Adam Marblestone and Dwarkesh Patel discuss evolved brain architecture, learning objectives, cellular diversity and what neuroscience might contribute to AI. Eight research notes examine specific experiments and theoretical proposals, distinguishing measured results from engineering forecasts. The conversation also covers connectomes, formal verification, scientific infrastructure and the limits of current explanations of cognition.
HostDwarkesh Patel
GuestAdam Marblestone
About these timings. Publisher-supplied VTT cue times for the official audio. Automated captions contain recognition errors; Discussion summaries use clear passages checked against the publisher’s editorial transcript. No independent audio audit or cross-platform offset is claimed. Duration is rounded up from publisher metadata.
Marblestone proposes that evolution shaped many specialised learning objectives and architectures, while acknowledging that their details remain uncertain.
Background and context
Marblestone, Wayne and Kording’s September 2016 article proposed three connected hypotheses: that brains optimise cost functions, that these objectives differ across regions and development, and that learning operates within structured architectures. The authors brought ideas from machine learning into conversation with evidence about neural circuits and suggested ways to refine and test the framework.
This is a theoretical synthesis and research proposal. It does not identify a complete list of biological objectives, demonstrate one universal credit-assignment mechanism, or show that every evolved behaviour can be reproduced by adding a term to an artificial system’s training loss. That distinction explains the interview’s repeated uncertainty. The framework can organise questions about innate circuitry and learned representations without resolving them. It is useful to distinguish the proposed language for explaining a brain from the experimental measurements needed to establish a particular mechanism.
The cited paper advances hypotheses and possible tests; it does not decode a complete biological learning algorithm.
Marblestone cites a model of early visual cortex as an example of neuroscience contributing useful constraints to artificial networks.
Background and context
Dapello and colleagues’ 2020 NeurIPS paper placed a fixed model of primary visual cortex in front of several image-classification networks. Its components represented simple and complex cell responses together with neuronal variability. The researchers tested the resulting systems on image-recognition benchmarks involving adversarial perturbations and common image corruptions. Their experiments found improved robustness and examined how different components contributed.
This supports a concrete example of biological inspiration helping with a specified engineering problem. It does not establish that the complete visual system has been explained, that the model learns as a brain does, or that inserting one cortical component supplies general intelligence. Benchmark design and the chosen network architectures remain part of the result. Marblestone’s broader suggestion that neuroscience could guide AI therefore has an empirical example here, while its scale and eventual importance remain open questions.
Robustness on image benchmarks is not a demonstration of general intelligence or a complete model of vision.
Marblestone discusses molecularly diverse cell types as possible clues to the brain’s inherited architecture.
Background and context
Yao and colleagues’ December 2023 atlas combined single-cell RNA sequencing with spatial measurements across the adult mouse brain. Approximately four million sequenced cells passed quality control, alongside about 4.3 million cells measured with MERFISH. The resulting hierarchy included 34 classes and 5,322 clusters, with substantial differences in how cell types were organised across brain regions.
These are classifications based on measured molecular and spatial features. They provide a resource for investigating circuits, development and evolution, but do not by themselves assign a reward function or behavioural purpose to every cell type. Nor is an adult mouse atlas a complete human wiring diagram. The distinction matters when Marblestone uses cellular diversity to motivate hypotheses about specialised innate systems. The data make that diversity concrete; the proposed link from cell identity to a particular computational objective still requires functional experiments.
Transcriptomic clusters are not demonstrated learning objectives, and mouse findings do not automatically establish human circuit function.
Marblestone connects reinforcement-learning ideas with dopamine signals and contrasts cached action values with model-based planning.
Background and context
Schultz, Dayan and Montague’s March 1997 paper connected behavioural learning experiments, primate dopamine recordings and quantitative reinforcement-learning theory. Its central interpretation was that fluctuations in some dopamine-neuron responses could signal discrepancies between expected and obtained rewarding events. This supplied a bridge between an experimentally observed pattern and a computational account of learning.
The paper is a synthesis and theoretical interpretation of physiological findings, rather than proof that one signal explains all motivation or every kind of learning. A reward-prediction error is also different from simply measuring pleasure: its meaning depends on expectations and changes in those expectations. That narrower distinction helps with the interview’s comparison between habitual action selection and planning. The comparison remains a model of interacting processes, not evidence that an entire brain implements a single textbook algorithm or that all dopamine neurons have identical roles.
Theoretical correspondence with physiological patterns does not establish one complete or uniform brain-wide learning mechanism.
Marblestone uses learned timing in a Purkinje cell to illustrate why cellular computation may matter beyond adjustable connection strengths.
Background and context
Johansson and colleagues’ September 2014 experiment trained cerebellar Purkinje cells in ferrets and measured learned patterns of firing over hundreds of milliseconds. They stimulated immediate inputs directly and used pharmacological interventions to examine whether temporally patterned inputs or surrounding interneurons were necessary. Their results led them to locate the timing mechanism within the Purkinje cell itself.
The experiment challenges a particular account in which the timing must already be present in the incoming activity pattern. It does not demonstrate that all memories reside inside individual cells, make synaptic plasticity irrelevant, or identify the complete molecular mechanism behind every learned response. The authors discuss a specialised form of learning with an unusually clear timed output. Marblestone’s example is therefore useful evidence that biological implementation can add computational detail, while its wider implications remain questions to investigate.
A specialised ferret conditioning result is not a universal account of memory or a complete molecular explanation.
Marblestone discusses optical and molecular approaches to mapping brains while distinguishing a wiring diagram from functional measurements.
Background and context
Chen, Tillberg and Boyden’s January 2015 expansion-microscopy paper demonstrated a way to enlarge labelled specimens physically inside a swellable polymer. This separated labels that conventional optics could not otherwise resolve. The researchers reported approximately 70-nanometre apparent lateral resolution in cultured cells and brain tissue, including three-colour imaging of a volume of mouse hippocampus with a conventional confocal microscope.
This was an experimental imaging advance. It was not a completed whole-brain connectome, a measurement of every connection’s effective strength, or validation of later cost estimates for mapping a human brain. Sample preparation, labelling, tracing and interpretation remain additional problems. The paper provides a concrete foundation for the interview’s engineering optimism without making the forecast itself a result. Marblestone’s distinction between anatomical information and activity measurements is essential when considering what a future map might explain.
Resolution in prepared samples does not establish complete tracing, functional weights, whole-brain scale or forecast costs.
Marblestone explains that verified software can still fail to capture the real-world property its designers intended to require.
Background and context
Klein and colleagues’ October 2009 seL4 paper described machine-checked verification of an operating-system microkernel from an abstract specification to its C implementation. The stated result was functional correctness: the implementation follows the specified kernel behaviour. The paper explicitly assumed that the compiler, assembly code and hardware behaved correctly.
Those boundaries are part of the accomplishment, not incidental footnotes. A proof relating implementation to specification does not establish that the specification expresses every desired safety property, that an entire deployed system meets it, or that assumptions about the surrounding machinery always hold. This provides a concrete precedent for Marblestone’s discussion of formal methods and AI. A machine can check a carefully defined claim while the task of choosing the right claim remains substantive. The 2009 result does not validate his broader forecast for automated mathematics or future AI safety.
The theorem is conditional on its specification and stated assumptions; it is not a guarantee about every property of an entire system.
Marblestone describes the Gap Map as a way to identify scientific infrastructure projects that existing institutions may struggle to support.
Background and context
The April 2025 launch announcement by Marblestone, Mary Wang and Joseph Fridman described the first Convergent Research Gap Map as an illustrative survey. It drew on conversations with hundreds of scientists and technologists and assembled roughly a hundred gaps with a few hundred proposed capabilities. The authors explicitly called it preliminary and noncomprehensive, rather than a ranking of where funding should go.
That method can surface neglected ideas without establishing a representative census of scientific needs or measured returns on every proposed investment. The announcement also says focused research organisations are only one possible institutional response, alongside academic programmes, companies, prizes and other mechanisms. These qualifications matter for the closing discussion. The map records a programme’s informed proposals and invites criticism; the existence of an entry is not independent evidence that its technology will work, its budget will suffice or its benefits will follow.
This is the organisation’s preliminary survey and proposal, not an independent evaluation of outcomes or a comprehensive priority ranking.
Remember a word from the conversation? Find an approximate listening time without reading through a transcript.
Times follow the Publisher audio edition. 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