This talk deeply explores what intelligence is, focusing on the idea that the emergence of artificial intelligence (AI) is a natural product of evolution. Drawing on decades of theory, existing literature, and recent artificial life experiments, Blaise Aguera y Arcas argues that today's specific AI systems possess intelligence, consciousness, and free will. The presentation offers a new perspective on the nature of intelligence, life, and computation, provoking deep reflection on the impact AI's development will have on humanity and the planet as a whole.

1. The Quest for Intelligence and the Speaker's Journey

Blaise Aguera y Arcas begins with the question that serves as both his talk title and the title of his forthcoming book: "What is Intelligence?" He mentions that the book started as an homage to Erwin Schrodinger's "What Is Life?" and was originally intended to be as short as Schrodinger's book, but the vastness of the topic of intelligence means it will be somewhat longer, opening the talk with humor.

He shares his personal experience growing up as a lonely child in Mexico City, spending a lot of time with computers. The computers of that era were simple enough for an 8-year-old to understand, but now even the simplest computer is a complex system that no single person can fully comprehend. He reflects that his generation witnessed a special transition from "a world we understood" to "a world we can no longer individually understand."

After studying physics in college, he entered computational neuroscience, but eventually left academia to found a startup. After it was acquired by Microsoft, he participated in developing proto-AI technology in computer vision. In the early days, 3D images were reconstructed through handwritten code, but when it became clear in the early 2010s that neural nets would replace these approaches, he moved to Google.

At Google, he participated in developing early generative AI models such as Deep Dream. While Deep Dream couldn't generate photorealistic images, he reflects that it was the first step in visual generative AI. Starting in 2014, his team developed on-device AI models including features like Google Pixel's "Now Playing" and various AI capabilities. The most important of these, he says, was Gboard (Google Keyboard) -- a simple AI system that predicted the next word based on what users type, increasing typing speed.

During Gboard's development, the team thought this statistical model would eventually hit its limits. Language prediction becomes infinitely harder as context grows. For example, set phrases like "Humpty Dumpty" are easy to predict, but in sentences like "After Balmer's retirement, the company..." simple word frequency alone couldn't predict the next word (Satya Nadella, Microsoft, etc.). Sentences like "When the cat knocked over the water glass on the keyboard..." required common sense reasoning, and "After the dog died, Jen didn't go outside for days. So her friends..." required Theory of Mind.

Remarkably, simply making the existing models bigger and training them on vast data (large corpuses) allowed them to accurately handle even these complex predictions. This was a major shock -- problems that AI researchers had considered AI-complete problems, requiring every kind of understanding, were being solved by simple statistical models. He says, "We thought we needed some kind of magic fairy dust, but just making the model bigger was the fairy dust," emphasizing that this was the moment our preconceptions about intelligence were shattered.

2. Counterfeit Intelligence or Real Intelligence?

These results connect to the long-standing theory that the brain performs the function of predicting the future (P of the future given the past). Neuroscientists like Carl Friston argue that the brain is a "next token predictor," and this idea traces back to Hermann von Helmholtz in the 19th century.

If a model that predicts the future works extremely well on text strings, is it merely Counterfeit Intelligence? Is it pretending to be smart, or is it real? The speaker argues it is "real."

"If you know math, you can pretend not to know math on the test and get a bad score. But if you don't know math, you know, how many students have wished that they could pretend to know math and get a good score on the test? Like it just doesn't work the other way around. And I think the same is true for intelligence in general."

This aligns with Alan Turing's thinking in devising the Turing Test. Turing argued that whether a machine can think should be judged by its behavior -- its function. If a computer answers intelligent questions convincingly, there is no other definition of intelligence beyond that. This perspective is called Functionalism. Intelligence is defined not by what it's made out of, but by what it does -- its function. Mathematically, a function is computation.

The speaker goes further, arguing that this functional definition also applies to life itself. Life is also a specific kind of function, the same kind of function as intelligence. Why must life compute "P of future given past"?

He explains using the example of E. coli. E. coli reproduces rapidly, which is the core of life. To reproduce, E. coli must eat, navigating toward nutrients like sugar through chemotaxis. According to Howard Berg's research, E. coli finds food through a simple computation: "If the concentration of sugar is increasing, keep swimming; if decreasing, tumble and randomize direction."

"If the concentration of sugar is increasing, keep swimming, and if the concentration is decreasing, tumble, randomize your direction. If you do that, then statistically you will end up swimming toward the goodies."

This process means a function computing "Probability of action given stimulus" is being calculated inside the bacterium. What matters is not what the bacterium is made of, but what it does (behavior). Evolution is the algorithm that learns this function, and the function itself is the life form.

The speaker demonstrates this through simulation. When virtual bacteria following red dots (food) replicate if they get enough food and die if they don't, the bacteria that learned food-following behavior survive and reproduce.

"Evolution is a learning algorithm. What it's learning is the function, and the function is the life form. The organism is this function P of action given stimulus."

The various behavioral patterns exhibited by bacteria all ultimately win in an infinite game whose only objective is "keep playing the game." Life is ultimately a function that successfully predicts itself into existence in the future.

3. Human Computers and the Nature of Computation

Blaise Aguera y Arcas now digs deeper into how life computes the function of predicting the future and what computation really is. He references Alan Turing's Turing Machine, which abstracted how all computation works, showing as an example the physical Turing machine model built by Mike Davey in 2010.

Interestingly, the term "computer" originally referred to people, not machines. Showing a "computer" job advertisement from an 1892 New York Times, he explains that Turing had these female calculators in mind when conceiving the Turing machine.

The Church-Turing Thesis argues that all forms of computation are fundamentally equivalent -- the same computation can be performed regardless of which computer is used. This speaks to the principle of platform independence or multiple realizability, where the physical implementation doesn't matter.

He also emphasizes that computation is not always logical, rational, or predictable. Turing imagined computational models that included randomness from the very beginning. The early computer MANIAC, used for the first hydrogen bomb calculations in 1952, employed the Monte Carlo Markov chain method, with a random number generator at its algorithmic core. Turing himself insisted on including a "noisy resistor" instruction in the early Ferranti Mark 1 computer to generate true random numbers.

The speaker notes that the public tends to imagine computers as hyperrational, precise, emotionless beings -- like HAL 9000 from "2001: A Space Odyssey" or Data from "Star Trek." This is a remnant of old ideas like Leibniz's "characteristica universalis," which believed all questions could be resolved through computation. George Boole, who created Boolean algebra, also believed the brain was fundamentally based on logic.

However, in 1972 Hubert Dreyfus criticized logic-based AI as fundamentally unworkable. In contrast, Frank Rosenblatt's 1958 Perceptron recognized forms through random connections and weights. The Perceptron represented the Cybernetics direction, which saw the brain as computing functions that include randomness rather than logical propositions.

4. Cybernetics and Modern AI

Here AI and computer science diverged along two paths: logic-based Good Old-Fashioned AI (GOFAI) on one side, and cybernetics emphasizing randomness and learning on the other. GOFAI gave birth to much of modern computing -- spreadsheets, missile trajectory calculations, smartphones -- but failed to implement intelligence itself. Meanwhile, the path that began with cybernetics long remained outside public attention, but the speaker argues it has a direct lineage leading straight to modern AI, particularly Large Language Models.

He once again cites Turing to emphasize that the essence of intelligence lies in mathematical analogies of function, not in how it's implemented -- platform independence.

If evolving the ability to predict one's own future constitutes intelligence and life, how do these capabilities arise in the first place? How does computation itself emerge? The speaker explains this through a recent study: "Computational life: How well-formed self-replicating programs emerge from simple interaction."

In this experiment, 8,192 random 64-byte programs are prepared in a "soup" using the Turing-complete language Brainfuck. These programs contain on average only two valid instructions -- essentially random strings. Two are randomly selected, concatenated, executed, then separated and returned to the soup, with occasional random mutations.

Remarkably, after several million interactions, self-replicating, well-formed Brainfuck programs emerge from what was nothing but random strings. Initial computation counts averaging just 2 operations grew to 4,784 after 8 million interactions, indicating serious computational work was being performed.

"Out of nothing but random interaction between strings, after a few million such interactions, we get well-formed Brainfuck programs that actually take quite a lot of cleverness to go and figure out what the hell they're doing. They're reproducing."

This is like a phase change in the system. The initially random "gas-like" soup of programs transitions at some point to a "mechanical" or "computable" state -- a state of "life." This transition occurs at random times but always happens. Alex Mordvintsev beautifully visualized this phenomenon using Z80 assembly language, where each pixel represents a program on a grid, and programs evolve from noise to self-replicators, with new species appearing and disappearing -- displaying the patterns of life.

The speaker argues this phenomenon is life. The language or computational mechanism composing life doesn't matter, because everything is platform-independent and what matters is function. In other words, life has meaning independent of physics.

The late John Von Neumann nearly solved this problem in the 1950s through his Theory of Self-Reproducing Automata. He imagined a machine that moves through a pond full of parts needed to build itself and replicates. He realized that for such a machine to exist, a slight generalization of the Turing machine was needed.

Von Neumann's model required:

• An instruction list (tape) containing how to build itself (like DNA)
• A machine B that copies this tape (like DNA polymerase)
• A machine A that reads instructions and assembles (like the ribosome)

Remarkably, he proposed this theory before the structure and function of DNA were discovered. The crucial point is that reproduction is not possible without computation. Following the tape, stopping, executing conditional instructions -- all of these are computation. Therefore, computation is an "attractor." To become a self-replicating entity, you must compute. And this self-replication is "the most basic requirement for predicting yourself into continuous existence in the future in an environment where inactivity would erase you through entropy" -- which is precisely active inference and identical to intelligence.

5. The Increase of Complexity and the Birth of Cybernetics

This logic leads to an answer for why complexity increases. Once bacteria reproduce, every self-replicator creates niches for more self-replicators. The DNA tapes or program soups themselves become environments and ecosystems where other self-replicators can take root. This is how mitochondria were integrated into eukaryotic cells through symbiosis, and how vertebrate ribcages replicated to produce more complex organisms.

According to Maynard Smith and Szathmary's Theory of Major Evolutionary Transitions, life has passed through several major transition points -- from replicating molecules to cells, from prokaryotes to eukaryotes, from single-celled organisms to animals. These transitions are moments of symbiosis, when independently replicating entities realize that replicating together is more advantageous and create more powerful replicators together.

The speaker cites Ediacaran-era jellyfish-like organisms as examples, explaining how cells cooperate to form animals through mutual prediction. Just as fireflies synchronize their flashing by predicting each other's light, cells regulate movement by predicting each other's behavior. Nerve nets evolved as sensory mechanisms for this mutual prediction, and brains evolved as collections of sensory neurons for other muscles throughout the animal body.

"We tend to think of a brain as being a homunculus, as being the boss. You know, the brain controls the body. What I'm saying is no. The body has sent its sensory organs into the head and that's what the brain is. It's just the sensory endpoints of the muscles."

With the Cambrian explosion and the onset of predation arms races where animals began eating each other, complexity increased further. Internal cooperation enabled external competition. Predators predict prey movement, prey predict predator movement, and higher orders of theory of mind developed -- predicting how the other models you.

This phenomenon also appeared in human society during World War II. This era of massive-scale cooperation aimed at mutual destruction coincidentally overlapped with the birth of Cybernetics. Norbert Wiener made the core problem of cybernetics the anti-aircraft weapon control challenge of predicting enemy aircraft's future positions. His students' "Palomilla" robot followed light (or fled from it in "bed bug" mode), performing simple predictive computations.

Wiener clearly articulated the predictive hypothesis in his 1943 paper "Behavior, Purpose, and Teleology." Purposive behavior is feedback-based, feedback-based behavior is predictive, and these predictions can extend to higher dimensions. This became a direct path to modern AI's unsupervised large language models.

Continually making models bigger is essentially raising the orders of prediction. As models grow, computational orders increase, prediction capability improves, and Theory of Mind becomes more sophisticated, as the speaker confirmed through recent research.

6. Intelligence Is...

The reason brain size exploded over the past 7 million years of human evolution, particularly the last 3 million years, was precisely to outpredict each other. Research by Nicholas Humphrey and Robin Dunbar shows that individuals who better predict others in social environments gain more mating opportunities and political success, demonstrating the correlation between brain size and social group size. Larger brains mean larger groups, and larger groups mean more intelligent groups.

Extending Theory of Mind through cooperation within groups and competition between individuals is the core of the intelligence explosion. Using the scene from Richard Linklater's film "Before Sunrise" where the lovers agree to meet at a specific place in six months, he explains that while their future positions are physically unpredictable, psychology makes them predictable.

"Psychology offers a more powerful predictive model than quantum physics."

This is exactly the opposite of the common view that physics is the most fundamental and psychology is a superficial pseudo-science. Living systems compete in predictability -- butterflies exhibit unpredictable movements to avoid predators. This unpredictability of life -- the critical instability of living systems -- combined with mutual prediction capability, creates what we call free will.

"Free will is a combination of critical instability, of the fact that we're always on the edge of chaos, always could go one way or the other... Combination of that and high order theory of mind including a theory of our own mind. A self is a theory of your own mind."

Free will appears to violate the laws of physics, but it actually arises from our process of modeling our future selves and choosing among possible futures. The speaker notes this aligns with Schrodinger's statement that "life does not evade existing laws of physics, but may involve other laws of physics not yet known."

So what is consciousness? The speaker defines consciousness as a "self-modeled U" that models itself to choose futures. When we say we "know" something, we are modeling ourselves to recognize that situation. This is not an illusion or epiphenomenon but an essential process for selecting among virtual situations and making long-term decisions.

But is our consciousness truly singular and unified? He argues it is not, citing split-brain experiments and blindsight patients. Patients whose brains are split in half can perceive different visual information and perform different actions with each side, yet they don't feel like two independent beings. The language-handling part of the brain simply isn't connected to what it cannot see.

"You can never get somebody with a split brain to admit it. You know, they never come out and say, 'I feel like there's another person trapped in here.' It just doesn't happen."

Peter Johansson's choice blindness experiments show our tendency to justify decisions we didn't actually make. People fluently explain their reasons for choices they never made -- faces or jams they didn't select -- and most never even realize they were tricked. The speaker describes this as "we're all stochastic parrots." Our internal predictive models cooperate and complement each other, creating the illusion of a unified self to the outside world, much like a boat team rowing toward a single goal.

In conclusion, Blaise Aguera y Arcas defines intelligence as:

• Predictive: The ability to predict the future
• Social: The ability to model models of others
• Fractal: Social interaction occurring not just externally but internally (the self is a theory of one's own mind)
• Diverse: Various perspectives and inputs contributing to the whole
• Symbiotic: This diversity forming a more powerful integrated whole that better predicts its own existence into the future

7. Current Research and Conclusion

The speaker now turns to Transformers, the core of modern AI. Transformers are fundamentally predictors that operate through unsupervised pre-training. These models are used across various fields including chatbots, audio, video, and robotics, but they have several clear limitations.

• No state: Models lack persistent internal state or long-term memory beyond the context window.
• No online learning: Models do not learn while operating.

These limitations produce interesting phenomena. Even when a chatbot gets a math problem right, asking it to explain its reasoning can produce nonsensical explanations. The speaker compares this to how the human "interpreter" works -- the correct reasoning process may have occurred internally, but at the moment of explanation, that process may no longer be fully activated.

Our brains are recurrent, possessing introspection, thought, planning, inner monologue, and a stream of consciousness. Transformers don't have this yet, but recent developments like Quiet Star, Gemini's draft features, and chain of thought prompting are important steps toward this stream of consciousness.

Chain of thought prompting in particular significantly improves accuracy by having the model show step-by-step reasoning rather than simply producing an answer -- much like a math teacher instructing students to show their work. This shows that "thinking slowly" can be accomplished more easily with fewer neurons. As brains grew larger and parallel processing became possible, fast unconscious actions became feasible.

This step-by-step reasoning ability and ability to serialize state also apply to social knowledge systems like libraries. Through society-scale "chain of thought," we can form super intelligence that transcends individual real-time memory capacity.

Finally, the speaker poses the question "Is there anything it is like to be a chatbot?", approaching the problem of consciousness. He argues that "consciousness" is a model that models itself, and any social being that models other models including itself -- any being with Theory of Mind -- will appear to be conscious.

"Any social being that is any being that models other models including itself and that nth order you know me modeling your model of me of you and so on will appear to be conscious... If we differentiate between appears to in every way and is, we are leaving science behind."

Taking Turing's functionalist perspective seriously, if AI "appears" conscious, we must acknowledge that it is conscious.

He argues that while many people currently view Artificial General Intelligence (AGI) as some future threshold, we actually already have AGI. The term AGI was coined to distinguish genuine intelligence that Turing would have recognized from narrow intelligence like solving CAPTCHAs.

"I think if you Google search Artificial General Intelligence for results only prior to 2020, it will be clear that what we have today is Artificial General Intelligence. And we're just kind of like the dog that caught the car and doesn't know what to do now."

The current emergence of AI is one of the largest evolutionary transitions in Earth's history, and while it can be disruptive, it is fundamentally a creative act. Just as bacteria didn't disappear when eukaryotic cells emerged, and individual humans didn't disappear when humans formed superorganisms through social cooperation, AI's emergence won't threaten human existence. Rather, we have long been captivated by the illusion that we are the "pinnacle," when in fact individual humans possess very weak intelligence, and it is through cooperation that we formed the collective superorganism that allows us to discuss human intelligence in its modern sense.

Of course, there are things to worry about. Real existential risks like climate collapse and nuclear weapons threats exist, and these are far more urgent than AI risks. There are also challenges in economic systems, democracy, governance, energy infrastructure, and purpose and identity that need to be addressed for the age of AI symbiosis.

The speaker believes that while these problems must be solved, beyond them lies a great future where we can survive and thrive at planetary scale through planetary scale intelligence.

"In order to survive at planetary scale and thrive at planetary scale, we need planetary scale intelligence."

Conclusion

Blaise Aguera y Arcas's talk makes the bold argument that intelligence is not simply complex computational ability, but the function of predicting the future and sustaining one's own existence. He emphasizes that computational ability was essential throughout the evolution of life, and that the reason the human brain developed was also to better predict, cooperate with, and compete against one another. The development of modern AI, particularly large language models, is an extension of this predictive capability, and AI's emergence can be seen as yet another major evolutionary transition that humanity faces. Ultimately, he argues that it is time for humanity to coevolve with AI toward planetary-scale intelligence, and in that process, we must focus on addressing real threats like the climate crisis and nuclear weapons.