The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence

Jeff Hawkins; Niels Leadholm; Viviane Clay

REVIEW 3 major objections 6 minor 2 cited by

The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes that a single repeating sensorimotor unit, the learning module, which models the world through spatial reference frames and embodied interaction, is sufficient to build general intelligence, and presents Monty as an…

desk verdict A clear, honest engineering white paper for the Thousand Brains sensorimotor architecture, with real code and no measurements; the universal-capability claim remains an act of faith. read the letter →

arxiv 2412.18354 v1 pith:RX2JYFMK submitted 2024-12-24 cs.AI q-bio.NC

Viviane Clay , Niels Leadholm , Jeff Hawkins This is my paper · ORCID

classification cs.AIq-bio.NC

keywords sensorimotorintelligenceneocortexreferenceframesworldmodelslearningmodulecorticalmessagingprotocolembodiedAIgeneral

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that general intelligence does not require scaling up deep neural networks but can be built from a single repeating computational unit: the learning module, modeled on the neocortex's cortical columns. Each module learns complete object models in explicit spatial reference frames from sensorimotor interaction, and modules communicate through a common protocol. The authors present Monty, an early open-source instantiation of this 'thousand-brains' architecture, and claim that the same mechanism should eventually implement any capability the human neocortex has, including abstract concepts, language, and theory of mind. If correct, this would offer an alternative path to AI that learns quickly, continuously, and without large datasets.

What carries the argument

The central object that carries the argument is the learning module: a repeating computational unit modeled on the cortical column, which represents every object as a graph of features anchored to poses in a spatial reference frame. The module updates an evidence score for each object-and-pose hypothesis as the sensor moves, using displacements to map new observations into the object's reference frame. The second load-bearing mechanism is the Cortical Messaging Protocol (CMP), the standardized feature-at-pose message that lets sensor modules, learning modules, and motor systems interoperate, so that voting across modules and hierarchical composition both reduce to the same message format.

What would settle it

Demonstrate a neocortical capability that cannot be expressed as features at poses in a reference frame—such as a purely relational, non-spatial abstract concept—or run Monty on the theory-of-mind and language tasks it lists in Section 2.4 and show that the same learning module used for object recognition cannot learn them; either result would refute the claim that one sensorimotor mechanism underlies all intelligence.

Watch

Extended reading notes

Core claim

This paper sets out to establish that the neocortex's organizing principle—thousands of semi-independent cortical columns, each learning complete models of the world in spatial reference frames through sensorimotor interaction—can be lifted out of the brain and used as the blueprint for a general-purpose AI. Its central claim is that the same repeating learning module, using features-at-poses and movement, can recognize objects, infer their orientation, act on the world, and eventually represent abstract concepts and other minds. The paper presents Monty, a first-generation implementation of this architecture, as evidence that the principles are concrete enough to build and test, and it argues that learning from a few sensorimotor interactions, rather than from huge static datasets, is the fundamental advantage this design offers.

Load-bearing premise

The framework rests on the assumption that a single repeating sensorimotor algorithm using explicit coordinate systems is enough to produce every kind of intelligence, including abstract concepts, language, and theory of mind; if some capability of the neocortex cannot be represented as features at poses in a reference frame, the central claim fails.

Editorial extensions

If this is right

AI built this way would learn a new object from a handful of movements instead of millions of labeled examples, because structure is stored explicitly in reference frames rather than in weights.
Learning and inference would become a single continuous process, with the system updating its world models during ordinary operation rather than in a separate training phase.
All sensory modalities—touch, vision, hearing—would run the same learning algorithm, with multimodal integration handled by voting and a shared protocol instead of modality-specific architectures.
The same models used for perception would also support planning and manipulation, because a learned graph can be queried both for 'what is here?' and for 'what action leads to the desired state?'

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

For this framework to cover all of intelligence, abstract concepts must inherit the geometry of concrete space; the most informative experiment will be to see whether a learning module can build a stable reference frame for something like a family tree or a melody—and whether that frame genuinely aids reasoning the way spatial frames aid recognition.
A quantitative prediction follows from the design: an agent that actively chooses informative next observations (via the hypothesis-testing policy) should recognize objects in fewer steps than one that moves randomly, since the graph-mismatch procedure is supposed to target the most disambiguating points; measuring that step-count gap would test whether the model-based policy is doing real epistem
The architecture implies that pose information is the premium commodity for intelligence: a cheap sensor that delivers accurate poses may outperform a rich sensor that does not, since the learning module's power comes from knowing where each feature sits relative to the body and to the object.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Katy —

Bottom line: this is a useful architecture and implementation white paper, not an empirical research result. The genuinely new content is the Monty codebase, the Cortical Messaging Protocol, and the detailed description of the evidence-based learning module. The conceptual material—reference frames, sensorimotor learning, voting, cortical columns—is largely a restatement of Hawkins et al. 2017/2019 and Lewis et al. 2019. The paper is transparent about that lineage and about being early stage, which I appreciate.

What it does well: the CMP is a sensible interface abstraction, and the buffer/graph-memory/evidence-update design is described in enough detail that someone could reimplement it from the text and the code on GitHub. The hypothesis-testing policy is a nice concrete example of model-based control. As a design document for the project, it is coherent and readable.

Now the soft spots, in proportion. Section 5 is titled “Experimental Evaluations” but contains zero evaluations. The surrounding sections describe a testbed (Habitat, YCB objects, episodes/epochs) but no accuracy numbers, no baselines, no ablations. That means the paper’s central claim—that this architecture is “uniquely suited” to learn a wide range of tasks and “eventually implement any capabilities the human neocortex has”—is unsupported by anything in the paper. The stress-test note gets this right: the implemented evidence LM is tied to 3D geometric morphology. The evidence updates use point normals, principal curvature directions, 3D displacements, and nearest-neighbor graph lookups (§9.9–§9.12). The paper itself defers abstract spaces to future work (§11.5.2) and says the exact structure of space “can potentially be learned” (§3.3). So the universality claim is not just missing benchmarks; the architecture as specified has no concrete construction for non-spatial domains. If the authors want to claim language or abstract concepts, they need at least a worked example showing how a family tree or logical relation would supply the geometry that the evidence update requires.

Also worth noting: the design is explicitly derived from the authors’ own Thousand Brains Theory, so until independent groups reproduce results or the authors publish controlled comparisons, circularity is a real concern. That said, this is a legitimate research program and the open-sourced code gives others a way to test it.

Who is this for? People working on embodied AI, world models, or neuro-inspired architectures should read it as a proposal and a code release. It does not deserve to be treated as a demonstration of the claims. For peer review: if a venue accepts white papers or position papers, send it to a referee with a request that the authors add at least one quantitative evaluation and a concrete abstract-domain construction. My own verdict would be skeptical until those appear.

Referee Report

3 major / 6 minor

Summary. The paper advocates a new paradigm for AI, 'Thousand Brains systems', based on the authors' Thousand Brains Theory of the neocortex. It specifies a modular architecture in which sensor modules, learning modules, and motor systems communicate through a common 'Cortical Messaging Protocol' (CMP). Each learning module builds structured, reference-frame-based models from sensorimotor experience; the first implementation, Monty, uses explicit 3D Cartesian graphs with an evidence-based matching procedure. The paper describes the architecture in detail, including the evidence update rule, voting between learning modules, hierarchical and lateral connectivity, and several action policies. It also provides a link to open-source code. The central claims are that this architecture is 'uniquely suited to quickly learn a wide range of tasks' and can 'eventually implement any capabilities the human neocortex has', including abstract concepts and language. The section titled 'Experimental Evaluations' (Section 5) contains no experiments, metrics, or comparisons, and the paper's own text repeatedly defers abstract spaces and hierarchical policies to future work.

Significance. If the capabilities claimed in the Abstract and Section 2.4 were substantiated, this would be a significant alternative to deep-learning-based embodied AI, with a concrete, open-source implementation and a clear set of design principles. The manuscript is valuable as a detailed architectural proposal and as a reference for the Monty codebase: it provides a sufficiently precise description of the CMP message format, the evidence-based learning module, voting, and hypothesis-testing policies that others could reimplement or extend the system. Those strengths are real. However, the paper currently offers no quantitative evidence for its claims, and the strongest claims about universal neocortical competence rest on an unspecified extension from 3D sensorimotor reference frames to abstract domains. The significance of the contribution therefore cannot yet be assessed at the level asserted by the authors.

major comments (3)

[Section 5, 'Experimental Evaluations'] Despite the section title, Section 5 contains no experimental results, no metrics, no benchmark comparisons, and no evaluation of any kind. It only describes the Habitat testbed, time discretization, and the observation that an episode can end on recognition, no-match, or timeout. This matters because the Abstract and Section 1 claim that Monty is 'uniquely suited to quickly learn a wide range of tasks' and Section 2.4 says it 'should be able to solve any task the neocortex can solve.' Without at least a basic object-recognition or pose-estimation evaluation on the stated YCB dataset, the central empirical claims are unsupported. The authors should either add a quantitative evaluation or explicitly reframe the paper as a position/architecture paper whose capability claims are hypotheses.
[Sections 1, 2.4, 3.3, 11.5.2 vs. Sections 9.9–9.12] The universality claim—that all aspects of intelligence, including abstract concepts and language, are produced by the same sensorimotor reference-frame mechanism—is not backed by any construction in the implemented system. The evidence-based LM described in Sections 9.9–9.12 updates evidence using point normals, principal curvature directions, 3D displacements, and nearest-neighbor matches in explicit 3D Cartesian graphs. No mechanism is specified for mapping an abstract domain (e.g., a family tree, logical relations, or language) onto the CMP's 'features at poses' or for supplying the geometric quantities that the evidence update requires. Section 3.3 says only that the exact structure of space 'can potentially be learned,' and Section 11.5.2 defers abstract spaces to future work. The paper should either provide a concrete example of such a mapping or substantially narrow the claims in the Abstract and Section 2.4.
[Sections 9.10, 10.1, 10.3, 11.3.1] The described system depends on several user-set parameters—percent-threshold, feature tolerance, number of sampled rotations N, momentum parameter alpha, and the minimum number of LMs that must reach a terminal state—yet the paper provides no sensitivity analysis, ablation, or experimental justification for any of them. In particular, Section 9.2 asserts that the evidence-based LM is 'most robust to noise and sampling new locations,' but no evidence for this robustness is presented. At a minimum, the authors should report how performance varies with these parameters, or qualify the robustness claim as a design expectation rather than an empirical finding.

minor comments (6)

[Section 2.1] The phrase 'can can vary in implementation detail' contains a duplicated word that should be corrected.
[Section 3.1] The sentence ending 'represent the systems interface with the environment.Fin a common reference frame' appears to be a truncated join of two sentences; it should be rephrased for clarity.
[Section 10.2] The heading 'V oting with Evidence' contains an erroneous space and should read 'Voting with Evidence'.
[Section 10.4.1] The phrase 'there can exit ‘skip connections’' should read 'there can exist ‘skip connections’'.
[Section 2] The project name is inconsistently rendered as 'Thousands Brains Project' in one place and 'Thousand Brains Project' elsewhere; the paper should use a single spelling throughout.
[Section 9.12] The text acknowledges that the feature-evidence mechanism is 'only halfway achieved right now' because relative evidence can implicitly subtract from hypotheses. This self-acknowledged limitation should be carried into the paper's high-level claims about robustness, since it undercuts the statement that features never reduce evidence.

Circularity Check

2 steps flagged · score 5.0 of 10

Universality claim is inherited from the authors' own Thousand Brains Theory and from a definition of 'object' broad enough to include abstract concepts; the implemented LM provides no independent derivation.

self citation load bearing [Section 1 (Introduction) and Section 2.2 (Core Principles)]
"Core to the design of thousand-brains systems are the principles laid out in the Thousand Brains Theory [Hawkins et al., 2019], a theory of intelligence derived from neuroscientific evidence of the anatomy and function of the neocortex. ... Most concepts presented here derive from the Thousand Brains Theory (TBT) [Hawkins et al., 2019] and experimental evidence about how the brain works."

The paper's central claim — that Monty 'should be able to solve any task the neocortex can solve' (§2.4) — rests on the premise that all intelligence is produced by a single sensorimotor reference-frame mechanism (§1). That premise is not demonstrated or tested here; it is the content of TBT, cited from the authors' own prior work. Monty is then presented as an instantiation of TBT, so the universality claim is inherited from the same source rather than independently established. Section 5, titled 'Experimental Evaluations', contains no experimental results that could break this self-citation chain. Mountcastle is cited for a common cortical algorithm, but the specific reference-frame and sensorimotor-universality claim comes from [Hawkins et al., 2019] and [Hawkins et al., 2017].
self definitional [Section 3.1 (Cortical Messaging Protocol) and Section 2.3 (Challenging Preconceptions)]
"An object is a discrete entity composed of a collection of one or more other objects, each with their own associated pose. ... equally, an object could be a scene, an abstract arrangement of concepts, or any other composition of sub-objects. ... These principles are used for modeling all discrete concepts in the world, from the simplest of physical objects to abstract concepts in society or mathematics."

The claim that reference frames can represent all abstract concepts is made true by definition: 'object' is stipulated to include any 'abstract arrangement of concepts', and pose is defined as location and orientation in 3D space. Yet the implemented evidence-based LM only defines morphological pose quantities — 3D locations, point normals, and principal curvature directions (§8, §9.12). No construction is given for what a pose of a family tree, logical relation, or language concept is. Section 3.3 says the exact structure of such spaces 'can potentially be learned' and §11.5.2 defers abstract spaces to future work, confirming that the universality claim is an assumed definition rather than a derived consequence of the mechanism.

full rationale

This is a position/white-paper rather than a formal derivation: it contains no equations, no fitted parameters, and no experimental results, so the ordinary 'fitted input called prediction' pattern is absent. The technical core — graph memory, evidence updates, voting, and CMP messaging — is described as a self-contained algorithm and is not circular in itself. However, the paper's most load-bearing claim, that the architecture will eventually implement any neocortical capability, is justified in two circular ways. First, it is imported from the authors' own Thousand Brains Theory via self-citations; the theory is the premise, not a result established in this paper. Second, 'object' is defined so broadly that abstract concepts are objects by stipulation, so the assertion that reference frames model all discrete concepts is true by definition rather than by demonstration. The paper itself acknowledges that abstract spaces are future work (§3.3, §11.5.2) and that Section 5 contains no evaluations, which prevents a higher score (no forced derivation or unique-theorem chain). The score of 5 reflects partial circularity: the central universality claim reduces to self-citation and definition, while the implemented low-level sensorimotor algorithm remains independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of domain assumptions inherited from the authors' prior theory and from neuroscience claims, plus several user-set hyperparameters. No new physical entities are introduced. The main free parameters are thresholds and sampling counts that would need tuning in any application; the axioms are the load-bearing assumptions that the sensorimotor, reference-frame algorithm is universal.

free parameters (5)

percent-threshold
User-set threshold that determines when a hypothesis is confident enough to end an episode (Section 10.1).
feature tolerance
Tolerance used to decide whether sensed and stored features match when adding evidence (Section 9.10).
sampled rotations N
When curvature direction is ambiguous, N possible rotations are sampled to initialize hypotheses (Section 9.10).
momentum parameter alpha
Controls how likely a random policy repeats the previous action, affecting exploration (Section 11.3.1).
minimum number of LMs for terminal state
Episode ends when a parameter-defined minimum number of learning modules reach a terminal state (Section 10.3).

assumptions (4)

domain assumption A common neural algorithm underlies all cortical functions and can be captured in a single learning module.
Invoked in Section 1: 'all aspects of intelligence are created by the same sensorimotor mechanism'.
ad hoc to paper Spatial reference frames are a sufficient representation for all objects and for abstract concepts.
Section 3.3 states models have a 3D inductive bias and that 'the exact structure of space can potentially be learned', which is required for abstract spaces.
domain assumption Sensorimotor interaction is the necessary foundation for all intelligence.
Section 1 and 2.3 argue learning from interaction with the world is foundational, unlike training on static datasets.
domain assumption The Cortical Messaging Protocol can convey all task-relevant information between modules without sharing internal models.
Section 3.1 defines CMP and claims it enables arbitrary combination of modules; this is a design assumption without proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence." pith.science (2026). https://pith.science/paper/RX2JYFMK

@misc{pith2026241218354,
  author       = {Pith},
  title        = {Pith review of: The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RX2JYFMK}},
  note         = {Machine review of arXiv:2412.18354}
}

read the original abstract

Artificial intelligence has advanced rapidly in the last decade, driven primarily by progress in the scale of deep-learning systems. Despite these advances, the creation of intelligent systems that can operate effectively in diverse, real-world environments remains a significant challenge. In this white paper, we outline the Thousand Brains Project, an ongoing research effort to develop an alternative, complementary form of AI, derived from the operating principles of the neocortex. We present an early version of a thousand-brains system, a sensorimotor agent that is uniquely suited to quickly learn a wide range of tasks and eventually implement any capabilities the human neocortex has. Core to its design is the use of a repeating computational unit, the learning module, modeled on the cortical columns found in mammalian brains. Each learning module operates as a semi-independent unit that can model entire objects, represents information through spatially structured reference frames, and both estimates and is able to effect movement in the world. Learning is a quick, associative process, similar to Hebbian learning in the brain, and leverages inductive biases around the spatial structure of the world to enable rapid and continual learning. Multiple learning modules can interact with one another both hierarchically and non-hierarchically via a "cortical messaging protocol" (CMP), creating more abstract representations and supporting multimodal integration. We outline the key principles motivating the design of thousand-brains systems and provide details about the implementation of Monty, our first instantiation of such a system. Code can be found at https://github.com/thousandbrainsproject/tbp.monty, along with more detailed documentation at https://thousandbrainsproject.readme.io/.

Figures

Figures reproduced from arXiv: 2412.18354 by the authors.

**Figure 2.** Learning modules learn structured models [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Conceptual sketch of how the learning module could be implementing possible mechanisms of cortical [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (22 more)

**Figure 4.** Figure 4: By using a common messaging protocol between sensor modules and learning modules, the system can easily [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: High-level overview of the architecture with all the main conceptual components mirroring Figure 4 applied [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Three ways time is discretized in our current experimental setup is into steps (one movement and one [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

**Figure 7.** Figure 7: Example of six sensors in Habitat. The view-finder is not connected to any sensor module or learning module [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

**Figure 8.** Figure 8: Difference between sensors and agents. Agents can move independently of each other while all sensors [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: The 77 objects of the YCB dataset at 0, 0, 0 rotation. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

**Figure 10.** Figure 10: Example Monty system with five sensor modules and learning modules. Each sensor patch perceives a small [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

**Figure 11.** Figure 11: Processing sensory inputs for the Cortical Messaging Protocol. The sensor patch captures a small area of the [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

**Figure 12.** Figure 12: First two episodes (separated by a vertical double line) during learning. After we recognize an object [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

**Figure 13.** Figure 13: A Graph of features (nodes), linked by displace [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

**Figure 14.** Figure 14: (top) Building a graph from a buffer of observations. First, similar observations (high spatial proximity [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

**Figure 15.** Figure 15: Sensing a Two-Toned Object A representation of a sensor (camera patch), moving over the surface of a cylinder that is colored red in some parts, and blue in others. We will show how hypotheses are initialized and how evidence is updated based on this example. In this …

**Figure 16.** Figure 16: Mechanism for initializing hypotheses from one observation. (Top row) Initializing possible poses. For this, [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]

**Figure 17.** Figure 17: Updating evidence using two steps. First, we move upwards and sense red and curved features (top row). [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]

**Figure 18.** Figure 18: Calculating the evidence update for one hypothesis. First, we calculate a search location given the hypothesis [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]

**Figure 19.** Figure 19: Information flow in a graph learning module. At each step the LM receives features and a pose as input. [PITH_FULL_IMAGE:figures/full_fig_p024_19.png]

**Figure 20.** Figure 20: Thresholding evidence values to obtain possible matches and poses. The highest pose evidence within an [PITH_FULL_IMAGE:figures/full_fig_p025_20.png]

**Figure 21.** Figure 21: Vote sent from LM 1 to LM 2 and used to update the evidence in LM 2. The two LMs receive input from [PITH_FULL_IMAGE:figures/full_fig_p026_21.png]

**Figure 22.** Figure 22: Using the same object and agent positions for all objects leads to objects covering different amounts of the [PITH_FULL_IMAGE:figures/full_fig_p028_22.png]

**Figure 23.** Figure 23: Comparison of different surface-agent policies. [PITH_FULL_IMAGE:figures/full_fig_p028_23.png]

**Figure 24.** Figure 24: Two samples of the curvature-informed surface policy being used during inference. Blue segments represent [PITH_FULL_IMAGE:figures/full_fig_p029_24.png]

**Figure 25.** Figure 25: The hypothesis-testing policy. The left and right plots show where the agent is in the external world at two [PITH_FULL_IMAGE:figures/full_fig_p030_25.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Finding Similar Objects and Active Inference for Surprise in Numenta Neocortex Model
q-bio.NC 2025-06 conditional novelty 5.0 of 10

The paper defines similarity-finding and surprise-response algorithms on Numenta's neocortex model, with a matching non-Bayesian probability update.
The Generalist Brain Module: Module Repetition in Neural Networks in Light of the Minicolumn Hypothesis
q-bio.NC 2025-07 conditional novelty 4.0 of 10

A review arguing that repeating a single generalist neural module, inspired by cortical minicolumns, yields robustness, scalability, and generalization benefits compared to monolithic networks.

Reference graph

Works this paper leans on

17 extracted references · 7 canonical work pages · cited by 2 Pith papers

[1]

A framework for intelligence and cortical function based on grid cells in the neocortex

Jeff Hawkins, Marcus Lewis, Mirko Klukas, Scott Purdy, and Subutai Ahmad. A framework for intelligence and cortical function based on grid cells in the neocortex . Frontiers in Neural Circuits, 2019. ISSN 16625110. doi:10.3389/fncir.2018.00121

arXiv 2019
[2]

Mountcastle

Vernon B. Mountcastle. The columnar organization of the neocortex . Brain, 120 0 (4), 1997. ISSN 00068950. doi:10.1093/brain/120.4.701

doi:10.1093/brain/120.4.701 1997
[3]

The mindful brain: Cortical organization and the group-selective theory of higher brain function

G Edelman and V Mountcastle. The mindful brain: Cortical organization and the group-selective theory of higher brain function. pp, 100, 1978. URL https://psycnet.apa.org/record/1979-25355-000

work page 1978
[4]

Danny Driess, Fei Xia, Mehdi S.M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied m...

work page 2023
[5]

GPT -4 technical report

OpenAI , Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berne...

arXiv 2023
[6]

pi0: A vision-language-action flow model for general robot control

Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky. pi0: A vis...

arXiv 2024
[7]

A generalist agent

Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent. Transactions on ...

arXiv 2022
[8]

SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy, Martin Engelcke, Ryan Faulkner, F...

arXiv 2024

Show all 17 references

[9]

Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex

Jeff Hawkins and Subutai Ahmad. Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex . Frontiers in Neural Circuits, 10, 2016. ISSN 16625110. doi:10.3389/fncir.2016.00023

2016
[10]

A Theory of How Columns in the Neocortex Enable Learning the Structure of the World

Jeff Hawkins, Subutai Ahmad, and Yuwei Cui. A Theory of How Columns in the Neocortex Enable Learning the Structure of the World . Frontiers in Neural Circuits, 11 0 (October): 0 1--18, 2017. ISSN 1662-5110. doi:10.3389/fncir.2017.00081. URL http://journal.frontiersin.org/artic...

2017
[11]

How Can We Be So Dense? The Robustness of Highly Sparse Representations

Subutai Ahmad and Luiz Scheinkman. How Can We Be So Dense? The Robustness of Highly Sparse Representations . ICML 2019 Workshop on Uncertainty and Robustness in Deep Learning, 2019

2019
[12]

Locations in the neocortex: A theory of sensorimotor object recognition using cortical grid cells

Marcus Lewis, Scott Purdy, Subutai Ahmad, and Jeff Hawkins. Locations in the neocortex: A theory of sensorimotor object recognition using cortical grid cells . Frontiers in Neural Circuits, 2019. ISSN 16625110. doi:10.3389/fncir.2019.00022

2019
[13]

Prasad, Briana J

Judy A. Prasad, Briana J. Carroll, and S. Murray Sherman. Layer 5 Corticofugal Projections from diverse cortical areas: Variations on a pattern of thalamic and extrathalamic targets . Journal of Neuroscience, 40 0 (30), 2020. ISSN 15292401. doi:10.1523/JNEUROSCI.0529-20.2020

2020 doi
[14]

Habitat: A P latform for E mbodied AI R esearch

Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A P latform for E mbodied AI R esearch. In Proceedings of the IEEE/CVF International Confe...

2019
[15]

Habitat 2.0: Training home assistants to rearrange their habitat

Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Ji...

2021
[16]

Habitat 3.0: A co-habitat for humans, avatars and robots, 2023

Xavi Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Ruslan Partsey, Jimmy Yang, Ruta Desai, Alexander William Clegg, Michal Hlavac, Tiffany Min, Theo Gervet, Vladimir Vondrus, Vincent-Pierre Berges, John Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan,...

2023
[17]

Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M. Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 International Conference on Advanced Robotics (ICAR), pages 510--517, 2015. doi:10.1109/...

2015

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

A framework for intelligence and cortical function based on grid cells in the neocortex

Jeff Hawkins, Marcus Lewis, Mirko Klukas, Scott Purdy, and Subutai Ahmad. A framework for intelligence and cortical function based on grid cells in the neocortex . Frontiers in Neural Circuits, 2019. ISSN 16625110. doi:10.3389/fncir.2018.00121

arXiv 2019

[2] [2]

Mountcastle

Vernon B. Mountcastle. The columnar organization of the neocortex . Brain, 120 0 (4), 1997. ISSN 00068950. doi:10.1093/brain/120.4.701

doi:10.1093/brain/120.4.701 1997

[3] [3]

The mindful brain: Cortical organization and the group-selective theory of higher brain function

G Edelman and V Mountcastle. The mindful brain: Cortical organization and the group-selective theory of higher brain function. pp, 100, 1978. URL https://psycnet.apa.org/record/1979-25355-000

work page 1978

[4] [4]

Danny Driess, Fei Xia, Mehdi S.M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied m...

work page 2023

[5] [5]

GPT -4 technical report

OpenAI , Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berne...

arXiv 2023

[6] [6]

pi0: A vision-language-action flow model for general robot control

Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky. pi0: A vis...

arXiv 2024

[7] [7]

A generalist agent

Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent. Transactions on ...

arXiv 2022

[8] [8]

SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy, Martin Engelcke, Ryan Faulkner, F...

arXiv 2024