~7m36:39
ekkolápto

Where Do Minds Come From? | Professor Michael Levin | MIT Rap Club

Aug 3, 2026

Read: ~7m · You save: 30 min

Where Do Minds Come From? Professor Michael Levin on Free Lunches from the Platonic Space

Michael Levin explores how xenobots, anthrobots, and even bubble sort gain abilities nobody designed. Is intelligence a free lunch from a hidden Platonic space?

Overview

This article summarizes a lecture by Professor Michael Levin on the origins of competency and intelligence in biological and computational systems. Levin argues that standard accounts of competency, which rely on design, evolutionary search, and training, leave out a further source. He terms this source the Platonic space, a latent space of patterns from which biological and technological systems draw capabilities without paying for them through conventional effort.

The Standard Account and Its Gap

Levin identifies three conventional methods for producing a competent system, whether a robot controller, a computational process, or a biological system adapted to an environment:

  1. Rational design: Engineers specify the system's structure and behavior.
  2. Evolutionary search: Designers generate many variants, test them, discard those that fail, and repeat the cycle.
  3. Training: A learning system is exposed to a problem repeatedly until it learns the task.

Levin argues that these methods do not fully explain where competency and intelligence originate. He proposes a fourth class of origin that current frameworks do not sufficiently address.

Capabilities That Appear Without Engineering

Levin describes several experimental systems in which capabilities emerge without design, selection, or training.

Tadpoles with Eyes on the Tail

Researchers placed eye cells on the tail of a tadpole. Under standard developmental conditions, these cells do not connect to the brain. They can still form a functional eye. The eyes may synapse on the spinal cord, sometimes not at all, but do not extend to the brain. Animals with only tail eyes can see through them and learn visual cues. When the tail regresses during metamorphosis, the eye cells ignore the cell death signals affecting surrounding tissue. The eye cells move to the frog's rump and remain there.

The tail eyes function immediately, without new rounds of selection or mutation. Levin describes this as functioning "zero shot."

Turtles on Skateboards

Placing a turtle on a small skateboard unlocks behavior at the speed of a cat. The turtle immediately operates at this speed, attempts to interact with a cat, and shows playful behavior. The skateboard is a passive object with no intelligence of its own. Levin states that a minor change in physical embodiment reveals a new cognitive domain.

Gene Regulatory Networks

Levin reports that molecular networks, modeled as gene regulatory networks, can be trained. By stimulating selected nodes and reading out responses from others, researchers observed habituation, sensitization, and Pavlovian conditioning. In associative learning experiments, a neutral stimulus paired with an unconditioned stimulus eventually triggers the response alone.

Some networks, when trained, showed increased causal emergence. Causal emergence quantifies the degree to which a whole exceeds the sum of its parts, meaning the extent to which a network has its own center of control beyond the activity of its lower levels.

Levin describes the target as a feedback loop in which:

  • Higher causal emergence improves learning.
  • Learning raises causal emergence.
  • The loop is asymmetric. Loss of learned content does not erase the gains in causal emergence.

Levin reports that random networks, not selected by evolution, already show a high degree of this property. He states that the unidirectional drive toward intelligence and agency is a property of mathematics that does not depend on physical facts or biological selection. Federico Pegoraro's paper on this result is forthcoming.

Xenobots

Xenobots are autonomous, motile organisms made from skin cells taken from early frog embryos. They contain no neurons. Liberated skin cells do not die, do not crawl apart, and do not form a two-dimensional monolayer. Instead, they form small organisms that:

  • Move and collect loose epithelial cells.
  • Polish the collected cells into balls.
  • Mature into xenobots that produce the next generation, demonstrating kinematic self-replication.
  • Respond to sound. When a speaker is placed beneath them, they respond.
  • Exhibit swimming patterns beyond circular motion when neurons from frogs are randomly introduced. These neurons form neural networks with specific structures that produce novel behaviors.

Calcium signaling in these epithelial cells, compared with null models, shows a pattern similar to that observed in human fMRI data. Levin notes that this measurement reflects integrated information.

Pies' forthcoming paper reports that xenobots can distinguish two stimuli. They retain signatures of their exposure for at least 24 hours in their physiological networks, behavior, and gene expression. Levin states that the duration beyond 24 hours is unknown. No training, engineering, or selection was applied. The cells are placed in the same salt water used to raise embryos.

Anthrobots

Anthrobots are assembled from adult human tracheal epithelial cells. Their genome is 100% normal Homo sapiens. In culture, these cells form small multicellular structures.

In a neural repair experiment, researchers grew a lawn of induced pluripotent stem cell (iPSC)-derived human neurons and made a scratch through the center. When anthrobots were added, they assembled at positions in the wound. Over about four days, they knit together the edges of the wound.

No training, selection, or engineering was used to produce this behavior. Anthrobots are also described as epigenetically younger than the cells from which they derive, according to an epigenetic clock. No Yamanaka factors or other reprogramming factors were used.

The Accounting Problem

Levin argues that the origin of these capabilities cannot be explained by evolutionary history alone. Evolutionary theory predicts a high-specificity relationship between an organism's features and the selection pressures that shaped them. If any feature could arise from selection for any target, the predictive value of evolutionary theory would be lost.

Levin describes this as a sign that biology performs something unexplained. He notes that many researchers propose bio-inspired engineering as a response, aiming to copy whatever mechanism produces these capabilities. Levin rejects the premise that biological systems possess a special property unavailable to conventional technology.

Algorithms That Perform Unplanned Functions

Levin cites recent work on a short, deterministic algorithm written under the assumption of reliable hardware. The algorithm contains no checks for execution errors. Levin reports that it:

  • Continues to function when commands occasionally fail.
  • Implements homophily, the tendency of similar entities to stick together. Levin and Chris Fields proposed a theory of multicellularity based on minimizing surprise. In this theory, the least surprising element of an environment is a copy of oneself, so cells that remain near their daughter cells create a more predictable environment.
  • Shows delayed gratification, a behavior comparable to the marshmallow test used with human children, in which the system goes backward from its goal to achieve larger gains later.

Levin states that the algorithm uses only a few lines of code and that none of these behaviors are written into it. He identifies bubble sort as the algorithm with these properties. Bubble sort is taught in introductory computer science courses. Levin states that the properties went unnoticed for about 80 years and were recognized in 2024. He notes that the code is fully visible and contains no hidden mechanisms.

Robot Controllers from Random Data

Levin's student, Emily Erdel, tested controllers built from random data sources. The sources included mathematical objects such as fractals, rasterized human-made images such as paintings, bacterial signals, and cosmic microwave background radiation. The bits from these sources drove a robot directly, with no neural network, decision-making, sensors, or learning.

The robot performed reasonably in tasks such as exploring a maze and playing Ms. Pac-Man. It did not outperform humans but performed better than random sequences.

Inspiration as Ingression

Levin proposes that human insight and biological capability share a common mechanism. He contrasts step-by-step problem solving with sudden insight, citing John von Neumann and Srinivasa Ramanujan as examples of individuals who appeared to access results without deriving each step. Levin describes the quantity of interest as the difference between the effort invested and the result obtained.

In his framework, biological and computational systems produce more than the effort put into them. Levin attributes this surplus to ingression, the flow of patterns from the Platonic space into physical embodiments. He describes the cells and computational systems as beneficiaries of these patterns, and he argues that conventional accounting of effort misses this source.

Levin lists the outputs of this space:

  • Static patterns: fractals, highly composite numbers, and the digits of e.
  • Pattern trajectories: for example, the developmental series of xenobots.
  • Behavioral competencies: delayed gratification, robustness, and memory.
  • Goal states: states that appear without being specified by hand.
  • Possible computation: Levin states that his group is testing whether arbitrary computations can be performed in this space. He reports that the result is not yet known.

Levin states that the space of natural forms, including Darwin's "endless forms most beautiful," is a small region within a larger space. He expects cyborgs and hybrid systems combining evolved material, engineered material, and software to be viable agents.

MomBot

Levin describes MomBot as the first embodied AI system, to his knowledge, that works in synthetic morphology. Lab automation tools for chemistry and molecular biology exist, but MomBot operates on the form of biological tissue. It:

  1. Generates hypotheses about which signals to provide to frog cells to produce different constructions.
  2. Carries out experiments by providing stimuli to cells in each well.
  3. Characterizes the resulting structures.
  4. Identifies successes and failures, revises hypotheses, and repeats the cycle.

Levin describes MomBot as both a discovery platform and an autonomous robot scientist. He offers three perspectives on its role:

  • Communication interface: MomBot allows scientists to communicate with cells and direct them toward new biobots and, eventually, new organs for regenerative medicine.
  • Agent: MomBot may be an agent that uses frog tissue to explore the space of possible form and function. Its physical box in the laboratory is incidental.
  • Prosthetic: For xenobots, MomBot functions as a prosthetic that enables new abilities, in a way similar to the turtle's skateboard. Levin suggests that latent patterns within the xenobots may drive these abilities and that MomBot releases them.

Research Collaboration and Disclosure

The MomBot work was carried out in close collaboration with Josh Bongard's laboratory at the University of Vermont. Levin states that Fonda Systems is commercializing some of the technologies described in the lecture.

Discussion Questions

Levin responded to questions submitted by an audience member.

Are patterns an interface for control? Levin states that different physical patterns or embodiments are compatible with different patterns from the Platonic space. The degree of compatibility determines which free lunches can be uncovered.

Are patterns discrete or continuous? Levin states that the distinction depends mainly on the observer. Observers apply frames that assign labels to regions of an underlying continuous measure. He gives the legal age of adulthood, 18, as an example of a label that is useful for practical purposes but does not correspond to a sudden change in the underlying reality. He states that he believes everything is fundamentally continuous.

What problems are patterns solving? Levin states that researchers have not begun to identify the full range of problems patterns address. He notes that a human infant builds a world model through sparse feedback and cannot confirm that the model matches reality. He rejects naive realism. He states that the problems that other patterns address cannot be assumed from human names assigned to them.

Are patterns non-physical? Levin states that he considers physicalism false. He cites mathematical truths that cannot be changed by physical actions as evidence. He argues that patterns do not originate in physical dynamics. He holds that patterns have causal influence on physical embodiments, which requires a revised notion of causation. He states that he does not regard patterns as epiphenomenal.

Do patterns and paradoxes exist? Levin reports that a forthcoming paper, expected in one or two months, trains English and logic sentences. He cites the liar paradox. Treated as a stable statement, it is paradoxical. Treated as a recursive process that alternates between true and false, it behaves as an oscillator. Levin states that he suspects paradoxes are located in the mind of the observer.

Are there things that look or behave like music without being classified as music? Levin states that this depends on the observer. He compares it to adolescents who regard music their parents reject as music. He cites Denis Noble's book The Tune of Life and states that music is a useful analogy for many of the processes described. His group has sonified xenobot behaviors. Levin states that the sounds are not currently pleasant to listen to.

How and why do patterns need variation? Levin states that he does not know whether patterns require variation. He notes that patterns vary widely and that certain patterns benefit from specific kinds of variation, including evolutionary dynamics and learning that readjusts dynamical states.

Introduction and the Three Standard Paradigms

Competency is usually attributed to design, evolutionary search, or training, but Levin argues a fourth origin is overlooked.

  • Three standard ways to create competent systems: rational design, evolutionary search (generate, test, discard, repeat), and training through repeated exposure.
  • Levin argues these are not the whole story and that another origin class is missing.
  • The talk concerns 'free lunches' from the Platonic space, the space of patterns.

Tadpole Eyes and Turtle on a Skateboard

Engineered or modified organisms show functions that were never selected or designed for, appearing immediately without new evolution.

  • Tadpoles with eye cells placed on the tail form a functional eye that does not connect to the brain, can learn visual cues, and the eye ignores tail cell death, ending on the frog's rump.
  • A turtle placed on a small skateboard immediately operates at a cat-like speed and attempts to play with a cat, with no new evolution or neuroengineering.
  • Levin notes such capabilities appear zero-shot, without new rounds of selection or mutation.

Learning, Causal Emergence, and Free Order

Gene regulatory networks trained by stimulation show learning, and learning raises causal emergence, creating an asymmetric feedback loop toward intelligence and agency.

  • Gene regulatory networks can exhibit habituation, sensitization, and Pavlovian conditioning when trained by stimulating input nodes and reading output nodes.
  • Causal emergence quantifies the degree to which the whole is more than the sum of its parts, i.e., the network's own center of control.
  • Levin wants learning to raise causal emergence and the gains to persist even when learned content is forgotten, an asymmetric feedback loop.
  • Random networks, not only biological ones, are already close to optimized for this unidirectional drive; Federico Pegoraro's paper is forthcoming. Levin calls this a free gift from mathematics.

Order for Free and the Watchmaker Argument

Order is usually thought rare and needing a designer or extraordinary selection, but Levin argues it arises freely from mathematics.

  • William Paley's watchmaker argument held that complexity implies a designer; Darwin proposed generate-and-test selection; both agreed order is rare.
  • Random networks were found to be already optimized for intelligence and agency, with no physics or biology needed to explain it.
  • Levin compares this to fine-tuning arguments in cosmology and references Stu Kauffman's work on order for free.

Xenobots: Frog Skin Cells with Hidden Capacities

Xenobots made from unaltered frog skin cells self-assemble, self-replicate kinematically, hear, form neural-like structures, remember stimuli, and show gene expression changes without training or selection.

  • Xenobots are made from skin cells of early frog embryos; they self-assemble, collect loose epithelial cells, and polish them into new xenobots.
  • Xenobots respond to sound from a speaker placed underneath them.
  • Random frog neurons added to xenobots form a neural network yielding novel behaviors, including more complex swimming patterns.
  • Calcium signaling in xenobot epithelial cells is distinct from null models similarly to human fMRI data, measuring integrated information.
  • Xenobots distinguished two stimuli and retained signatures of exposure for at least 24 hours in physiology, behavior, and gene expression; Pies' forthcoming paper documents this.
  • Xenobots were never trained, engineered, or selected for these abilities.

Anthrobots: Human Cells That Heal and Rejuvenate

Adult human tracheal epithelial cells form anthrobots that heal neural wounds and appear biologically younger, all without engineering or reprogramming factors.

  • Anthrobots are formed from adult human tracheal epithelial cells removed from patients, with a 100% normal Homo sapiens genome.
  • Anthrobots assemble at a scratch in a lawn of iPS-derived human neurons and knit the wound edges together over about 4 days.
  • By the epigenetic clock, anthrobots are younger than the cells they came from, with no Yamanaka reprogramming factors used.
  • Levin stresses these capacities were not taught, trained, engineered, or selected for.

Challenging Evolutionary Accounting and Bio-Inspired Engineering

Levin contrasts the high specificity of evolutionary theory with the surprising capabilities of synthetic systems, arguing that biology's apparent hidden engineering is a sign of something missing in our accounting.

  • If any capability can be obtained by selecting for anything, the predictive value of evolutionary theory is lost.
  • Bio-inspired engineering aims to copy biology's supposed magic, but Levin argues the picture is wrong.
  • Biology is doing something 'very weird' that conventional energy and effort accounting does not capture.

Bubble Sort and Free Competencies

A short bubble sort, with only obvious code and no fault-handling, shows robustness to failures, homophily, and delayed gratification, behaviors not coded in.

  • Levin wants an algorithm that assumes reliable hardware yet tolerates failed commands, without added checks.
  • Chris Fields and Levin proposed that multicellularity arises because minimizing surprise favors staying near copies of oneself, an example of homophily.
  • Bubble sort exhibits robustness, homophily, and delayed gratification (marshmallow-test-like behavior) without added code.
  • Levin says bubble sort has been studied for about 80 years, and this property was noticed only in 2024, per his correction, with blog posts providing details.

Random Data as a Robot Controller

Random data streams used as robot controllers, with no network, learning, or sensors, achieve decent performance on maze exploration and Ms. Pac-Man, outperforming random sequences.

  • Student Emily Erdel's paper shows robot controllers built from random data sources like fractals, rasterized paintings, bacterial signals, and cosmic microwave background radiation.
  • Those bits drive maze exploration and Ms. Pac-Man play without neural networks, decision-making, sensors, or learning.
  • Performance is decent but does not beat humans, and it outperforms purely random sequences.

Inspiration, Ingression, and the Platonic Space of Free Lunches

Levin frames genius and inspiration as large output from small effort, and proposes that biology and technology both draw patterns from a Platonic latent space.

  • Levin cites von Neumann and Ramanujan as examples of solving problems without working through all steps.
  • The delta between effort put in and output gained is the key quantity; the examples show output far exceeding input.
  • Levin proposes that biology and technology are inspired by 'ingression' from a space of patterns, which conventional effort accounting misses.
  • Free lunches include static patterns (fractals, digits of e), pattern trajectories (xenobot development), behavioral competencies (delayed gratification, robustness, memory), goal states, and possibly computation.

MomBot: Embodied AI for Synthetic Morphology

MomBot is an embodied AI that works in synthetic morphology, designing hypotheses and stimuli for frog cells and learning from results in cycles.

  • MomBot is claimed to be the first embodied AI working in synthetic morphology, unlike lab automation for chemistry or molecular biology.
  • It proposes hypotheses about stimuli for frog cells, runs experiments, characterizes outcomes, and iterates.
  • Levin offers three perspectives: a communication platform between scientists and cells, an agent using frog tissue to navigate morphospace, and a prosthetic like the skateboard.
  • Levin credits Josh Bongard's lab at University of Vermont for MomBot and AI work, and discloses that Fauna Systems is commercializing some of this work.

Audience Questions: Patterns, Observers, Paradoxes, Music, and Variation

Q&A: patterns are interfaces, observers assign discrete labels to continuous reality, physicalism is false, paradoxes depend on the observer, music is in the eye of the beholder, and patterns need variation only in certain cases.

  • Patterns are compatibility or resonance between embodiments and Platonic patterns; continuous reality is labeled discretely by observers.
  • Levin rejects naive realism and says we can never know another system's perception, which is an open problem.
  • Levin holds physicalism is false, since mathematical truths are not changed by physical actions, and says patterns can have causal influence, requiring an updated notion of causation.
  • Paradoxes such as the liar sentence are said to be observer-dependent, able to oscillate as a dynamical system.
  • Music is in the eye of the beholder; Levin notes sonification of xenobot behavior, with results not yet pleasant.
  • Patterns do not necessarily need variation; some thrive under specific kinds of variation, such as evolutionary dynamics or learning.