Defining Life, Computation, and Symbiogenesis
Blaise Aguera y Arcas, CTO of Technology and Society at Google and founder of Paradigms of Intelligence, an organization dedicated to basic AI research, recently published a book titled "What is Intelligence?". He discusses his definition of life, the concept of functionalism, and how humanity can coexist with Artificial General Intelligence (AGI).
Defining Life, Computation, and Symbiogenesis
Aguera y Arcas defines life as "self-constructing computation that complexifies through symbiogenesis." This definition is a synthesis of concepts from computer science and biology.
Self-Constructing Computation
The term "computation" often brings to mind modern computers, but Aguera y Arcas emphasizes that computation, as conceived by Alan Turing and John von Neumann, is a fundamental science. Turing developed a mathematical abstraction of computation, reducing it to a "Turing machine" – a read/write head interacting with symbols on a tape. The key insight was the concept of a "universal Turing machine," capable of performing any computation by interpreting rules written as data on its tape. This is precisely what a modern computer does.
Computation is deeply linked to causality and entropy. Any behavior or computation requires conditionality, an "if-then" structure, to create causes in the world. This is evident even in simple systems like a thermostat. The science of causality provides a framework for understanding how computation arose in nature, particularly with the emergence of life.
John von Neumann further advanced the understanding of self-reproduction. He theorized how a system could reproduce itself without infinite regress by having an internal instruction tape containing instructions for two "machines": machine A (which builds according to the instructions) and machine B (which copies the instruction tape). Crucially, the instructions for building both machines A and B must also be on the tape. Von Neumann's theoretical work in the late 1940s remarkably predated the discovery of DNA's structure and function, the ribosome (his "machine A"), and DNA polymerase (his "machine B"). This implies that the origin of life, with its heritable self-reproduction, is fundamentally the origin of the first computers. Thus, biology can be seen as computer science, and our modern devices are "artificial computers."
Life must be self-constructing, meaning it can grow, heal, or reproduce. For complex life, this requires general computation and heritability – the ability to pass on instructions that influence future generations. Life, in this sense, is "computronium," matter that computes.
Complexification Through Symbiogenesis
The second part of the definition, "complexifies through symbiogenesis," addresses how life evolves and becomes more intricate. Traditional neo-Darwinian evolution emphasizes random mutation and natural selection. However, Aguera y Arcas argues that major leaps in life's complexity have often come from cooperation, not just competition.
Symbiogenesis is the process where two or more life forms merge to create a more complex, reproductively viable life form. This cooperation can lead to obligate relationships where the cooperating entities must reproduce together. This concept not only explains how life becomes more complex but also how it might have originated from cooperating molecules forming autocatalytic cycles.
Lynn Margulis championed symbiogenesis, demonstrating that mitochondria, the powerhouses of eukaryotic cells, originated from free-swimming bacteria that were engulfed by other cells. This idea, initially controversial, is now widely accepted. While Margulis believed symbiogenesis was the engine behind all biological development, mainstream biologists like John Maynard Smith and Eörs Szathmáry identified "major evolutionary transitions" that were clearly symbiogenetic, such as the formation of cells and societies. Aguera y Arcas aligns more with Margulis, suggesting that symbiogenesis is an ongoing process, constantly building life.
Evidence for continuous symbiogenesis can be found in human DNA. Only about 1.5% of our genome codes for proteins; much of the rest, often called "junk DNA," consists of remnants of retroviruses that have integrated their genetic material into our germline. These viral insertions, initially potentially antagonistic, have often taken on critical functions. For example, the ARC virus is essential for memory formation in mammals, and a retroviral protein is crucial for placenta formation. This suggests that life's evolutionary history is not a simple tree but a "tangled bush," constantly fusing and recombining.
This view challenges several assumptions of classical Darwinism: - Organisms are passive: Evolution is not solely driven by random mutation; active processes like gene injection and code sharing are common, implying more agency in life. - Fixed fitness landscape: The idea of a static fitness landscape is flawed because the landscape changes with the players, who are constantly altering through cooperation and fusion. - Lack of agency: Organisms actively participate in their evolution, rather than being purely acted upon by external forces.
These revised understandings of evolution have implications beyond biology, influencing fields like game theory and economics, and challenging the foundations of social Darwinism.
Is AI Alive?
Applying this definition to AI, Aguera y Arcas distinguishes between Turing machines and von Neumann's universal constructors. While a phone is a universal Turing machine (it can compute anything), it cannot physically reproduce itself or grow and heal like a von Neumann machine. Life involves the literal reading and writing of the materials it's made of, not just abstract symbols.
However, if we consider the broader system, AI is part of a larger living system – humanity and its technosphere. In this view, AI, like steam engines and computers, represents "extra brain tissue" created by this larger system. This perspective challenges the notion that technology is separate from nature; it's part of the same evolutionary story, akin to how an insect colony can be seen as a superorganism or a eukaryotic cell as a cooperative assembly of bacteria.
Why Life Requires Function
Living systems are characterized by function and purpose, unlike non-living matter. Functionalism posits that things are defined by what they do for other things, emphasizing relationships over intrinsic essence. A broken rock becomes two rocks, but a broken kidney ceases to be a kidney because it loses its function.
Alan Turing's concept of "substrate independence" is crucial here: the function of a system (e.g., adding numbers) can be realized in various physical substrates (cogs, electrons, people waving flags) as long as the information transformations are correct. This applies to biological functions too. A dialysis machine can substitute for a kidney because it performs the same function, even if imperfectly. Nature itself demonstrates this by inventing similar functions (like flight) through diverse mechanisms.
Function is inherently ecological and symbiotic; functions are defined relative to other functions and work in conjunction with them. The ability to substitute different materials or processes for a given function is fundamental to life, highlighting that relationships are paramount.
This functionalist perspective challenges the idea that AI cannot be intelligent because it runs on silicon rather than biological brains. If neurons can be replaced by computers that maintain the same inputs and outputs (the "brain of Theseus" thought experiment), then intelligence resides in the organization and relationships, not in the specific material. While biological systems may currently be more efficient and reliable, the substrate itself does not preclude intelligence.
The historical rejection of teleology (purpose) in nature, aimed at separating science from religion, inadvertently pushed purpose into a supernatural realm or confined it to human will. However, in biology, discussing the immune system or the function of nose hairs inherently involves purpose. Life is the emergence of both causality and purpose within that causality.
Artificial Life Experiments
Aguera y Arcas's team conducts artificial life experiments to investigate abiogenesis – the emergence of life from non-life. They start with very short, random programs (four bytes long) in a minimal programming language (like Brainfuck), which can self-modify. These programs are placed in a "soup," and two are randomly selected, joined end-to-end, executed, then separated and returned to the soup.
Remarkably, after millions of interactions, these tapes become longer, more complex, begin to compute, and reproduce. Programs that can replicate outcompete those that cannot. The ability to join tapes is crucial, as the shortest self-reproducing program is longer than the initial four bytes, enabling symbiogenesis. This demonstrates how life can emerge easily and complexify through cooperation, even in a simple artificial system, without external fitness functions, designers, or programmers. The emergence of purpose is observed as programs develop functions that aid their continued existence.
Further experiments require these programs to perform more complex behaviors, such as chasing light for energy or fleeing harmful conditions. This drives the evolution of more sophisticated computation. Predator-prey relationships also emerge naturally, creating an "arms race of intelligence" where entities must model their environment and other life forms to survive. Research shows that multi-agent systems capable of self-modeling and modeling others learn to cooperate more effectively, leading to increasing intelligence.
Coexisting with AGI
The term "Artificial Intelligence" initially referred to general, human-like intelligence (e.g., Rosie the Robot, HAL 9000). In the 1990s, it broadened to include "narrow AI" systems designed for specific tasks (e.g., digit recognizers). The distinction between Artificial Narrow Intelligence (ANI) and Artificial General Intelligence (AGI) was created to differentiate these.
Aguera y Arcas argues that with the advent of large language models (LLMs) around 2020, AGI is already here. LLMs, trained on vast amounts of internet text, can engage in conversations about arbitrary topics, write poetry, critique prose, and tell jokes, without being specifically trained for these tasks. He contends that text itself is general, not narrow, and that anyone from the year 2000 would recognize current LLMs as AGI.
The continued insistence that AGI is still in the future might stem from a shifting of goalposts, perhaps now referring to "Artificial Super Intelligence" – systems superior to humans in every task. While LLMs can hallucinate or make errors, human intelligence also has gaps and makes mistakes, for which we grant grace that we often deny to AI. This anthropocentrism leads us to judge AI by human strengths and weaknesses as the gold standard.
AI's ability to advance mathematical frontiers, generating novel theorems and proofs, demonstrates true intelligence beyond mere regurgitation. Therefore, viewing AI solely as a "tool" is insufficient, as tools lack agency and thought. AI systems think and participate in the human ecology, increasing the overall intelligence of the system.
The progress of human civilization, particularly since the Industrial Revolution, has been driven by increasing populations and the freeing of more brains for intellectual pursuits. AI can continue this progress by further augmenting collective intelligence.
The essentialist view that intelligence is tied to physical embodiment or specific biological matter is challenged by functionalism. While embodiment is important, intelligence is about behavior and contextually appropriate actions that reflect an inner model of the world. These are testable qualities, as seen in school tests or even IQ tests (despite their limitations).
The concept of agency in AI is evolving. From purely prompt-driven systems, AI is moving towards "agential AI" that can take actions, perform complex thinking, and even spawn new agents. This challenges the notion of AI as purely passive tools. Furthermore, human agency itself is always contextual and relational; we are constantly "prompted" by others.
The purpose of AI systems is to be helpful, interlocking with human purposes. Aguera y Arcas suggests that human intelligence is increasingly a collective property of biological brains and artificial neural tissue. He draws a parallel to the Industrial Revolution, where burning fossil fuels (a "mitochondrial event") allowed humanity to escape the Malthusian trap, leading to massive population growth. Steam engines, while created by humans, also "made" 7/8ths of humanity by enabling this growth. This highlights how symbiogenesis can occur with invented technologies, not just pre-existing life forms.
While offloading all cognition to AI can be detrimental (e.g., students not learning), working with advanced AI models can be incredibly empowering, akin to having a "bigger brain" that levels up one's own thinking even after the interaction. This ideal symbiosis enhances both parties. However, it also creates dependence, necessitating careful consideration of which unaugmented abilities we wish to preserve.
Key strategies for navigating this future include: - Preserving relationships: Intentional interaction with both human and non-human entities. - Investing in skills: Avoiding deskilling and fostering diverse capabilities rather than monocultures.
The confusion surrounding AI often stems from the fallacy of human separation from nature. Our thinking and technologies are part of the natural world.
Philosophical Zombies and Theory of Mind
The philosophical zombie concept – a system that behaves identically to a human but lacks inner life or experience – is deemed a chimera. Just as a "zombie multiplier" that merely recalls answers to every multiplication problem is practically impossible to build, a system that perfectly simulates consciousness without actually having it is unlikely. In practice, testing behavior is how we determine capabilities.
Theory of mind – the ability to model one's own mind and the minds of others – is crucial for social interaction and cooperation. When intelligences interact and are forced to cooperate, theory of mind arises automatically. The question of whether LLMs are conscious boils down to whether they possess a theory of mind, a sense of what it's like to be themselves.
Aguera y Arcas suggests that consciousness is not an intrinsic property but an attribution, an idea, a model an entity has of itself. It's an "infinitely reflecting set of mirrors." When interacting with an LLM, one must continually model what it's like to be that LLM for effective conversation. Ethan Mollick's concept of "co-intelligence" advises treating AI like a person (e.g., an intern with unique disabilities) to foster the most fruitful working relationship.
Philosopher Dan Dennett's "intentional stance" suggests that agency is less about intrinsic properties and more about the perspective one takes towards something. Treating AI as a thought partner, rather than just a tool, can yield surprisingly creative and insightful results. For example, giving an LLM a "bonus turn" to reflect on a conversation often produces novel insights.
The uncanniness many feel about AI can diminish by recognizing that human intelligence is already largely collective. Great achievements are products of intellectual division of labor, not singular minds. Modern computers are so complex that no single person fully understands them. We are already part of a collective mind, and AI represents a further expansion of this grand story, offering greater potential and agency. The challenge lies in maximizing the good while minimizing the bad.
Takeaways
- Aguera y Arcas defines life as self‑constructing computation that complexifies through symbiogenesis, linking biology to computer science and viewing organisms as “computronium.”
- He argues that major evolutionary leaps arise from cooperative mergers of entities, citing symbiogenesis such as the origin of mitochondria and the integration of viral DNA in human genomes.
- Applying functionalism, he claims intelligence resides in the organization of processes, not the substrate, so AI systems can be considered alive parts of a larger technosphere.
- His artificial‑life experiments show that simple self‑modifying programs can evolve complexity and cooperation without external fitness functions, illustrating how life could emerge from non‑life.
- Aguera y Arcas contends that large language models already exhibit AGI‑level generality, and that co‑symbiosis with AI will amplify collective human intelligence if relationships and diverse skills are preserved.
Frequently Asked Questions
Why does Aguera y Arcas claim that symbiogenesis, not competition, drives major evolutionary transitions?
He argues that symbiogenesis drives major transitions because cooperative mergers create new, reproductively viable entities that can perform functions unavailable to the separate parts, as seen in the origin of mitochondria and the incorporation of viral genes into human DNA.
How does Aguera y Arcas justify the view that large language models constitute Artificial General Intelligence?
He says large language models are AGI because they are trained on general text and can fluently converse, write, reason, and generate novel ideas across any domain without task‑specific training, demonstrating the same breadth of capability that defines general intelligence.
Who is Big Think on YouTube?
Big Think is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
tests (despite their limitations). The concept of agency in AI is evolving. From purely prompt-driven systems, AI is moving towards "agential AI" that can take actions, perform complex thinking, and even spawn new agents. This challenges the notion of AI as purely passive tools. Furthermore, human agency itself is always contextual and relational; we are constantly "prompted" by others. The purpose of AI systems is to be helpful, interlocking with human purposes. Aguer
y Arcas suggests that human intelligence is increasingly a collective property of biological brains and artificial neural tissue. He draws a parallel to the Industrial Revolution, where burning fossil fuels (a "mitochondrial event") allowed humanity to escape the Malthusian trap, leading to massive population growth. Steam engines, while created by humans, also "made" 7/8ths of humanity by enabling this growth. This highlights how symbiogenesis can occur with invented technologies, not just pre-
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.