Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Memory makes computation universal, remember?

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Memory plus reliable history access is enough to make any state-keeping system a universal computer, the paper argues.

desk verdict The formal claim collapses on the space accounting — the history log grows linearly with runtime — so the paper is a clearly written essay with a broken theorem. read the letter →

arxiv 2412.17794 v1 pith:UIOXTIYH submitted 2024-12-23 cs.LG

Erik Garrison This is my paper · ORCID
classification cs.LG
keywords memoryuniversalcomputationrecursivestatemaintenancereliablehistoryaccessTuringcompletenessneuralnetworkschain-of-thoughtreasoning
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

The paper tries to establish that universal computation comes from memory, not from the power of processing units. Its formal claim is that any system able to (1) recursively update and retain its own state and (2) retrieve any past state reliably, in correct temporal order, can simulate a universal Turing machine with at most logarithmic overhead in time and space; the paper asserts these two abilities are necessary and sufficient. A sympathetic reader should care because the claim would unify how we explain progress in neural networks, biological cells, and language models: capability grows from better state maintenance, not from smarter basic operations. The paper applies this lens to chain-of-thought reasoning, recent benchmark gains on abstract reasoning tasks, and cellular memory mechanisms.

What carries the argument

The load-bearing construction is a time-stamped history ledger. Each tape write is recorded as a tuple $h_p=(p,a_p,t_p)$, and the system answers "what is on the tape at position $p$ now?" by finding the recorded entry with maximum $t_p$ not exceeding the current time $t$. This turns the tape into an addressable log, which is exactly what the "reliable history access" requirement supplies. The other half is the recursive update loop $s \mapsto f(s)$, where $s=(q,p,a,t)$; the step counter makes each configuration uniquely addressable in time. The author argues that this combination, not the complexity of $f$, is what delivers universal computation.

What would settle it

Take a recurrent network with finite precision and a memory window that grows only logarithmically with time; if it cannot answer "what was the last symbol written at position p?" without error and within logarithmic time, then the condition of reliable history access is not met, and the theorem does not apply to it. A concrete check: measure retrieval error and retrieval time as the simulated history length grows; linear growth of time or non-vanishing error would falsify the applicability of the theorem to that system.

Watch

Extended reading notes

Core claim

The paper's central discovery is Theorem 1: a system with recursive state maintenance and reliable history access can simulate a universal Turing machine with logarithmic overhead. The simulation encodes each machine configuration as a tuple $(q, p, a, t)$ holding the current state, tape-head position, symbol under the head, and step counter, and stores every tape write as a history entry $h_p=(p,a_p,t_p)$; reading a cell means retrieving the entry with the largest $t_p$ not exceeding the current time. Two inductive invariants, state coherence and history consistency, show that the simulated configuration always matches the machine being simulated. The paper presents this as evidence that simple or parallel systems, including threshold units and neural networks, become universal once they can keep and retrieve state across iterations.

Load-bearing premise

The proof assumes a system can implement unbounded, error-free, addressable history access at logarithmic cost per retrieval; no physical mechanism shows that real neural networks, cells, or language models actually provide this.

Editorial extensions

If this is right

  • Neural networks with simple threshold units become universal if they maintain reliable state across iterations; adding memory is more decisive than adding unit complexity.
  • Chain-of-thought prompting works by providing scaffolding for state maintenance, so models can execute sequential computations that single-pass pattern matching cannot.
  • Scaling model size alone will not remove sequential reasoning limits; architecture should be judged by how robustly it maintains and accesses history.
  • Systems that score highly on abstract visual-reasoning benchmarks likely do so by storing and retrieving partial results over a search, not by having better base processors.

Reading between the lines

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

  • If the thesis holds, an inexpensive testable extension is to add an explicit external history index to a transformer and measure whether the gain per parameter exceeds gains from added depth; a positive result would support the memory-over-architecture claim.
  • The proof's logarithmic overhead depends on an addressable-history primitive; until a neural or molecular implementation shows $O(\log t)$ retrieval cost, the strongest reading of the theorem applies to abstract systems rather than current hardware.
  • The framework suggests a benchmark-based test: sort reasoning tasks by how much explicit state they require; performance with and without chain-of-thought should separate along that axis.
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.

Referee Report

4 major / 4 minor

Summary. The paper argues that universal computation is enabled by two capabilities—recursive state maintenance and reliable history access—and claims to prove that these are both necessary and sufficient for simulating a Universal Turing Machine with at most logarithmic overhead in space and time. The proof constructs a simulation in which the machine state, tape contents, and history are encoded, and then discusses implications for neural networks, biological cells, and large language models, including chain-of-thought reasoning and OpenAI's o3 system. The paper contains a formal-looking theorem (Theorem 1) and corollary (Corollary 4), but the proof of the theorem is only a sketch, the complexity analysis omits the storage cost of the history log, the claimed necessity direction is never proved, and the cost model for 'reliable history access' is left unspecified. The latter half of the paper is largely speculative and connects the framework to empirical AI results without a rigorous bridge.

Significance. If the central claim were correct, it would provide a unifying principle for computation across biological and artificial systems, potentially reframing how the field thinks about the role of memory in AI. The paper also gestures at an important recent literature on memory checking and lower bounds (e.g., Boyle et al.). However, as written, the formal result is not established: the space accounting in the proof of Theorem 1 is internally inconsistent, the necessity direction is asserted rather than proved, and the logarithmic overhead depends on an unstated cost model for history access. The paper's value is therefore more as a position essay than as a rigorous contribution to computational theory. The constructive idea—that reliable history access is sufficient for universality—is standard and plausible, but it is not new and is not developed enough to support the paper's strong conclusions.

major comments (4)
  1. [§2, Theorem 1 proof, 'Tape Simulation' and 'complexity bounds'] The proof's space accounting omits the stored history log. The construction maintains an entry h_p = (p, a_p, t_p) for every write operation; the number of entries grows with the number of write events, which can be as large as the runtime T. For example, a UTM that alternates writing two symbols on a single tape cell forever uses a constant-size tape but generates one new history entry per step, consuming Ω(T log T) bits. The proof counts only O(log t) bits for the step counter and O(log n) bits for the head position, so the claimed O(log T) space overhead is incorrect. This is a load-bearing flaw: the theorem's main quantitative claim fails for the construction as described.
  2. [§2, definition of 'reliable history access' and Theorem 1 proof] The time complexity bound assumes that 'reliable history access' supports the positional, time-thresholded query 'retrieve the entry with maximum t_p ≤ t for position p' in O(log t) time. However, the definition in Section 2 only requires that previous states can be referenced reliably and in order; it does not specify any data structure, cost model, or mechanism for this query. Without such a cost model, the statement 'O(log t) for history access operations' is an unproven assumption rather than a derived bound. This affects both the time and space claims, since a naive implementation of the query could require scanning the entire history.
  3. [Abstract and §2, Corollary 4] The abstract states that the two capabilities are 'necessary and sufficient' for universal computation, and the discussion around Corollary 4 implies the same. However, Theorem 1 and its proof establish only sufficiency: they show that a system with the two capabilities can simulate a UTM. No argument is given that every universal system must possess recursive state maintenance and reliable history access. The necessity claim is not proved and is not obviously true under any standard definition of universality; it would require a separate theorem with a converse construction. This overgeneralization is central to the paper's framing.
  4. [§3, 'Neural architectures face a fundamental computational barrier'] The paper states that 'neural architectures are restricted to TC0 complexity' because of parallel training, citing refs. [3,4]. Those references establish specific results about log-precision transformers under certain assumptions, not a general impossibility theorem for all neural architectures. The paper's own axiom—that parallel training at scale forces TC0—is presented without proof and is not derived from the cited works. Since the later practical discussion depends on this claimed barrier, it should be either proved or substantially qualified; as written, it is an unsubstantiated premise that supports the paper's biological and AI conclusions.
minor comments (4)
  1. [Throughout] The text contains numerous typographical artifacts from the PDF extraction, such as 'r principle', 'enab les', 'bo th', 'sufficient', and 'straightforward' split across lines. These should be cleaned before any resubmission.
  2. [References] Several references are incomplete or malformed, e.g., [16] gives 'arXiv:2024.268' without a title or author, and [47] contains an embedded line break in the URL. The paper also cites [15,16,9] as proving tightness of the logarithmic bounds, but those works do not analyze the particular construction in this paper; the tightness claim is not supported.
  3. [§2, Corollary 4 discussion] The example of a threshold unit that 'becomes capable of counting' with memory is illustrative but not connected rigorously to the theorem: no proof is given that the described counter can be composed into a full UTM simulation, and the phrase 'this counter then serves as a fundamental building block' is asserted rather than demonstrated.
  4. [§4, o3 discussion] The claims about o3's performance and cost are drawn from a non-archival blog post and a technical report, and the interpretation of o3's chain-of-thought as 'reliable history access' is speculative. The paper should clearly distinguish empirical observations from its own framework's predictions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 1 is a constructive simulation from stated memory capabilities, not a restatement of its conclusion; the main flaws are correctness gaps, not circular dependencies.

full rationale

The paper's central derivation is Theorem 1, which claims that recursive state maintenance plus reliable history access suffice to simulate a Universal Turing Machine with logarithmic overhead. The proof gives an explicit encoding of the UTM configuration s=(q,p,a,t) and a construction of tape contents from history entries h_p=(p,a_p,t_p), with induction lemmas for state coherence and history consistency. This is a genuine constructive reduction: the assumptions do not mention UTM simulation, and the simulation is built out of them rather than being inserted into them. The assumptions are admittedly strong—reliable history access already grants unbounded, error-free, temporally ordered storage—but that makes the theorem close to a standard random-access-memory simulation result, not circular. The abstract's 'necessary and sufficient' claim is not supported by a necessity proof, and the proof of the logarithmic space bound omits the stored history entries from the accounting; however, these are correctness and rigor problems, not cases where a prediction reduces by construction to a fitted input or to a self-citation chain. The paper explicitly credits classical universality results (Rule 110, mov, etc.) and uses external empirical work as validation, and there are no load-bearing self-citations. No circular step meeting the required evidentiary standard can be identified.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No numeric free parameters are fitted. The load-bearing axioms are the standard UTM framework, the assumed implementability of unbounded cheap addressable history, and the overgeneralized TC0 claim about neural networks. No new physical or mathematical entities are introduced.

assumptions (3)
  • standard math Church-Turing thesis and standard UTM definition
    Section 2 defines universality via a Universal Turing Machine with state set Q, tape alphabet Gamma, and transition delta; this is taken as the standard model without proof.
  • domain assumption S can implement unbounded, error-free, addressable history
    Theorem 1's construction requires storing history entries h_p=(p,a_p,t_p) and retrieving the most recent write to any position in O(log t) time; no implementation is provided and real biological and neural systems are finite.
  • ad hoc to paper Neural architectures are restricted to TC0 because of parallel training
    Section 3 asserts this as proven and uses it to argue memory mechanisms determine computational power, but cited results are about log-precision transformers, not all neural networks; the link to parallel training is asserted without derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Memory makes computation universal, remember?." pith.science (2026). https://pith.science/paper/UIOXTIYH

@misc{pith2026241217794,
  author       = {Pith},
  title        = {Pith review of: Memory makes computation universal, remember?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UIOXTIYH}},
  note         = {Machine review of arXiv:2412.17794}
}
read the original abstract

Recent breakthroughs in AI capability have been attributed to increasingly sophisticated architectures and alignment techniques, but a simpler principle may explain these advances: memory makes computation universal. Memory enables universal computation through two fundamental capabilities: recursive state maintenance and reliable history access. We formally prove these requirements are both necessary and sufficient for universal computation. This principle manifests across scales, from cellular computation to neural networks to language models. Complex behavior emerges not from sophisticated processing units but from maintaining and accessing state across time. We demonstrate how parallel systems like neural networks achieve universal computation despite limitations in their basic units by maintaining state across iterations. This theoretical framework reveals a universal pattern: computational advances consistently emerge from enhanced abilities to maintain and access state rather than from more complex basic operations. Our analysis unifies understanding of computation across biological systems, artificial intelligence, and human cognition, reminding us that humanity's own computational capabilities have evolved in step with our technical ability to remember through oral traditions, writing, and now computing.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Hierarchical Domain Generalization

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Over infinite domains, hierarchy-uniform domain generalization is impossible for every nontrivial hypothesis class; a length-generalization bound is a property of the length hierarchy, not a hierarchy-free guarantee.

Reference graph

Works this paper leans on

56 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    Molecular Biology of the Cell: Seventh International Stude nt Edition with Registration Card

    Bruce Alberts, Rebecca Heald, Alexander Johnson, David Morgan, Martin Raff, Keith Roberts, and Peter Walter. Molecular Biology of the Cell: Seventh International Stude nt Edition with Registration Card . WW Norton & Company , 2022

    work page 2022

  2. [2]

    On computable numbers, with an application to t he entscheidungs problem

    A T uring. On computable numbers, with an application to t he entscheidungs problem. Proceedings of the London Mathematical Society Series/2 (42) , pages 230–42, 1936

    work page 1936

  3. [3]

    The parallelism t radeoff: Limitations of log-precision transformers

    William Merrill and Ashish Sabharwal. The parallelism t radeoff: Limitations of log-precision transformers. T ransactions of the Association for Computational Linguis tics, 11:531–545, 2023. doi:10.1162/tacl_a_00562

    doi:10.1162/tacl_a_00562 2023

  4. [4]

    On limitations of the transformer architec- ture

    Binghui Peng, Srini Narayanan, and Christos Papadimitr iou. On limitations of the transformer architec- ture. COLM, 2024

    work page 2024

  5. [5]

    Universality in elementary cellular auto mata

    Matthew Cook. Universality in elementary cellular auto mata. Complex Systems, 15(1):1–40, 2004

    work page 2004

  6. [6]

    mov is turing-complete

    Stephen Dolan. mov is turing-complete. Cambridge Unive rsity Computer Laboratory , July 2013. URL: https://www.cl.cam.ac.uk/~sd601/papers/mov.pdf

    work page 2013

  7. [7]

    Parallel molecular computation on digital data stor ed in dna

    Boya Wang, Siyuan Stella Wang, Cameron Chalk, Andrew D El lington, and David Solove- ichik. Parallel molecular computation on digital data stor ed in dna. PNAS, 120(37), 2023. doi:10.1073/pnas.2217330120

    work page doi:10.1073/pnas.2217330120 2023

  8. [8]

    Space-time tradeoffs f or linear recursion

    Sowmitri Swamy and John E Savage. Space-time tradeoffs f or linear recursion. Mathematical Systems Theory, 16:9–27, 1983. doi:10.1007/BF01744566

    work page doi:10.1007/bf01744566 1983

Show all 56 references
  1. [9]

    Memory checking requires logarithmic overhead

    Elette Boyle, Ilan Komargodski, and Neekon Vafa. Memory checking requires logarithmic overhead. In Proceedings of the 56th Annual ACM Symposium on Theory of Com puting (STOC ’24) , pages 1712– 1723, New Y ork, NY , USA, June 2024. ACM. doi:10.1145/3618260.3649686

    2024

  2. [10]

    Introduction to the theory of computat ion

    Michael Sipser. Introduction to the theory of computat ion. ACM Sigact News, 27(1):27–29, 1996

    1996

  3. [11]

    Space-time tradeoffs

    John E Savage. Space-time tradeoffs. Models of Computation , 1994

    1994

  4. [12]

    Time/space trade-offs for reversib le computation

    Charles H Bennett. Time/space trade-offs for reversib le computation. SIAM Journal on Computing , 18(4):766–776, 1989

    1989

  5. [13]

    Unive rsality in quantum computation

    David Deutsch, Adriano Barenco, and Artur Ekert. Unive rsality in quantum computation. arXiv:quant- ph/9505018, 1995

    1995

  6. [14]

    On the complexity of space bounded computations

    Maciej Liskiewicz and R ¨ udiger Reischuk. On the complexity of space bounded computations. Theoretical Computer Science, 147:1–45, 1994

    1994

  7. [15]

    Memory lower boun ds and impossibility results for anonymous dynamic broadcast

    Garrett Parzych and Joshua J Daymude. Memory lower boun ds and impossibility results for anonymous dynamic broadcast. arXiv:2407.09714, 2024

    2024 arXiv

  8. [16]

    A new approach to generic lower bounds: Clas sical/quantum mdl, quantum factoring, and more

    Minki Hhan. A new approach to generic lower bounds: Clas sical/quantum mdl, quantum factoring, and more. arXiv:2024.268, 2024

    2024

  9. [17]

    Molecular complexi ty calculated by fractal dimension

    Modest von Korff and Thomas Sander. Molecular complexi ty calculated by fractal dimension. Scientific Reports, 9:1746, 2019. doi:10.1038/s41598-018-37253-8 . 9

    2019 doi

  10. [18]

    Information maintenance in working memory: An integrated presentation of cognitive an d neural concepts

    Markus Martini, Marco R Furtner, Thomas Maran, and Pier re Sachse. Information maintenance in working memory: An integrated presentation of cognitive an d neural concepts. Frontiers in Systems Neuroscience, 9:104, 2015. doi:10.3389/fnsys.2015.00104

    2015

  11. [19]

    Differential brain mechanisms of selection and maintenance of information during working memory

    Romain Quentin et al. Differential brain mechanisms of selection and maintenance of information during working memory . Journal of Neuroscience , 39(19):3728–3740, 2019. doi:10.1523/JNEUROSCI.2764-18.2019

    2019 doi

  12. [20]

    Epigenetic cell memory: The gene’s inner chromatin modification circuit

    Simone Bruno, Ruth J Williams, and Domitilla Del Vecchi o. Epigenetic cell memory: The gene’s inner chromatin modification circuit. PLoS Computational Biology , 18(4), 2022. doi:10.1371/journal.pcbi.1009961

    2022 doi

  13. [21]

    A. M. T uring. The chemical basis of morphogenesis. Philosophical T ransactions of the Royal Society of London. Series B, Biological Sciences , 237(641):37–72, 1952

    1952

  14. [22]

    Analyz ing biological and artificial neural net- works: challenges with opportunities for synergy? Current Opinion in Neurobiology , 55:55–64, 2019

    David GT Barrett, Ari S Morcos, and Jakob H Macke. Analyz ing biological and artificial neural net- works: challenges with opportunities for synergy? Current Opinion in Neurobiology , 55:55–64, 2019. doi:10.1016/j.conb.2019.01.007

    2019 doi

  15. [23]

    Deep learning , volume 1

    Y oshua Bengio, Ian Goodfellow, and Aaron Courville. Deep learning , volume 1. MIT press Cambridge, MA, USA, 2017

    2017

  16. [24]

    Efficiently modeling long sequences with structured stat e spaces

    Albert Gu, Karan Goel, and Christopher R ´e. Efficiently modeling long sequences with structured stat e spaces. arXiv preprint arXiv:2111.00396 , 2021

    2021 arXiv

  17. [25]

    Mamba: Linear-time sequence mode ling with selective state spaces

    Albert Gu and T ri Dao. Mamba: Linear-time sequence mode ling with selective state spaces. arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  18. [26]

    T ransformers are ssms: Generaliz ed models and efficient algorithms through structured state space duality

    T ri Dao and Albert Gu. T ransformers are ssms: Generaliz ed models and efficient algorithms through structured state space duality . arXiv preprint arXiv:2405.21060 , 2024

    2024 arXiv

  19. [27]

    Hyenadna: Long-range genomic se- quence modeling at single nucleotide resolution

    Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Y oshua Bengio, et al. Hyenadna: Long-range genomic se- quence modeling at single nucleotide resolution. Advances in neural information p...

    2024

  20. [28]

    Were rnns all we needed? arXiv preprint arXiv:2410.01201 , 2024

    Leo Feng, Frederick T ung, Mohamed Osama Ahmed, Y oshua B engio, and Hossein Hajimirsadeghi. Were rnns all we needed? arXiv preprint arXiv:2410.01201 , 2024

    2024 arXiv

  21. [29]

    Efficient computation by molecular competition networks

    Haoxiao Cai, Xiaoran Zhang, Rong Qiao, Xiaowo Wang, and Lei Wei. Efficient computation by molecular competition networks. Physical Review Research, 6:033208, 2024

    2024

  22. [30]

    scgrn: a comprehens ive single-cell gene regulatory network platform

    Shuhua Fu, Anqi Wang, and Kin Fai Au. scgrn: a comprehens ive single-cell gene regulatory network platform. Nucleic Acids Research , 52(D1):D293–D303, 2023. doi:10.1093/nar/gkad885

    2023 doi

  23. [31]

    What is the total number of protein molecules p er cell volume? a call to rethink some published values

    Ron Milo. What is the total number of protein molecules p er cell volume? a call to rethink some published values. BioEssays, 35(12):1050–1055, 2013. doi:10.1002/bies.201300066

    2013 doi

  24. [32]

    Molecular tu mbling rate effects on t1 and t2

    Roland Zhang, Zach Drew, and Jeremy Jones. Molecular tu mbling rate effects on t1 and t2. Radiopae- dia.org, 2023. URL: https://radiopaedia.org/articles/60742, doi:10.53347/rID-60742

    2023 doi

  25. [33]

    The expressive p ower of transformers with chain of thought

    William Merrill and Ashish Sabharwal. The expressive p ower of transformers with chain of thought. ICLR, 2024

    2024

  26. [34]

    Crispr dna elements controlling site-specific spacer integration and proper repeat length by a type ii crispr–cas system

    Jenny G Kim, Sandra Garrett, Y unzhou Wei, Brenton R Grav eley , and Michael P T erns. Crispr dna elements controlling site-specific spacer integration and proper repeat length by a type ii crispr–cas system. Nucleic Acids Research , 47(16):8632–8648, 2019

    2019

  27. [35]

    Digital data storage on dna tape u sing crispr base editors

    Afsaneh Sadremomtaz, Robert F Glass, Jorge Eduardo Gue rrero, Dennis R LaJeunesse, Eric A Josephs, and Reza Zadegan. Digital data storage on dna tape u sing crispr base editors. Nature Com- munications, 14(1):6497, 2023. doi:10.1038/s41467-023-42223-4 . 10

    2023 doi

  28. [36]

    How computation is helping unravel the dynamics of morphogenesis

    David Pastor-Escuredo and Juan C del ´Alamo. How computation is helping unravel the dynamics of morphogenesis. Frontiers in Physics, 8:31, 2020. doi:10.3389/fphy.2020.00031

    2020

  29. [37]

    T oward s a bioinformatics of patterning: a com- putational approach to understanding regulative morphoge nesis

    Daniel Lobo, T aylor J Malone, and Michael Levin. T oward s a bioinformatics of patterning: a com- putational approach to understanding regulative morphoge nesis. Biology Open , 2(2):156–169, 2012. doi:10.1242/bio.20123400

    2012 doi

  30. [38]

    M ultiplex recording of cellular events over time on crispr biological tape

    Ravi U Sheth, Sung Sun Yim, Felix L Wu, and Harris H Wang. M ultiplex recording of cellular events over time on crispr biological tape. Science, 358(6369):1457–1461, 2017. doi:10.1126/science.aao0958

    2017 doi

  31. [39]

    How far can you trust chain-of-thought pro mpting? BD T ech T alks, 2024

    Ben Dickson. How far can you trust chain-of-thought pro mpting? BD T ech T alks, 2024

    2024

  32. [40]

    Recursive chain-of-feed back prevents performance degradation from redundant prompting

    Jinwoo Ahn and Kyuseung Shin. Recursive chain-of-feed back prevents performance degradation from redundant prompting. arXiv:2402.02648, 2024

    2024 arXiv

  33. [41]

    Openai o1 system card

    OpenAI. Openai o1 system card. T echnical report, OpenA I, December 2024

    2024

  34. [42]

    Mind your step (by step): Chain-of-thought can reduce performan ce on tasks where thinking makes humans worse

    Ryan Liu, Jiayi Geng, Addison J Wu, Ilia Sucholutsky , T a nia Lombrozo, and Thomas L Griffiths. Mind your step (by step): Chain-of-thought can reduce performan ce on tasks where thinking makes humans worse. arXiv preprint arXiv:2410.21333 , 2024

    arXiv

  35. [43]

    Mea- suring the effects of data parallelism in neural network tra ining

    Chris Shallue, Jaehoon Lee, Joe Antognini, Jascha Sohl -Dickstein, Roy Frostig, and George Dahl. Mea- suring the effects of data parallelism in neural network tra ining. Google Research Blog , 2019

    2019

  36. [44]

    Arc prize 2024: T echnical report

    Franc ¸ois Chollet, Mike Knoop, Gregory Kamradt, and Br yan Landers. Arc prize 2024: T echnical report. T echnical report, ARC Prize, December 2024

    2024

  37. [45]

    Openai o3 breakthrough high score o n arc-agi-pub, 12 2024

    Franc ¸ois Chollet. Openai o3 breakthrough high score o n arc-agi-pub, 12 2024. URL: https://arcprize.org/blog/2024/o3-breakthrough

    2024

  38. [46]

    Deliberative alignment: reasoning enables sa fer language models, 12 2024

    OpenAI. Deliberative alignment: reasoning enables sa fer language models, 12 2024. URL: https://openai.com/index/deliberative-alignment/

    2024

  39. [47]

    Delibera- tive alignment: Reasoning enables safer language models

    Melody Y Guan, Manas Joglekar, Eric Wallace, Saachi Jai n, Boaz Barak, et al. Delibera- tive alignment: Reasoning enables safer language models. T echnical report, 12 2024. URL: https://assets.ctfassets.net/kftzwdyauwt9/4pNYAZteAQXWtloDdANQ7L/0aedc43a8f2d1e5c71c5e114d287593

    2024

  40. [48]

    Learning to reason with LLMs, 9 2024

    OpenAI. Learning to reason with LLMs, 9 2024. URL: https://openai.com/index/learning-to-reason-with-ll ms/

    2024

  41. [49]

    Openai o3 and the future of language mod els, 12 2024

    Simon Willison. Openai o3 and the future of language mod els, 12 2024. URL: https://simonwillison.net/2024/Dec/21/o3

    2024

  42. [50]

    New parallel computing algorithm of molecular dynamics for extremely huge scale biological sys tems

    Jaewoon Jung, Chigusa Kobayashi, Kento Kasahara, et al . New parallel computing algorithm of molecular dynamics for extremely huge scale biological sys tems. Journal of Computational Chemistry , 42(4):231–241, 2020. doi:10.1002/jcc.26450

    2020 doi

  43. [51]

    Overcoming memory constraints in quan tum circuit simulation with a high-fidelity compression framework

    Ruomin Zhu et al. Overcoming memory constraints in quan tum circuit simulation with a high-fidelity compression framework. arXiv:2410.14088, 2024

    2024 arXiv

  44. [52]

    The molecular ba- sis of cell memory in mammals: The epigenetic cycle

    Menc´ ıa Espinosa-Mart´ ınez, Mar´ ıa Alc ´azar-Fabra, and David Landeira. The molecular ba- sis of cell memory in mammals: The epigenetic cycle. Science Advances , 10(8), 2024. doi:10.1126/sciadv.adl3188

    2024 doi

  45. [53]

    Chain-of-thought prompting elicits reason ing in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma , Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reason ing in large language models. In NeurIPS, 2022

    2022

  46. [54]

    Aut oregressive large language models are com- putationally universal

    Dale Schuurmans, Hanjun Dai, and Francesco Zanini. Aut oregressive large language models are com- putationally universal. arXiv preprint arXiv:2410.03170 , 2024

    2024 arXiv

  47. [55]

    Language models are few-shot learners

    T om B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah , Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 , 2020

    2005 arXiv

  48. [56]

    Emergent abilities of large language models

    Jason Wei, Yi T ay , Rishi Bommasani, Colin Raffel, Barret Zoph, et al. Emergent abilities of large language models. T ransactions on Machine Learning Research, 2022. 11

    2022

Pith tools

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