise between what it expects and what it senses. Don’t derive the equation — just name it as Bayes’ rule: the posterior belief about the world is the prior updated by sensory evidence. Plant the flag here: prediction is the common thread from cortex to transformer, and I’ll cash this out after coffee.

Walk the loop once, left to right, and let the colours do the work — they match the equation underneath. Start with a prior: a broad, uncommitted belief about the world. Multiply by the likelihood, which is what the senses say. Out comes a posterior: the same belief, sharper. Then the move that matters, the dashed arrow: that posterior becomes the next prior, and the loop runs again. Note the two curves are the same claim drawn twice — the prior is a fat bump, the posterior a narrow spike. Perception isn’t a snapshot; it’s this loop, running continuously, all the way down.

If someone asks why “free energy” rather than just Bayes: free energy is the quantity you minimise to make this loop tractable when the exact posterior is out of reach. Say that, then stop — the maths past that point is a lecture of its own.

II. Human Intelligence as Machine Learning

Why Intelligence Evolved

An unstable niche — the world kept moving
Two bets: brittle instinct or costly inference
We paid to learn within a lifetime
Tools are the evidence of that inference
Three capacities: world-modelling, transmission, practice

Out of Africa into everything — and the handaxe that went with us, barely changed for 300,000 years

This slide answers the question Section II otherwise skips: why did this kind of mind evolve at all? Open with the trade-off and land the one line — instinct is a bet that tomorrow looks like today, and for us it didn’t. Terrain, climate, food supply and social arrangement all moved faster than genomes could track, so the cheap compiled answer kept going stale.

Point at the map for what that bet cost us: the same animal ends up in the Arctic and the Kalahari with no specialised instinct for either, paying for a big brain and a childhood far longer than any other ape’s. Then the handaxe — this is the slide’s best moment, so slow down. Making one means modelling a cause you cannot see, separating means from ends, planning for a need that hasn’t arrived, and carrying a technique onto unfamiliar material. Those are exactly the operations a novel problem demands, which is why a worked lump of flint counts as evidence of a mind. Say plainly that this is the standard adaptive account, not a settled one — same hedge as the Messi story.

Close by naming the three capacities and telling them to hold onto the names: world-modelling, transmission, practice. Self-supervised, supervised and reinforcement learning are their machine analogues a few slides from here, and I’ll call back to this list at the Unifying View.

Human Intelligence Properties

Short-Term and Long-Term Memory
Sample Efficiency — learn from a few examples
Generalization & Transfer
Compositionality & Abstraction
Causal Reasoning, Metacognition, Grounding

Supervised Learning

Map inputs → labeled outputs
Told the answer; adjust to cut error
Bottleneck: the cost of labels
Human parallel: a student with an answer key

Self-Supervised Learning

Data supplies its own labels — predict the masked word
The engine behind foundation models
The internet becomes supervision, unannotated
Human parallel: an infant learning by observation

Unsupervised Learning

No labels — find structure in the data itself
Clustering, dimensionality reduction, density estimation
Reveals hidden groupings and manifolds
Human parallel: noticing categories no one named

Reinforcement Learning

Learn from reward and consequence
Powers game-players and chat-model tuning
The agent generates its own training signal
Human parallel: a child learning by trial, reward, and consequence

Unifying View

Self-supervised prediction builds the world model
Reinforcement fine-tunes on top
The same recipe in human brain and in models

III. “AI”: An Ambiguous Term

Beginning: Dartmouth, 1956

McCarthy chose “artificial intelligence”
The conjecture: intelligence can be “so precisely described that a machine can be made to simulate it”
They thought a summer and 10 assistants would do it.

with Nathaniel Rochester — the four
who signed the 1955 proposal

The origin story, and it is evidence for the whole section — so tell it as evidence, not trivia. Summer 1956, Dartmouth College: a six-week workshop proposed the year before by McCarthy, Minsky, Rochester and Shannon. It is conventionally the founding moment of the field.

The detail that matters is the second bullet, and it is the reason this slide exists: McCarthy picked the words “artificial intelligence” because they were neutral — he wanted to avoid being annexed by cybernetics and its analog-feedback framing, and he wanted to avoid a fight with Norbert Wiener, who was formidable. So the name of our entire discipline was a political choice made to dodge a rival school and a difficult colleague, before there was any science to name. That is the ambiguity thread with a birth certificate: “AI” was a label first and a field second.

Land the last bullet lightly — about twenty people passed through, daily attendance ran three to eight, and only Solomonoff, Minsky and McCarthy stayed the whole summer. The founding event of the field was a loosely-attended summer workshop, not a moment of revelation.

Four Acronyms

ML — algorithms that improve at a task with more data
ANI — narrow: superhuman at one task, useless outside it
AGI — general: human-level breadth, novel tasks included
ASI — beyond: exceeds the best humans at essentially everything

ANI is here · AGI is the aspiration · ASI is the artwork

Pause here and pin the vocabulary down — the rest of the talk depends on these four not blurring together. ML is the method: systems that improve from data rather than hand-coded rules; it powers everything on this slide. ANI is what we actually have — chess engines, translators, LLMs: superhuman in a lane, helpless outside it. AGI is the aspiration — human-level breadth, and stress the word novel: problems it wasn’t trained for. ASI is the speculation beyond that; mention it once, don’t let the talk drift into sci-fi. Use the picture against itself for that — notice that ANI gets a robot arm playing chess (a real thing that exists), AGI gets a humanoid juggling tasks (a hope), and ASI gets a glowing cosmic figure rising over the Earth (a painting). The artwork escalates exactly as the evidence evaporates. That is the whole slide’s caution in one image, and it lets you make the point with a smile rather than a lecture. The trap to name out loud: “AI” in headlines slides between all four. When someone says AI, ask which one they mean — that’s the ambiguity thread again.

Make Coffee

Functionalist test: a machine does what a human did — intelligent?
Once we see the mechanism, we stop calling it AI
Chess, OCR, translation — “AI” until it worked
Make coffee in any kitchen in the world
“AI” often means: not yet understood

Coffee, yes — but in a rig built for it, not any kitchen

This is the “AI effect” and the emotional centre of the section. The pattern: chess, OCR, translation were all “AI” right up until they worked — then we saw the mechanism and demoted them to “just software.” “Make coffee in any unfamiliar kitchen” (Wozniak’s test) is the vivid version of a task still on the far side of that line. Use the clip against itself: yes, that arm makes coffee — but look where. It is a fixed barista rig built around the machine, every cup and spout in a known spot. That is the easy half of Wozniak’s test. The hard half — walking into a stranger’s kitchen and finding the beans — is exactly what no robot can do, and the gap between the two is the whole point. Land the punchline: “AI” too often just labels whatever we don’t yet understand — a moving goalpost, not a category. This directly teases the thesis slide.

Turing Test

Imitation: indistinguishable in text → passes
Says nothing about internal states
Cited often, engineered against rarely
A floor, not a ceiling

Russell & Norvig: Agents

Act to achieve the best expected outcome
The textbook definition — agent, environment, measure
Says nothing about how — only that it decides well

Chollet: Intelligence as Adaptation

Efficiency at acquiring new skills
Adaptation, not static performance
The machine echo of human skill acquisition
Memorize well, adapt poorly → not intelligent

One Definition of Knowledge

Data — raw, uncompressed
Compression — redundancy found
Abstraction — the reusable structure it yields
Prediction — the test on unseen data
Knowledge — abstraction that keeps passing

Process data to generate knowledge

Knowledge as Compression

Three traditional sources of knowledge:

Genes — survival encoded by evolution
Experience — direct interaction
Culture — collective wisdom, transmitted

A fourth new source, over the last 50 years:

Computers — synthesis at scale

☕ Coffee Break

15 minutes — please join us in the foyer

IV. Generative AI & LLMs

What Gen AI Is

Simulation from a partially learned distribution; produces new samples
Text, image, audio, code
Self-supervised pre-training + the transformer
Planetary scale recursive classification simulation

Noise → sample: a diffusion model inventing characters

Welcome them back from coffee, re-anchor quickly. Generative AI learns a probability distribution over data and samples new points from it — text, image, audio, code, all the same trick. Two ingredients did it: self-supervised pre-training plus the transformer. The phrase to plant, because it pays off two slides later: this is “System 1 at planetary scale” — fast, fluent, intuitive pattern-completion, with no deliberate reasoning behind it yet.

Let it loop while you talk — it is the first bullet happening in front of them. The counter runs from 1% to 99% denoised: the model starts from pure noise and walks it, step by step, into a sample. That is what “simulation from a learned distribution” means, and no equation lands it as fast.

The detail worth pausing on: these are not real Chinese characters. The model learned the distribution of characters — strokes, radicals, balance, the way they sit in a square — and drew new ones from it. They look right to anyone who doesn’t read Chinese, and wrong to anyone who does. That is generative AI in one image: fluent in the form, indifferent to the meaning. It also pre-loads “What Gen AI Is Not,” the next slide — fluency is not knowledge.

What Gen AI Is Not

No guarantee of truth, cause, or grounding
Fluency is not knowledge
It compresses culture and lets us query it
New and powerful — not a mind

The LLM

A next-token predictor over human text
No persistent memory, weak causal models
Unreliable deliberate reasoning
A statistical reflection of us

Attention, layer by layer — every token weighing every other

The most precise technical description in the deck — slow down and be exact, some of the audience are specialists. Strip it to the mechanism: an LLM predicts the next token over human text, nothing more. From that fall the real limits — no persistent memory across sessions, weak causal models, unreliable multi-step reasoning. Close on the resonant line: it’s a statistical reflection of us, our text mirrored back. That framing sets up the entire “state of the art” gaps section.

Let the animation loop while you talk — it is attention itself, every token weighing every other, layer by layer. Point at the crossing lines and say: that is the entire mechanism. No memory, no plan, no model of the world — weighted lookups over the context window, stacked a few dozen times. Everything the machine knows, it knows by re-reading its own input.

One caveat, and concede it fast if a specialist raises it: this is BertViz, so the tokens are BERT’s — [CLS] and [SEP] — and BERT is a masked model, not a next-token predictor. The attention mechanism is identical; only the masking differs. If that trade bothers you, the alternative is an architecture diagram, which is more precise and much less alive.

V. The Five Tribes

Symbolists — Logic and Deduction

Roots: logic — Newell & Simon’s symbol system hypothesis
Learning = inverse deduction; rules, trees, expert systems
Precise and auditable — but brittle
Optimization: combinatorial search over rule/hypothesis space
Modern echo: neuro-symbolic AI, formal verification
EXAMPLE: Routing Email Complaints

Every answer comes with its own audit trail

Josh Tenenbaum

Framing for the whole section — this is Pedro Domingos’ “five tribes” from The Master Algorithm. The symbolists came first: intelligence is the manipulation of symbols by explicit rules, and learning is inverse deduction — inferring the rules from examples. Their systems are precise and auditable, which is exactly why they’re brittle when reality doesn’t fit the rules. The email example: a rule tree that routes complaints by keywords — transparent, but it shatters on a phrasing nobody wrote a rule for. Modern echo: neuro-symbolic AI and formal verification.

Let the tree animation loop and use it for the third bullet — it is the tribe’s whole case in one picture. A record walks the tree, and at every node you can see exactly which rule fired and why: home owner, marital status, income, verdict. That is what “auditable” means, and not one neural network in this deck can show you that. Then turn it over in the same breath: the rigidity is the brittleness. The tree has no answer for a record it was never given a branch for — it cannot hedge, it cannot guess, it just falls off the end.

Connectionists — Brain Inspired

Roots: McCulloch–Pitts (1943), Rosenblatt’s perceptron (1958)
Backpropagation (1986) → CNNs → transformers
Optimization: stochastic gradient descent via backpropagation
Scales with data and compute — and dominates today
EXAMPLE: Making DNA viral capsids

Dendrites in, axon out — the unit they copied

Connectionists: Bolivarian Contribution

Two inspiring contributors from the Bolivarian nations!

Evolutionaries — Selection Algorithm

Roots: Darwin, formalized by Holland (1975)
Evolve a population — mutation, crossover, selection
No gradient, no labels — only a fitness score
Optimization: genetic algorithms / evolutionary search
Niche but creative: architecture search, robot morphology
EXAMPLE: Avoiding immune response for gene envelope

An evolved gait — most of the population falls over

Simulated annealing — escaping local optima

Darwin as algorithm, formalized by Holland in 1975. Use the walker: every one of those figures is a candidate gait, and almost all of them fall over. Nobody told any of them what walking is — there is no label, no gradient, only a score for how far you got before you hit the ground. Keep the winners, mutate, repeat. That is the whole tribe in one loop, and it is also why it is slow. You keep a population of candidate solutions and apply mutation, crossover, and selection against a fitness score — no gradient, no labels, just “did it work better?” That makes it the tool of choice when the problem is non-differentiable or you can’t even write down the objective: neural architecture search, evolving robot body shapes. The example ties back to the capsid: evolving a gene-therapy envelope that slips past the immune system, an objective too messy for gradients. Niche today, but genuinely creative.

Bayesians — Belief as Probability

Roots: Bayes (1763), Laplace
Learning = updating a distribution over hypotheses
Prior + likelihood → posterior
Optimization: probabilistic inference (MCMC, variational inference)
Calibrated uncertainty, at a computational cost
EXAMPLE: Soldier Death Benefit Pensions

\[\underset{\color{#f0a868}{\text{posterior}}}{\color{#f0a868}{P(H \mid D)}} \;=\; \frac{\underset{\color{#63b3a6}{\text{likelihood}}}{\color{#63b3a6}{P(D \mid H)}}\;\cdot\;\underset{\color{#7aa0e0}{\text{prior}}}{\color{#7aa0e0}{P(H)}}}{\underset{\color{#9aa0aa}{\text{evidence}}}{\color{#9aa0aa}{P(D)}}}\]

\[\underset{\color{#e06c75}{\text{predictive}}}{\color{#e06c75}{P(x^{*} \mid D)}} \;=\; \int \color{#63b3a6}{P(x^{*} \mid H)}\; \color{#f0a868}{P(H \mid D)}\; dH \qquad \text{— average over every hypothesis, don't pick one}\]

Analogizers — Similarity

Roots: psychology — reasoning from prior cases
k-nearest neighbors, SVMs, the kernel trick
Optimization: constrained optimization to maximize margin (SVMs)
Legacy: recommendation, few-shot via embeddings
EXAMPLE: Locating outperforming stock pairs

Vladimir Vapnik — co-inventor of the SVM

The margin, widened until it cannot be

The fifth tribe: reason by similarity to past cases rather than by rules or models. Roots in psychology; the workhorses are k-nearest neighbours and SVMs, whose kernel trick measures similarity in high-dimensional space. The legacy is everywhere — recommendation systems (“people like you”), and few-shot learning via embeddings, which is how modern LLMs retrieve. The stock example: pairs trading, finding two stocks that historically move together and betting on the gap when they diverge — pure reasoning from analogous past behaviour. Set up the synthesis slide next: today’s systems quietly use all five.

The flag is a callback to Dartmouth, and worth thirty seconds. Vapnik and Chervonenkis built statistical learning theory — VC dimension, the margin, eventually the SVM — in Moscow through the 1960s and 70s, at an institute for control sciences, under the disciplinary banner of cybernetics. That is the very word McCarthy picked “artificial intelligence” to avoid. So the same intellectual territory got two names on two sides of the Iron Curtain, and the Western half spent decades not reading the Eastern half: VC theory only really landed here after Vapnik came to Bell Labs around 1990. The point for the ambiguity thread: the name a field carries is an accident of politics, and it determines who reads you.

The animation is the tribe’s actual idea: two classes, and infinitely many lines that separate them — the SVM picks the one that sits as far from both as possible. Widen the margin until it jams. Say that the margin is not a detail, it is the whole theory: Vapnik’s claim was that the widest gap is the one that generalises, and he had the proof, not just the intuition.

The Frontier Borrows From All Five

No production system is one tribe
LLMs: connectionist core, Bayesian uncertainty, symbolic tools, RL tuning
Domingos’ provocation: one master algorithm?
Which synthesis will your work advance?

VI. State of the Art

Brittle AI & the Reasoning Gap

Narrow, brittle outside the training distribution
The frontier problem is reasoning, not scale
Performance can mask absent understanding
The $1M ARC Prize still stands — no system has claimed it

Set the honest tone for Section VI — this is the reality check after the tribes. Today’s systems are narrow and brittle the moment you step outside the training distribution. The reframe that matters: the frontier problem is reasoning, not scale — more parameters won’t close this gap. And the warning to the researchers in the room: benchmark performance can mask a total absence of understanding, so measure carefully and distrust a single headline number. The animation is an ARC-AGI puzzle being solved — gesture at it when you name the reasoning gap: a child infers the rule from two or three examples, and this family of tasks is precisely where frontier models still fall over. Then point at the leaderboard: on ARC-AGI-3 the best frontier systems sit under 8%, and only by spending thousands of dollars per run — the score axis tops out at 20% and the cost axis is logarithmic. That one chart is the whole slide in numbers. And the last bullet is the stakes made concrete — the $1M ARC Prize for a system that cracks these tasks efficiently is still unclaimed; the gap isn’t rhetorical, it’s a standing bounty nobody has collected.

System 1 vs. System 2

After Kahneman: fast intuition vs. slow reasoning
LLMs excel at System 1, stumble at System 2
“Reasoning” models narrow the gap at inference time
Open and fast-moving

Reality Gets a Vote

Knowledge is what survives a test
Code and maths moved fastest — they can be checked
Fluency with no verifier → hallucination
Essays, strategy, taste, most real tasks: no oracle

Mike Tyson standing over a floored opponent, captioned "Everyone has a plan 'till they get punched in the mouth."

Pearl’s Ladder

See — P(y | x)
Do — P(y | do(x))
Imagine — what if I had?
LLMs just see, no doing or imagining without a verifier!
Can’t watch your way to cause

The Experiment Is the Intervention

do(x) — reach in, set x by hand, cutting its usual causes
That severed graph is what an experiment is — randomise, don’t just watch
Verifying and causing are one move
To do experiments and find cause, you need a body

VII. What We Want from AGI

Generality, Efficiency, Reasoning

Generality — competence transfers to the untrained
Sample Efficiency — human-like, from few examples
Robust Reasoning — multi-step, causal, self-correcting

Memory, Grounding, Aligned Autonomy

Persistent Memory — no catastrophic forgetting
Grounded World Models — causal, ideally embodied
Autonomy + Alignment — open-ended yet controllable
The last is a constraint — and the highest-stakes one

VIII. The Future

Embodied AI & Moravec’s Paradox

Intelligence may require a body
Moravec: reasoning is cheap, sensorimotor skill is dear
Crossing a room is harder than grandmaster chess

Foundation Models Meet Robotics

Model “brains” in capable “bodies”
Surgery, deep-sea exploration, autonomous logistics
The loop closes: sense → predict → act → learn
Where the next decade concentrates

Unitree G1 — a body waiting for a better brain

This is where I place my bet on the next decade. Let the robot loop — it is a Unitree G1, a genuinely capable body you can buy today, and it is the point of the whole slide: the hardware has arrived, the brain has not. Sensorimotor competence is nearly solved; what it lacks is the grounded, causal world-model from Section VI. That gap is the bet. Put the foundation-model “brain” into a capable “body” — surgery, deep-sea and space exploration, autonomous logistics. The unifying idea is the closed loop: sense, predict, act, learn — which is Friston’s active inference and the RL loop from Section II, now running in the physical world. Tell the candidates plainly: this is where I’d concentrate, the convergence of foundation models and robotics. But hold it lightly — a bet, not a prophecy.

AI Ethics — Beyond My Expertise

Cheap Dopamine — engineered attention capture
Economic Effects — labour, wealth concentration
Children — growing up AI-saturated
Distribution — who reaps the fruits of the technology
A prior, held lightly: fewer regulations tend to serve better

Key Takeaways

“AI” is an ambiguous term — today it means LLMs
Minds run on prediction — octopus to cortex to transformer
Human intelligence is the ML task list, written first in our biology
The open gaps — reasoning, memory, grounding, alignment — are the agenda

Closing Remarks

This PhD program is not about mastering today’s tools — it is about building the future.

You are the architects of the world models that will define the coming decade. The questions left open tonight — what consciousness is, how memory should persist, what closes the reasoning gap, what we should and should not want from AGI — are not gaps in this lecture.

They are your thesis directions.

Key References I

Foundations & cognition

Turing, A. (1950). Computing Machinery and Intelligence. Mind.
Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.)
Kahneman, D. (2011). Thinking, Fast and Slow.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.

Key References II

Symbolic & neural roots

Newell, A. & Simon, H. (1976). Computer science as empirical inquiry: symbols and search. CACM.
McCulloch, W. & Pitts, W. (1943). A logical calculus of ideas immanent in nervous activity.
Rosenblatt, F. (1958). The perceptron. Psychological Review.
Minsky, M. & Papert, S. (1969). Perceptrons.

Key References III

Learning at scale

Rumelhart, D., Hinton, G. & Williams, R. (1986). Learning representations by back-propagating errors. Nature.
Vaswani, A. et al. (2017). Attention Is All You Need. NeurIPS.
Hinton, G., Bengio, Y. & LeCun, Y. — 2018 ACM A.M. Turing Award, for deep neural networks
Dean, J. et al. (2012). Large scale distributed deep networks. NeurIPS.

Key References IV

Defining intelligence

Holland, J. (1975). Adaptation in Natural and Artificial Systems.
Sutton, R. (2019). The Bitter Lesson.
Chollet, F. (2019). On the measure of intelligence. arXiv:1911.01547.
Domingos, P. (2015). The Master Algorithm.
Moravec, H. (1988). Mind Children.

Image Credits

Bonobo — natataek, CC BY-SA 3.0
Octopus — Diego Delso, CC BY-SA 4.0
L. Messi goal clip (Getafe, 2007) — broadcast footage, rights holder unidentified
Mirror self-recognition (bear, infant, child animations) — via Medium
G. Hinton — Cmichel67, CC BY-SA 4.0
Y. LeCun — J. Barande, CC BY-SA 2.0
Y. Bengio — Xuthoria, CC BY-SA 4.0
J. Schmidhuber — ITU/R. Farrell, CC BY 2.0
D. Kahneman — nrkbeta, CC BY-SA 2.0
J. Pearl — Better Than Bacon, CC BY 2.0
J. McCarthy — null0, CC BY-SA 2.0
M. Minsky — OLPC, CC BY 3.0
C. Shannon — Tekniska museet, CC BY 2.0
F. Chollet — Ramosset, CC BY-SA 4.0
A. Turing — Elliott & Fry, public domain
T. Bayes — public domain (attribution disputed)
Soviet flag — public domain
Early human migrations map — Dbachmann, CC BY-SA 4.0
Acheulean handaxe (Saint-Acheul) — Didier Descouens, CC BY-SA 4.0

Simulated annealing — Kingpin13, CC0
Map boundaries — Natural Earth, public domain
J. Holland, V. Vapnik, J. Tenenbaum, K. Friston — no free licence identified; used under fair dealing for education
Portrait with sculpture — personal photograph, used with permission
Attention animation — BertViz, via comet.com
Diffusion denoising animation — via Towards Data Science
Decision tree animation — via Medium
Biological neuron animation — via Telefónica Tech
Evolved walker animation — via kottke.org
SVM margin animation — via Medium
ANI/AGI/ASI illustration — via Medium
AI history timeline — via LinkedIn
Unitree G1 robot animation — via roboticgizmos.com
Coffee-making robot animation — via Mothership
M. Tyson quote poster — rights holder unidentified; used under fair dealing for education

Photographs from Wikimedia Commons unless noted. Titled slide image: Piedras del Tunjo.

Discussion

What is the most important open problem in AI — and how will you spend the next five years on it?