Pith. sign in

REVIEW 2 major objections 4 minor 53 references

Distributed and Decentralised Training: Technical Governance Challenges in a Shifting AI Landscape

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Progress in low-communication training is moving AI model building beyond the single data centre, eroding assumptions that underpin compute governance.

desk verdict Useful, honest taxonomy for distributed vs. decentralised training, but the no-off-switch scenario rests on a P2P architecture that has not been demonstrated. read the letter →

arxiv 2507.07765 v1 pith:RQ4MAGUM submitted 2025-07-10 cs.CY cs.LG

classification cs.CYcs.LG
keywords distributedtrainingdecentralisedlow-communicationcomputegovernanceshutdownabilitycapabilityproliferationstructuringdataparallelism
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

Recent advances in low-communication training algorithms are making it possible to train large AI models over slow, long-distance connections, and the paper argues this enables two distinct regimes that policy discussions tend to conflate: distributed training, where one central entity coordinates multiple data centres, and decentralised training, where community-contributed compute trains models with no central coordinator. The authors claim that if these trends continue, key assumptions of compute governance weaken: decentralised training could erode detectability and shutdownability, lower barriers to capability proliferation, and make compute structuring—splitting a training run to stay under regulatory size thresholds—easier. They also argue that compute still matters, that export controls remain relevant, and that decentralised training could bring benefits such as privacy-preserving access to private data and reduced concentration of power. The paper's aim is to give policymakers a more precise vocabulary and analysis for compute, capability proliferation, and decentralised AI development.

What carries the argument

The carrying technical mechanism is low-communication data-parallel training, exemplified by DiLoCo and DeMo: workers each hold a full model replica, train on a data shard, and synchronise parameters only every few hundred steps rather than every step, cutting communication needs by a factor of up to 500 and making training over ordinary Internet connections feasible. The paper adds a two-axis taxonomy—number of contributing parties against number of compute locations—that defines the distributed and decentralised regimes, and identifies the peer-to-peer network as the design in which nodes join and leave trustlessly without a master node, producing the 'no off-switch' governance problem.

What would settle it

A concrete check: run a peer-to-peer pre-training protocol at a scale of hundreds of billions of parameters over ordinary Internet connections, with nodes freely joining and leaving and no master node, and compare convergence and downstream benchmarks against a centralised run of the same total compute; if the peer-to-peer run cannot match convergence or reliability, the strongest governance scenario the paper describes does not materialise at scale.

Watch

Extended reading notes

Core claim

The paper's central claim is that the algorithmic shift toward low-communication training changes where and by whom frontier AI models can be built. It proposes a taxonomy that separates distributed training (one coordinating entity, many physical sites, driven by energy and siting constraints) from decentralised training (many contributors, many sites, no central coordinator), and argues that the latter, in a fully peer-to-peer form, could run with no obvious point of accountability or intervention, undermining the shutdownability that compute governance relies on. The paper does not claim this fully decentralised outcome has arrived: current demonstrations are at the ten-billion-parameter scale, the competitiveness gap relative to centralised training is uncertain, and no training protocol has yet achieved full decentralisation. What it argues is that the trajectory is real and fast, that the two regimes have distinct governance profiles, and that policy should be designed with both in view while continuing to treat compute quantity and export controls as relevant levers.

Load-bearing premise

The load-bearing premise is that fully decentralised, peer-to-peer training with no central coordinator will scale beyond today's ten-billion-parameter demonstrations to frontier-relevant sizes while remaining convergent, reliable, and competitive; if that scaling fails, the no-off-switch and capability-proliferation risks the paper warns about lose most of their force.

Editorial extensions

If this is right

  • Frontier labs can continue scaling by spreading training across multiple data centres, reducing the force of energy and siting constraints as bottlenecks to growth.
  • Decentralised training lowers barriers to capability proliferation, especially for post-training reasoning models, where forward-to-backward pass ratios can reach 1000:1 and consumer-grade hardware can contribute.
  • A fully peer-to-peer training network would have no clear point of accountability or intervention, making off-switch style policy interventions inapplicable to the training run itself.
  • Low-communication methods make compute structuring—dividing one large run into smaller parallel workloads to stay under regulatory FLOP thresholds—more feasible, and detection will require information sharing between providers.
  • Compute quantity, quality, and export controls remain relevant: low-communication training does not reduce the total compute needed, and hyperscalers retain advantages from GPU memory and scale.

Reading between the lines

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

  • If low-communication methods continue to improve with model scale, the strongest technical governance lever may shift from monitoring training clusters to monitoring synchronisation events: the frequency, destinations, and volume of parameter-exchange traffic.
  • A decisive test the paper leaves open is a head-to-head benchmark of fully peer-to-peer pre-training against centralised training at equal total compute and data; a persistent gap would keep the no-off-switch scenario hypothetical.
  • The paper's privacy-preserving-data benefit implies that heterogeneous private data from many owners must train models that generalise as well as models trained on pooled central data; that comparison is not yet demonstrated.
  • The same low-communication methods that enable community training could also let a single well-resourced actor hide a large training run across many small clusters, which suggests detection problems apply to distributed as well as decentralised regimes.
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

2 major / 4 minor

Summary. This paper argues that recent low-communication training algorithms (DiLoCo, DeMo, and successors) are enabling a shift from centrally located, single-cluster training toward two distinct regimes: distributed training across multiple data centres under a single coordinating entity, and decentralised training using community-provided compute without a central coordinator. It proposes a clear terminological distinction between these two regimes, reviews the algorithmic evidence (GPT-4.5, Gemini-1.5, INTELLECT-1, DisTrO), and draws out governance implications: an increased risk of compute structuring, capability proliferation, and erosion of detectability and shutdownability, while also acknowledging potential benefits such as privacy-preserving training and reduced power concentration. The paper is explicitly cautious about remaining uncertainty, particularly regarding the scalability and competitiveness of decentralised training, and it stresses that compute-based policy levers such as export controls remain relevant.

Significance. If the central claim is correct, the paper provides a valuable corrective to policy discourse that conflates geographically distributed training with fully decentralised training. The conceptual framework is clear, the treatment of benefits and risks is balanced, and the paper is transparent about the uncertainty of its projections. It is a synthesis rather than an empirical study, but it cites primary technical sources for the algorithmic developments, and its deductive chain from algorithmic progress to governance risk is logical. The main weakness is that the most consequential claim—erosion of shutdownability through peer-to-peer training—rests on a protocol archetype that has not yet been demonstrated, as detailed below.

major comments (2)
  1. [2.3.1; 3.2] In Section 2.3.1, the statement that new nodes can contribute to training 'trustlessly' and that the legitimacy of their work is 'attested through cryptographic proofs and verified by other nodes' cites Prime Intellect (2025d). That reference (Toploc) is a locality-sensitive hashing scheme for verifiable inference, not a mechanism for verifying gradient updates or checkpoint integrity during training. This matters because Section 3.2's 'true P2P network' scenario and its 'no obvious point of accountability or intervention' conclusion depend on trustless contribution to training. The authors should either cite a proof-of-learning or verifiable-training mechanism that applies to the training setting, or explicitly rephrase Section 3.2 as conditional on a protocol capability that has not yet been demonstrated. As written, the cited evidence overstates the support for the no-off-switch scenario.
  2. [2.3.1; 3.2] The 'true P2P network' scenario also assumes that a coordinator-free protocol can maintain convergence, reliable checkpointing, and sybil resistance while still constituting a single training run. Section 2.3.1 itself states that 'no training protocol has achieved a fully decentralised status as of today', and the concrete examples discussed (INTELLECT-1, DisTrO) involve centralisation under the paper's own taxonomy. The manuscript should therefore state whether the shutdownability risk is meant to apply to currently feasible architectures or to a hypothetical future architecture; in the latter case, the risk should be labelled as a scenario-analysis result rather than a developing trend. This distinction is important because the policy response to an established trend differs from the response to a plausible but unproven contingency. The paper's later remark that the competitiveness gap 'remains uncertain' (Section 2.3) does not fully resolve this, since Section 3.2 proceeds to treat the P2P case as the decisive governance challenge.
minor comments (4)
  1. [2.3] There is a typo in Section 2.3: 'decentralisedtraining' should be 'decentralised training'.
  2. [Figure 1] The fourth quadrant of Figure 1 is labelled only with '?'. The caption does explain this, but consider making clearer in the figure itself or the caption that this quadrant denotes a logically possible but currently unattested regime, rather than a missing example.
  3. [1, footnote 1] The estimate of roughly $145 million in venture funding would benefit from an explicit statement that it covers announced rounds only and the date as of which the total was computed; the footnote lists rounds spanning 2023–2025 but no aggregation date.
  4. [Appendix B] The claim that a single M3 Ultra platform with 512 GB of memory is sufficient to run inference on DeepSeek-R1 (671B parameters) should specify the precision or quantization assumed, since 512 GB is below the unquantized parameter footprint of that model and the statement as written is potentially misleading.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the governance implications are conditional extrapolations from external technical work, with the only self-citation peripheral and non-load-bearing.

full rationale

This paper is a policy synthesis rather than a derivation from fitted or self-authored premises. The central claim that low-communication algorithms enable distributed/decentralised training is supported by external, independently published results (DiLoCo, DisTrO, INTELLECT-1, Moshpit SGD), none authored by the present authors. The distributed/decentralised distinction is proposed as an explicit definitional contribution (Sec 2.1), and the shutdownability discussion is explicitly conditional: Sec 3.2 says 'it might be possible' to run a 'true P2P network as described in Sec. 2.3.1,' while Sec 2.3.1 concedes 'no training protocol has achieved a fully decentralised status as of today.' The 'no off-switch' scenario is thus an analytic consequence of the stated P2P definition plus a stated premise about compute access as an intervention point, not a prediction that is forced by fitting or by construction. The only self-citation (Egan & Heim, 2023) is used for a secondary point about the difficulty of detecting compute structuring without cross-provider information sharing; it is independent prior work and not the load-bearing premise for the paper's main governance conclusions. Remaining weaknesses, such as the cryptographic-verification citation (Prime Intellect, 2025d) concerning inference rather than training, and the paper's own admission that the competitiveness gap is 'uncertain' (Sec 2.3), are evidentiary or correctness concerns, not circularity. No equation or argument reduces to its own input by definition.

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

The paper introduces no free parameters and no invented entities; it is a synthesis. Its analytic contribution is a new terminology and taxonomy. The central claims rest on future-oriented domain assumptions about the scalability of low-communication training and community compute, which the authors often flag as open questions.

assumptions (4)
  • domain assumption Low-communication methods (DiLoCo, DeMo) match or beat standard data-parallel training at scale.
    Invoked in Sec 2.2.1 and Appendix A as the key algorithmic enabler; relies on preprints (e.g., Charles et al. 2025) that the paper does not independently verify.
  • domain assumption Community compute pools can reach frontier-relevant scales, as illustrated by Folding@Home.
    Sec 2.3 uses Folding@Home's 280k GPUs as a precedent, while conceding AI training workloads differ and leaving precise estimates to future work.
  • domain assumption Reasoning-model post-training has a very high forward-to-backward pass ratio, making RL training low-communication compatible.
    Appendix B argues this ratio (~1000:1) enables consumer-GPU participation in decentralised post-training; presented as analysis, not measured.
  • domain assumption Single-site power and construction constraints will drive corporates to distribute training across sites.
    Sec 2.2 projects multi-data-centre adoption from power scaling estimates by Sevilla et al. and Fist & Datta; these projections are uncertain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distributed and Decentralised Training: Technical Governance Challenges in a Shifting AI Landscape." pith.science (2026). https://pith.science/paper/RQ4MAGUM

@misc{pith2026250707765,
  author       = {Pith},
  title        = {Pith review of: Distributed and Decentralised Training: Technical Governance Challenges in a Shifting AI Landscape},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQ4MAGUM}},
  note         = {Machine review of arXiv:2507.07765}
}
read the original abstract

Advances in low-communication training algorithms are enabling a shift from centralised model training to compute setups that are either distributed across multiple clusters or decentralised via community-driven contributions. This paper distinguishes these two scenarios - distributed and decentralised training - which are little understood and often conflated in policy discourse. We discuss how they could impact technical AI governance through an increased risk of compute structuring, capability proliferation, and the erosion of detectability and shutdownability. While these trends foreshadow a possible new paradigm that could challenge key assumptions of compute governance, we emphasise that certain policy levers, like export controls, remain relevant. We also acknowledge potential benefits of decentralised AI, including privacy-preserving training runs that could unlock access to more data, and mitigating harmful power concentration. Our goal is to support more precise policymaking around compute, capability proliferation, and decentralised AI development.

Figures

Figures reproduced from arXiv: 2507.07765 by the authors.

Figure 1
Figure 1. Possible training regimes organised by the number of contributing parties and compute locations. Each quadrant con￾tains example models: Grok 3 represents physically centralised compute with a single coordinating party (xAI, 2025); GPT-4.5 il￾lustrates centralised coordination across many compute sites (Ope￾nAI, 2025); INTELLECT-1 exemplifies decentralised training with many contributors and locations (Jaghouar et a… view at source ↗
Figure 2
Figure 2. In data-parallel training, there are N workers which train local replicas of the model on their assigned data shards. After updating the weights, these changes need to be synchronised across all workers. In the most basic implementation, this can be done through a single ‘parameter server’, but more efficient strategies are used in practice. For example, the core idea of DiLoCo is to detach local optimisation from g… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 37 canonical work pages

  1. [1]

    and Scher, A

    Barnett, P. and Scher, A. AI governance to avoid extinction: The strategic landscape and actionable research questions, 2025. URL https://techgov.intelligence.org/research/ai-governance-to-avoid-extinction

  2. [2]

    Introducing B erkeley C ompute: Revolutionizing AI with decentralized GPU power, 2024

    Berkeley Compute . Introducing B erkeley C ompute: Revolutionizing AI with decentralized GPU power, 2024. URL https://blog.berkeleycompute.com/introducing-berkeley-compute/

  3. [3]

    Decentralized training isn't a policy nightmare -- yet, 2025

    Bye, L. Decentralized training isn't a policy nightmare -- yet, 2025. URL https://www.transformernews.ai/p/decentralized-training-policy-implications

  4. [4]

    Communication-efficient language model training scales reliably and robustly: Scaling laws for D i L o C o

    Charles, Z., Teston, G., Dery, L., Rush, K., Fallen, N., Garrett, Z., Szlam, A., and Douillard, A. Communication-efficient language model training scales reliably and robustly: Scaling laws for D i L o C o. arXiv preprint arXiv:2503.09799, 2025

  5. [5]

    Exclusive: P luralis raises \ 7.6 million from prominent investors to take on O pen AI with big decentralized models, 2025

    Crypto, F. Exclusive: P luralis raises \ 7.6 million from prominent investors to take on O pen AI with big decentralized models, 2025. URL https://fortune.com/crypto/2025/03/19/pluralis-research-7-6-million-openai-coinfund-union-square-ventures/

  6. [6]

    Gensyn AI raises \ 43 million in S eries A funding led by a16z to revolutionize decentralized machine learning, 2023

    Cryptopolitan. Gensyn AI raises \ 43 million in S eries A funding led by a16z to revolutionize decentralized machine learning, 2023. URL https://www.cryptopolitan.com/gensyn-ai-raises-43-million-in-series-a-funding-led-by-a16z/

  7. [7]

    A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., and Shen, J

    Douillard, A., Feng, Q., Rusu, A. A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., and Shen, J. D i L o C o: Distributed low-communication training of language models. arXiv preprint arXiv:2311.08105, 2023

  8. [8]

    Streaming D i L o C o with overlapping communication: Towards a distributed free lunch

    Douillard, A., Donchev, Y., Rush, K., Kale, S., Charles, Z., Garrett, Z., Teston, G., Lacey, D., McIlroy, R., Shen, J., et al. Streaming D i L o C o with overlapping communication: Towards a distributed free lunch. arXiv preprint arXiv:2501.18512, 2025

Show all 53 references
  1. [9]

    and Heim, L

    Egan, J. and Heim, L. Oversight for frontier AI through a know-your-customer scheme for compute providers. arXiv preprint arXiv:2310.13625, 2023

  2. [10]

    and Besiroglu, T

    Erdil, E. and Besiroglu, T. Introducing the distributed training interactive simulator, 2024. URL https://epoch.ai/blog/introducing-the-distributed-training-interactive-simulator

  3. [11]

    EU AI A ct: A rticle 51 -- classification of general-purpose AI models as general-purpose AI models with systemic risk, 2024

    European Parliament . EU AI A ct: A rticle 51 -- classification of general-purpose AI models as general-purpose AI models with systemic risk, 2024. URL https://artificialintelligenceact.eu/article/51/

  4. [12]

    Exo G ym: Open-source P ython toolkit to facilitate distributed AI research., 2025

    Exo Labs . Exo G ym: Open-source P ython toolkit to facilitate distributed AI research., 2025. URL https://github.com/exo-explore/gym. Accessed: 2025-05-02

  5. [13]

    and Datta, A

    Fist, T. and Datta, A. Compute in A merica: A policy playbook, 2025. URL https://ifp.org/special-compute-zones/

  6. [14]

    I., Burnell, R., Bai, L., Gulati, A., Tanzer, G., Vincent, D., Pan, Z., Wang, S., et al

    Gemini Team , Georgiev, P., Lei, V. I., Burnell, R., Bai, L., Gulati, A., Tanzer, G., Vincent, D., Pan, Z., Wang, S., et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024

  7. [15]

    A sober look at progress in language model reasoning: Pitfalls and paths to reproducibility

    Hochlehnert, A., Bhatnagar, H., Udandarao, V., Albanie, S., Prabhu, A., and Bethge, M. A sober look at progress in language model reasoning: Pitfalls and paths to reproducibility. arXiv preprint arXiv:2504.07086, 2025

  8. [16]

    M., Basra, M., Obeid, F., Straube, J., Keiblinger, M., Bakouch, E., Atkins, L., Panahi, M., Goddard, C., et al

    Jaghouar, S., Ong, J. M., Basra, M., Obeid, F., Straube, J., Keiblinger, M., Bakouch, E., Atkins, L., Panahi, M., Goddard, C., et al. INTELLECT -1 technical report. arXiv preprint arXiv:2412.01152, 2024

  9. [17]

    Eager updates for overlapped communication and computation in D i L o C o

    Kale, S., Douillard, A., and Donchev, Y. Eager updates for overlapped communication and computation in D i L o C o. arXiv preprint arXiv:2502.12996, 2025

  10. [18]

    On the societal impact of open foundation models

    Kapoor, S., Bommasani, R., Klyman, K., Longpre, S., Ramaswami, A., Cihon, P., Hopkins, A., Bankston, K., Biderman, S., Bogen, M., et al. On the societal impact of open foundation models. arXiv preprint arXiv:2403.07918, 2024

  11. [19]

    Unpacking decentralized training, 2025

    KNOWER. Unpacking decentralized training, 2025. URL https://theknower.substack.com/p/unpacking-decentralized-training

  12. [20]

    Gradual disempowerment: Systemic existential risks from incremental AI development

    Kulveit, J., Douglas, R., Ammann, N., Turan, D., Krueger, D., and Duvenaud, D. Gradual disempowerment: Systemic existential risks from incremental AI development. arXiv preprint arXiv:2501.16946, 2025

  13. [21]

    Flower L abs raises \ 20 M S eries A , 2024

    Labs, F. Flower L abs raises \ 20 M S eries A , 2024. URL https://flower.ai/blog/2024-02-15-announcing-series-a/

  14. [22]

    and Drago, L

    Laine, R. and Drago, L. The intelligence curse, 2025. URL https://intelligence-curse.ai/breaking/

  15. [23]

    Unpacking decentralized training, 2025

    Lehman, S. Unpacking decentralized training, 2025. URL https://www.symbolic.capital/writing/frontier-training

  16. [24]

    A., Shen, J., Szlam, A., and Ranzato, M

    Liu, B., Chhaparia, R., Douillard, A., Kale, S., Rusu, A. A., Shen, J., Szlam, A., and Ranzato, M. Asynchronous local- sgd training for language modeling. arXiv preprint arXiv:2401.09135, 2024

  17. [25]

    Protocol learning, decentralized frontier risk and the no-off problem

    Long, A. Protocol learning, decentralized frontier risk and the no-off problem. arXiv preprint arXiv:2412.07890, 2024

  18. [26]

    Nous D is T r O , 2024

    Nous Research . Nous D is T r O , 2024. URL https://distro.nousresearch.com/

  19. [27]

    Open AI GPT-4.5 S ystem C ard, 2025

    OpenAI. Open AI GPT-4.5 S ystem C ard, 2025. URL https://openai.com/index/gpt-4-5-system-card/

  20. [28]

    Introduction to GPT-4.5

    OpenAI . Introduction to GPT-4.5 . https://www.youtube.com/live/cfRYp0nItZ8?t=457s, 2025. Accessed April 29, 2025. Quoted at 7:37 (457 seconds)

  21. [29]

    Patel, D., Nishball, D., and Ontiveros, J. E. Multi-datacenter training: Open AI ’s ambitious plan to beat G oogle’s infrastructure, 2024. URL https://semianalysis.com/2024/09/04/multi-datacenter-training-openais/

  22. [30]

    Peng, B., Quesnelle, J., and Kingma, D. P. Decoupled momentum optimization. arXiv preprint arXiv:2411.19870, 2024

  23. [31]

    F., Heim, L., and Brown, N

    Pilz, K. F., Heim, L., and Brown, N. Increased compute efficiency and the diffusion of AI capabilities. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 27582--27590, 2025 a

  24. [32]

    F., Sanders, J., Rahman, R., and Heim, L

    Pilz, K. F., Sanders, J., Rahman, R., and Heim, L. Trends in AI supercomputers, 2025 b . URL https://arxiv.org/abs/2504.16026

  25. [33]

    He who can pay top dollar for HBM memory controls ai training, 2024

    Prickett Morgan, T. He who can pay top dollar for HBM memory controls ai training, 2024. URL https://www.nextplatform.com/2024/02/27/he-who-can-pay-top-dollar-for-hbm-memory-controls-ai-training/

  26. [34]

    Planetary-scale inference: Previewing our peer-to-peer decentralized inference stack, 2025 a

    Prime Intellect . Planetary-scale inference: Previewing our peer-to-peer decentralized inference stack, 2025 a . URL https://www.primeintellect.ai/blog/inference

  27. [35]

    Decentralized training in the inference-time-compute paradigm, 2025 b

    Prime Intellect . Decentralized training in the inference-time-compute paradigm, 2025 b . URL https://www.primeintellect.ai/blog/intellect-math

  28. [36]

    Introducing P rime I ntellect’s protocol & testnet: A peer-to-peer compute and intelligence network, 2025 c

    Prime Intellect . Introducing P rime I ntellect’s protocol & testnet: A peer-to-peer compute and intelligence network, 2025 c . URL https://www.primeintellect.ai/blog/protocol

  29. [37]

    Toploc - a locality sensitive hashing scheme for trustless verifiable inference, 2025 d

    Prime Intellect . Toploc - a locality sensitive hashing scheme for trustless verifiable inference, 2025 d . URL https://www.primeintellect.ai/blog/toploc

  30. [38]

    \ 15m to build a peer-to-peer AI protocol, 2025 e

    Prime Intellect . \ 15m to build a peer-to-peer AI protocol, 2025 e . URL https://www.primeintellect.ai/blog/fundraise

  31. [39]

    H ivemind: D ecentralized D eep L earning in P y T orch, April 2020

    Ryabinin, M., Borzunov, A., Diskin, M., Gusev, A., Mazur, D., Plokhotnyuk, V., Bukhtiyarov, A., Samygin, P., Sinitsin, A., and Chumachenko, A. H ivemind: D ecentralized D eep L earning in P y T orch, April 2020. URL https://github.com/learning-at-home/hivemind

  32. [40]

    Moshpit SGD : Communication-efficient decentralized training on heterogeneous unreliable devices

    Ryabinin, M., Gorbunov, E., Plokhotnyuk, V., and Pekhimenko, G. Moshpit SGD : Communication-efficient decentralized training on heterogeneous unreliable devices. Advances in Neural Information Processing Systems, 34: 0 18195--18211, 2021

  33. [41]

    F., Aleksandrov, P., Qiu, X., et al

    Sani, L., Iacob, A., Cao, Z., Marino, B., Gao, Y., Paulik, T., Zhao, W., Shen, W. F., Aleksandrov, P., Qiu, X., et al. The future of large language model pre-training is federated. arXiv preprint arXiv:2405.10853, 2024

  34. [42]

    K., Ngo, R., Pilz, K., et al

    Sastry, G., Heim, L., Belfield, H., Anderljung, M., Brundage, M., Hazell, J., O'Keefe, C., Hadfield, G. K., Ngo, R., Pilz, K., et al. Computing power and the governance of artificial intelligence. arXiv preprint arXiv:2402.08797, 2024

  35. [43]

    and Fist, T

    Seferis, M. and Fist, T. Detecting compute structuring in AI governance is likely feasible, 2025

  36. [44]

    and Roldán, E

    Sevilla, J. and Roldán, E. Training compute of frontier AI models grows by 4-5x per year, 2024. URL https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year. Accessed: 2025-04-29

  37. [45]

    Can AI scaling continue through 2030?, 2024

    Sevilla, J., Besiroglu, T., Cottier, B., You, J., Roldán, E., Villalobos, P., and Erdil, E. Can AI scaling continue through 2030?, 2024. URL https://epoch.ai/blog/can-ai-scaling-continue-through-2030. Accessed: 2025-04-29

  38. [46]

    The shutdown problem: an AI engineering puzzle for decision theorists

    Thornley, E. The shutdown problem: an AI engineering puzzle for decision theorists. Philosophical Studies, pp.\ 1--28, 2024

  39. [47]

    Nonproliferation is the wrong approach to AI misuse, 2025

    Toner, H. Nonproliferation is the wrong approach to AI misuse, 2025. URL https://helentoner.substack.com/p/nonproliferation-is-the-wrong-approach

  40. [48]

    Nous R esearch secures \ 50m from P aradigm to build decentralized AI on S olana, 2025

    Trading View . Nous R esearch secures \ 50m from P aradigm to build decentralized AI on S olana, 2025. URL https://www.tradingview.com/news/cointelegraph:54f1a44d2094b:0-nous-research-secures-50m-from-paradigm-to-build-decentralized-ai-on-solana/

  41. [49]

    Interim report: Governing AI for humanity, 2023

    United Nations . Interim report: Governing AI for humanity, 2023. URL https://www.un.org/sites/un2.un.org/files/un_ai_advisory_body_governing_ai_for_humanity_interim_report.pdf

  42. [50]

    Vana and F lower L abs partner to build the world’s first user-owned AI model, 2025

    Vana. Vana and F lower L abs partner to build the world’s first user-owned AI model, 2025. URL https://www.vana.org/posts/vana-flower-labs-partnership

  43. [51]

    Grok 3 B eta — the age of reasoning agents, 2025

    xAI. Grok 3 B eta — the age of reasoning agents, 2025. URL https://x.ai/news/grok-3

  44. [52]

    A guide to dark fiber networks, 2024

    Zayo. A guide to dark fiber networks, 2024. URL https://www.zayo.com/resources/a-guide-to-dark-fiber/

  45. [53]

    I., Porter, J

    Zimmerman, M. I., Porter, J. R., Ward, M. D., Singh, S., Vithani, N., Meller, A., Mallimadugula, U. L., Kuhn, C. E., Borowsky, J. H., Wiewiora, R. P., Hurley, M. F. D., Harbison, A. M., Fogarty, C. A., Coffland, J. E., Fadda, E., Voelz, V. A., Chodera, J. D., and Bowman, G. R....

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.