Michael Levin discusses biological organisation, regeneration and intelligence in unfamiliar systems, then presents his broader account of minds and mathematical patterns. Eleven research notes distinguish controlled experiments from proposed interpretations, including cellular robots, memory through metamorphosis, sorting algorithms, molecular aging clocks and regulatory-network simulations. The complete Discussion map also follows the debate about physical explanations, collective goals, cancer research and scientific creativity.
HostLex Fridman
GuestMichael Levin
About these timings. Publisher-supplied paragraph starts explicitly linked to the YouTube recording. Paragraph ends are unknown and matching words may occur later. Duration comes from the official YouTube listing badge for this video, independently inspected by the lead author. Original repeated starts and transcription irregularities are retained. No audio audit is claimed.
Levin proposes comparing systems by the kinds of interventions that change their behaviour.
Background and context
Levin’s March 2022 TAME article presents a framework for investigating cognition across biological and engineered systems. Its proposed scale runs from direct physical modification toward training and communication. The operational question is which approach most effectively predicts and changes a particular system’s behaviour. Composition or evolutionary origin alone does not settle that question in the framework.
This is a useful way to turn a broad philosophical disagreement into experiments with specified interventions, outcomes and comparisons. It is also important to preserve the article’s status: the journal labels it a hypothesis and theory contribution. It does not demonstrate that every physical system learns, feels or understands an argument. Showing that a behavioural description helps control one system would support that application; a claim about subjective experience would require an additional argument.
Intervention success does not by itself establish conscious experience.
Levin describes regeneration as collective activity directed toward an anatomical outcome.
Background and context
Durant and colleagues’ May 2017 planarian study provides a concrete example of persistent changes in regenerative pattern. A temporary intervention in electrical communication altered the outcomes of later amputations. Some animals regenerated with two heads; others appeared normal but retained an altered probability of producing two heads after another cut. Subsequent regeneration occurred without repeating the original treatment.
The authors interpret these results as evidence for physiological storage of patterning information. The striking feature is that present appearance did not always reveal the altered regenerative state. This supports investigating how cell communication helps stabilise anatomical outcomes. It does not establish that a worm consciously imagines a target body, nor that the same intervention could regrow a human limb. The experimental result and the cognitive vocabulary used to describe it remain distinguishable.
Planarian patterning is not a demonstrated human regeneration treatment.
Levin describes Xenobots that gather cells and Anthrobots made from adult human airway cells.
Background and context
Two experiments underlie these examples. The 2021 Xenobot study showed that moving assemblies of frog cells could collect supplied, dissociated cells into additional functional assemblies. Replication depended on living starting material and the laboratory environment; it was not the creation of life from nonliving chemicals. Computer-assisted designs extended the number of replication rounds.
The Anthrobot paper, first published in November 2023, used adult human bronchial cells that formed mobile structures under selected culture conditions without introducing genetic modifications. The authors also reported repair across a scratch in a cultured neural layer. That is an experimental observation in a dish, not demonstrated recovery from a human brain injury. Together, the studies show how changing cellular arrangements and environments can uncover capabilities beyond the cells’ usual roles in an intact animal.
Both results depend on controlled laboratory preparations.
Levin uses memory across metamorphosis to illustrate continuity despite extensive bodily change.
Background and context
Blackiston, Silva Casey and Weiss tested tobacco hornworm caterpillars in a study published in March 2008. Pairing an odor with electric shock late in larval development produced avoidance that persisted in the adult moths. Controls addressed exposure to odor alone, shock alone and chemical residues on the pupal case. Training at an earlier larval stage did not produce the same adult retention.
These details refine the conversation’s butterfly and coloured-leaf example: this experiment concerned moths and an odor association. It also did not establish that a completely destroyed brain recreated memory from nowhere. The authors discuss persistence and remodelling of neural structures as possible explanations, while leaving the specific storage mechanism unresolved. The result is evidence for retention through substantial developmental transformation, with a meaningful dependence on when training occurred.
The experiment does not locate the precise memory substrate.
Levin connects his pattern-space proposal with the Platonic Representation Hypothesis in machine learning.
Background and context
Huh and colleagues’ 2024 ICML position paper asks whether independently trained models increasingly represent data in similar ways. Alongside a survey, it compares representations across image and language models and reports greater alignment as models become more capable. The proposed destination is a shared statistical representation of reality, expressed through relationships among data points.
That hypothesis is narrower than the claim that minds exist in a separate nonphysical realm and enter biological bodies. Similar internal representations can motivate questions about constraints imposed by shared data, tasks and the world without deciding that metaphysical issue. The paper itself discusses limitations and counterexamples rather than presenting universal convergence as established. Its relevance here is a concrete research comparison: investigators can measure alignment among specified models and test when the proposed trend does or does not hold.
Representational alignment does not establish a nonphysical source of minds.
Levin argues that mathematical and cognitive patterns may supply causal structure beyond physical descriptions.
Background and context
Levin’s later Philosophies article, published in September 2026 after this conversation, develops the position under the title Ingressing Minds. It explicitly argues for causal contributions from nonphysical patterns and acknowledges connections to dualist and panpsychist positions. Its proposed route forward is to investigate how different physical embodiments relate to patterns of cognition and behaviour.
This helps identify the strength of the claim being made. It is more than the familiar observation that mathematical models describe biological systems, and it is not an experimental demonstration that consciousness arrives from another realm. For an empirical comparison, a proposed mapping would need specified predictions and a way to distinguish its successes from competing accounts. The biological experiments discussed elsewhere in this episode can motivate that programme without uniquely establishing its metaphysical interpretation.
The cited article postdates the interview and presents a philosophical argument.
Levin describes sorting systems that temporarily reduce their sortedness while navigating defects.
Background and context
Zhang, Goldstein and Levin’s paper, first published online in August 2024, studies variants of sorting systems as a minimal model of biological organisation. The investigators changed two assumptions: elements could apply local sorting rules, and some elements could be unable to execute their moves. They also examined arrays containing elements governed by different algorithms.
Reported behaviours included temporary decreases in sortedness while navigating a defect and transient clustering among elements using the same rule. Those are measured properties of the implemented system. Terms such as delayed gratification or preference add an interpretation of the trajectory; they do not show that the elements consciously wait or desire company. Levin also explicitly acknowledges in the conversation that the code is followed. Unexpected higher-level patterns can invite useful analysis without supplying an exception to the underlying mechanics.
These computational observations do not establish subjective experience.
Levin says Anthrobot formation changed an epigenetic estimate of the cells’ age.
Background and context
The June 2025 Anthrobot lifecycle paper measured DNA methylation in cells from a 21-year-old donor, comparing starting cells with cells collected after formation of the mobile structures. The estimates were lower in the Anthrobot samples. This is a specific molecular-clock result, alongside the paper’s observations of movement, structural decline and transcriptional changes.
The figure caption and main text report somewhat different baseline and early-sample means, so a precise percentage reduction should not be treated as a uniquely established value. More fundamentally, a clock estimate is not a direct measurement of every aspect of cellular aging. This experiment did not rejuvenate a person, show an extension of human lifespan or establish durability in a patient. It supports further investigation of how cellular state and organisation affect the measured methylation pattern.
The clock experiment used cells from one donor; internal numerical differences remain visible.
Levin describes an association between learning and causal integration in chemical-network models.
Background and context
Pigozzi, Goldstein and Levin’s July 2025 study simulated 29 regulatory-network models drawn from a biological-model database. It tested combinations of input and response nodes for an associative-memory task, then analysed the qualifying circuits with Integrated Information Decomposition. The methods report 808 qualifying circuits drawn from 19 of the networks, rather than every tested combination succeeding.
The authors found distinct patterns in how their measure changed around training. This supports studying relationships between learned responses and integration in these specified dynamical models. It does not measure consciousness or directly demonstrate the same training effect in living tissue. A February 2026 correction, published after the interview, fixed interchanged titles and legends for two figures. Consulting that correction preserves the distinction between a repaired presentation error and a claim that the study’s entire result was withdrawn.
The result depends on model dynamics and the chosen information measure.
Levin describes altering cell communication to restore coordinated tissue behaviour in cancer-related experiments.
Background and context
Chernet and Levin’s May 2014 study provides one concrete example of bioelectric effects on tumour development. Researchers introduced human oncogenes into Xenopus tadpoles and observed tumour-like structures. Changing electrical properties of cells at distant sites reduced the incidence of those structures. The experiments investigated ion-channel interventions and a possible connection through histone-deacetylase regulation and the local biochemical environment.
This supports the idea that surrounding tissue physiology can influence oncogene-driven growth in that model. Its endpoint was not a demonstrated cure for human cancer, and reduced incidence is different from safely reversing an established malignancy in a patient. The broader language of restoring collective goals should therefore remain a research interpretation. Translating such work would require identifying reproducible mechanisms, appropriate delivery and effects across relevant tumour types, followed by evidence for safety and clinical benefit.
A tadpole tumour model does not establish a human treatment.
Levin describes walking and photographing nature as ways to support his creative process.
Background and context
Oppezzo and Schwartz’s study, first published online in April 2014, compared performance on creative tasks while participants sat or walked. Across four experiments, walking benefited divergent tasks, including generating alternative uses and creative analogies. The effect also appeared indoors on a treadmill, and some benefit persisted shortly after walking ended.
The results did not generalise equally to a task requiring a single convergent answer. That distinction fits a limited interpretation of Levin’s routine: movement may help with generating possibilities, while evaluating them and completing experiments are different activities. The study did not test his photographic practice, establish the source of scientific inspiration or measure the eventual quality of research publications. It supplies an experimentally examined component of a creative routine without turning a personal anecdote into a universal recipe for discovery.
Short task performance is not a measure of eventual scientific discovery.
Remember a word from the conversation? Find an approximate listening time without reading through a transcript.
Transcript paragraph starts; matching words may occur later. All times are approximate. Use up to six words; matching uses 30-second windows, not exact phrases. Transcription can be wrong, so listen to check. Searches stay in your browser.
PodCompanions