Pith. sign in

REVIEW 2 major objections 5 minor 48 references

Brief analysis of DeepSeek R1 and its implications for Generative AI

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This report argues that DeepSeek-R1's capability gains come from algorithmic efficiency—Mixture of Experts, reinforcement learning, and distillation—rather than brute-force scaling, making competitive reasoning models dramatically cheaper…

desk verdict A well-written briefing note on DeepSeek R1 with no new research; the cost advantage claim is shakier than the framing suggests. read the letter →

arxiv 2502.02523 v3 pith:6LBQ6BIO submitted 2025-02-04 cs.LG

Sarah Mercer , Samuel Spillard , Daniel P. Martin This is my paper · ORCID
classification cs.LG
keywords DeepSeekR1MixtureofExpertsreinforcementlearningmodeldistillationchain-of-thoughtreasoningalgorithmicefficiencyopen-weightmodelsAIcostanalysis
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 report argues that DeepSeek-R1's competitive performance at roughly one-fiftieth the reported training cost of comparable models is best explained by algorithmic choices—Mixture of Experts architecture, pure reinforcement learning with Group Relative Policy Optimization, and distillation of reasoning traces into smaller models—rather than by a step change in compute. If the authors are right, the bottleneck for frontier-level reasoning has shifted from access to hardware toward access to training recipes, making competitive models cheaper and more widely available. The report then traces consequences: price cuts and product releases from incumbent labs, questions about whether distillation from closed models is happening, and a lowered barrier for smaller actors with security and safety implications. It is explicitly a rapid think piece and concedes that vendor-reported figures are hard to verify.

What carries the argument

The load-bearing mechanism is the combination of Mixture of Experts (splitting a large model into specialized sub-models to cut training cost), Group Relative Policy Optimization (a simple reinforcement learning scheme that rewards correct reasoning without a separate reward model), and distillation (transferring long chain-of-thought reasoning from big models into much smaller ones). The report also highlights emergent self-reflection in chain-of-thought traces as a phenomenon that carries much of the capability gain and needs further study.

What would settle it

An independent audit of DeepSeek-V3's training compute (via GPU procurement, power records, or architecture-based FLOP estimates) that showed total cost near the hundreds of millions would undercut the efficiency claim. Likewise, evidence that R1's AIME performance requires traces or data from closed models, or disappears when the pure-RL recipe is run without such data, would falsify the claim that simple RL on a strong base model is the source of the capability.

Watch

Extended reading notes

Core claim

The paper's central claim is that the wave of Chinese reasoning models—DeepSeek-R1 above all—demonstrates that algorithmic efficiency and clever engineering can substitute for brute-force scaling. DeepSeek-V3's Mixture of Experts design and reported $5.6 million training cost, R1's pure-RL route to chain-of-thought reasoning, and the finding that reasoning can be distilled into small models together indicate a different cost curve than the field assumed. The authors treat this as a technical response to data and compute limits, and as evidence that open-weight models may now be 'good enough' for widespread adoption. They do not claim to settle the truth of the numbers; they frame the report as introductory material that identifies open questions.

Load-bearing premise

The assessment depends on DeepSeek's reported numbers—the $5.6 million training cost, benchmark scores around 79.8% on AIME, and the distillation results—being accurate and comparable; the authors concede they cannot verify these figures.

Editorial extensions

If this is right

  • If the efficiency narrative holds, frontier-competitive training costs fall by roughly an order of magnitude, weakening the assumption that only actors with vast GPU fleets can build top models.
  • Open-weight releases plus distillation place meaningful reasoning ability on local machines and small devices, changing privacy, adoption, and threat baselines.
  • Incumbents respond with price cuts and product releases, and may restrict access to reasoning traces, potentially pushing users toward open alternatives.
  • Simple RL with small datasets demonstrates that verifiable skills like math and coding can be upskilled cheaply, though the paper questions whether the recipe extends to open-ended tasks.
  • Distillation of reasoning from large to small models yields better small models than RL alone, and raises unresolved questions about whether model values survive distillation intact.

Reading between the lines

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

  • If the same RL-upskilling recipe transfers to other verifiable task families, expect a wave of cheaply specialized small models for formal verification, tool use, and structured data work, bypassing general-purpose scaling.
  • The report's own caveat about unverifiable costs suggests a direct test: estimate DeepSeek-V3's training FLOPs from its published architecture and release timeline, then compare against the reported $5.6 million figure to bound the true cost.
  • Distillation may transfer not only reasoning but also alignment behavior, meaning the personality and refusal patterns of a closed teacher model could propagate into open-weight derivatives without any public documentation.
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 / 5 minor

Summary. The paper is a short expository think piece about DeepSeek R1 and recent Chinese reasoning models. It summarizes DeepSeek-V3 and R1, the GRPO/RL recipe, the reported emergence of self-reflection in R1-Zero, distillation into smaller models, and independent replication efforts (HKUST simple RL, Hugging Face Open R1). It then surveys Doubao, iFlytek Spark X1, Kimi k1.5, Qwen2.5-VL, and OpenAI Deep Research, and discusses market reactions, security concerns, and political implications. The central claim is that competitive reasoning performance can be achieved at a small fraction of the cost of prior frontier models, through algorithmic efficiency (MoE, RL, distillation) rather than brute-force scaling.

Significance. If its cost-effectiveness claim holds, the paper points to a genuine shift in the affordability and accessibility of frontier-level reasoning models, with implications for export controls, market concentration, and risk thresholds for misuse. The paper's value is as a timely, broad survey rather than a technical contribution: it contains no original experiments or formal analysis, but it usefully collects sources and highlights independent replication (HKUST 7B RL, HuggingFace Open R1) that supports the algorithmic-efficiency thesis. Its candour about uncertainty and its identification of research gaps (data provenance, effect of distillation on values, emergence of CoT) are useful. However, the headline quantitative claim about cost is built on vendor-reported figures and is not reconciled with other evidence printed in the paper.

major comments (2)
  1. [Abstract and §1; §4.2 and §5] The central claim that R1 was 'developed at a fraction of the cost' and that the $5.6M figure is '1/50th of the costs of other comparable models' is load-bearing, but the paper never defines or defends the cost basis. The $5.6M is a vendor-reported cost for one final training run of V3, not the full cost of the R&D program, hardware, data, salaries, electricity, or prior models that produced R1. Section 4.2 itself cites Alexandr Wang's estimate of ~50,000 H100 GPUs, which implies an infrastructure investment far larger than $5.6M, yet the paper does not integrate this into the cost comparison. Section 5 concedes that 'it is obviously difficult for us to provide accurate and reliable conclusions,' but the abstract and introduction nevertheless present the 'fraction of cost' framing unqualified. The authors should either restate the claim as a marginal training-run cost estimate with explicit scope, or substantially revise the strength of the cost-advantage conclusion.
  2. [§2.2 and §5] The benchmark narrative needs technical precision. The text says R1-Zero improved from 15.6% to 71.0% on AIME 2024, 'comparable to openAI-o1-0912, which was then exceeded when the DeepSeek team tweaked the RL (majority voting) scoring 86.7%.' This leaves unclear whether the 86.7% and the final R1 79.8% are pass@1 or majority-vote scores, and how the comparison to o1 was made. Since the paper's claim of 'competitive' performance depends on these numbers, the authors should specify the evaluation protocol and cite the original tables.
minor comments (5)
  1. [§2.2] 'openAI' should be 'OpenAI' for consistency.
  2. [§3] 'Duobao' should be 'Doubao' in both places where it appears.
  3. [References] Reference [9] is attributed to 'Dia et al.'; the published DeepSeekMoE paper is by Dai et al., so the citation should be corrected.
  4. [§4.3] The phrase 'a recent data breach' for the Wiz Research finding may overstate what was reported: the researchers described an exposed/misconfigured database rather than a confirmed external breach. Consider wording that matches the source more neutrally.
  5. [§4.2] Several observations such as 'it can be brittle, and difficult to prompt' and the 'self-doubt' characterization of CoT are anecdotal; they should be attributed or explicitly labelled as community impressions rather than verified findings.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found: the paper is an expository commentary that makes no fitted predictions or self-referential derivations.

full rationale

This paper is a descriptive think piece about DeepSeek R1 and related models. It contains no equations, no fitted parameters, no trained models, and no derivation chain of its own. The central claims—that DeepSeek-R1 was developed at low cost and is competitive with OpenAI's models—are taken from vendor reports, press coverage, and external replication efforts, not from any argument internal to the paper. The authors explicitly disclaim certainty about cost and pipeline details in Section 5, noting it is 'obviously difficult for us to provide accurate and reliable conclusions.' That is an acknowledged limitation about external data quality, not circularity. There are no self-citations used as load-bearing evidence: the cited DeepSeek technical reports, HKUST replication, HuggingFace Open R1, and press items are external sources. Even if one doubts the vendor-reported cost figures, that is a factual accuracy concern, not a circularity concern. The paper does not define any concept in terms of its conclusion, does not fit parameters to data and then re-predict the same data, and does not import a uniqueness theorem from the authors' own prior work. Therefore the circularity score is 0.

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

The paper introduces no fitted parameters and no invented entities. Its conclusions rest on the accuracy of external reports: DeepSeek model papers, vendor statements, press coverage, and social media posts. The axiom entries list those background assumptions.

assumptions (3)
  • domain assumption Publicly reported cost, benchmark, and pipeline figures (e.g., $5.6M training cost, AIME scores, GRPO details) are accurate enough for the comparisons.
    Sections 1 and 2.2 use these numbers as the basis for the cost advantage and competitiveness claims.
  • domain assumption The reports of emergent self-reflection and chain-of-thought behavior in R1-Zero and the 7B replication describe real model behavior, not artifacts or cherry-picked examples.
    Section 2.2 treats the 'aha moment' as a key finding requiring further research.
  • domain assumption External reactions cited from press and social media (e.g., the 50,000 H100 claim, OpenAI distillation allegation) are attributed accurately.
    Section 4 summarizes these reactions without independent verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Brief analysis of DeepSeek R1 and its implications for Generative AI." pith.science (2026). https://pith.science/paper/6LBQ6BIO

@misc{pith2026250202523,
  author       = {Pith},
  title        = {Pith review of: Brief analysis of DeepSeek R1 and its implications for Generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6LBQ6BIO}},
  note         = {Machine review of arXiv:2502.02523}
}
read the original abstract

In late January 2025, DeepSeek released their new reasoning model (DeepSeek R1); which was developed at a fraction of the cost yet remains competitive with OpenAI's models, despite the US's GPU export ban. This report discusses the model, and what its release means for the field of Generative AI more widely. We briefly discuss other models released from China in recent weeks, their similarities; innovative use of Mixture of Experts (MoE), Reinforcement Learning (RL) and clever engineering appear to be key factors in the capabilities of these models. This think piece has been written to a tight timescale, providing broad coverage of the topic, and serves as introductory material for those looking to understand the model's technical advancements, as well as its place in the ecosystem. Several further areas of research are identified.

Figures

Figures reproduced from arXiv: 2502.02523 by the authors.

Figure 1
Figure 1. A comparison of model outputs to highlight value differences between the two models 4.3 Political Commentary Many have commented on the model’s refusal to answer questions on certain topics, related to the censorship of the CCP [40]. From a national security point of view, this raises several concerns. In particular, how the risk profile changes if the majority of users go from using an American aligned LLM, to a CC… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 41 canonical work pages

  1. [1]

    DeepSeek Homepage,

    DeepSeek, “DeepSeek Homepage,” 2025, accessed: 2025-02-03. [Online]. Available: https: //www.deepseek.com/

    work page 2025

  2. [2]

    DeepSeek-V3 Technical Report,

    DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv, December 27 2024. [Online]. Available: https://arxiv.org/abs/2412.19437

    arXiv 2024

  3. [3]

    Nvidia drops nearly 17% as China’s cheaper AI model DeepSeek sparks global tech sell-off,

    J. Reid, “Nvidia drops nearly 17% as China’s cheaper AI model DeepSeek sparks global tech sell-off,” CNBC, January 27 2025, accessed: 2025-02-03. [Online]. Available: https://www .cnbc.com/2025/01/ 27/nvidia-falls-10percent-in-premarket-trading-as-chinas-deepseek-triggers-global-tech-sell-off .html

    work page 2025

  4. [4]

    Nvidia shares sink as Chinese AI app spooks markets,

    P. Hoskins and I. Rahman-Jones, “Nvidia shares sink as Chinese AI app spooks markets,” BBC News, January 27 2025, accessed: 2025-02-03. [Online]. Available: https://www .bbc.co.uk/news/articles/ c0qw7z2v1pgo

    work page 2025

  5. [5]

    The race for

    G. Marcus, “The race for "AI Supremacy" is over - at least for now,” Marcus on AI (Substack), January 26 2025, accessed: 2025-02-03. [Online]. Available: https://garymarcus .substack.com/p/the-race-for-ai- supremacy-is-over

    work page 2025

  6. [6]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,

    DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning,” arXiv, vol. abs/2501.12948, January 22 2025. [Online]. Available: https://arxiv.org/abs/2501.12948

    arXiv 2025

  7. [7]

    China’s Cheap, Open AI Model DeepSeek Thrills Scientists,

    E. Gibney , “China’s Cheap, Open AI Model DeepSeek Thrills Scientists,” Nature, 2025, accessed: Feb. 3,

    work page 2025

  8. [8]

    Mixtral of Experts,

    Jiang et al., “Mixtral of Experts,” arXiv, vol. abs/2401.04088, Jan. 2024. [Online]. Available: https://arxiv.org/abs/2401.04088

    arXiv 2024

Show all 48 references
  1. [9]

    DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models,

    Dia et al., “DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models,” arXiv, vol. abs/2401.06066, Jan. 2024. [Online]. Available: https://arxiv.org/abs/2401.06066

    2024 arXiv

  2. [10]

    Simple Reinforcement Learning for Reasoning,

    HKUST-NLP, “Simple Reinforcement Learning for Reasoning,” GitHub repository , 2025, accessed: Jan. 28, 2025. [Online]. Available: https://github.com/hkust-nlp/simpleRL-reason

    2025

  3. [11]

    7B Model and 8K Examples: Emerging Reasoning with Reinforcement Learning is Both Effective and Efficient,

    Zeng et al., “7B Model and 8K Examples: Emerging Reasoning with Reinforcement Learning is Both Effective and Efficient,” Notion, Jan. 25 2025, accessed: Feb. 3, 2025. [Online]. Available: https://hkust-nlp.notion.site/simplerl-reason

    2025

  4. [12]

    rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking,

    Guan et al., “rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking,” arXiv, vol. abs/2501.04519, Jan. 8 2025. [Online]. Available: https://arxiv.org/abs/2501.04519

    2025 arXiv

  5. [13]

    Open R1,

    Hugging Face, “Open R1,” GitHub repository , 2025, accessed: Jan. 31, 2025. [Online]. Available: https://github.com/huggingface/open-r1

    2025

  6. [14]

    Doubao-1.5-Pro,

    Doubao Team, “Doubao-1.5-Pro,” Available at: https://team .doubao.com/en/special/doubao_1_5_pro, 2025, accessed: Jan. 27, 2025

    2025

  7. [15]

    ByteDance AI Introduces Doubao-1.5-Pro Language Model with a ‘Deep Think- ing’ Mode and Matches GPT-4o and Claude 3.5 Sonnet Benchmarks at 50x Cheaper,

    A. Razzaq, “ByteDance AI Introduces Doubao-1.5-Pro Language Model with a ‘Deep Think- ing’ Mode and Matches GPT-4o and Claude 3.5 Sonnet Benchmarks at 50x Cheaper,” MarkTechPost, Jan. 25 2025, accessed: Jan. 27, 2025. [Online]. Available: https: //www.marktechpost.com/2025/01/...

    2025

  8. [16]

    DeepSeek Has Rattled the AI Industry . Here’s a Look at Other Chinese AI Models,

    C. ZenSoo, “DeepSeek Has Rattled the AI Industry . Here’s a Look at Other Chinese AI Models,” TIME, Jan. 28 2025, accessed: Jan. 28, 2025. [Online]. Available: https://time .com/7210521/deepseek- chinese-ai-models/

    2025

  9. [17]

    Efficient and Accurate Prompt Optimization: The Benefit of Memory in Exemplar-Guided Reflection,

    Yan et al., “Efficient and Accurate Prompt Optimization: The Benefit of Memory in Exemplar-Guided Reflection,” arXiv, vol. abs/2411.07446, Nov. 2024. [Online]. Available: https://arxiv .org/pdf/ 2411.07446 Page 7 of 9

    2024 arXiv

  10. [18]

    LSH-MoE: Communication-Efficient MoE Training via Locality-Sensitive Hashing,

    Nie et al., “LSH-MoE: Communication-Efficient MoE Training via Locality-Sensitive Hashing,” arXiv, vol. abs/2411.08446, Nov. 2024. [Online]. Available: https://arxiv.org/abs/2411.08446

    2024 arXiv

  11. [19]

    iFlytek Releases the Xunfei Spark Deep Reasoning Model X1,

    AIbase, “iFlytek Releases the Xunfei Spark Deep Reasoning Model X1,” Available at: https:// www.aibase.com/news/14723, 2025, accessed: Jan. 28, 2025

    2025

  12. [20]

    Kimi k1.5,

    Kimi Team, “Kimi k1.5,” GitHub repository , 2025, accessed: Jan. 28, 2025. [Online]. Available: https://github.com/MoonshotAI/Kimi-k1.5

    2025

  13. [21]

    Kimi k1.5: Scaling Reinforcement Learning with LLMs,

    KimiTeam et al., “Kimi k1.5: Scaling Reinforcement Learning with LLMs,” arXiv, vol. abs/2501.12599, Jan. 22 2025. [Online]. Available: https://arxiv.org/abs/2501.12599

    2025 arXiv

  14. [22]

    Kimi k1.5: How China’s New AI Powerhouse is Redefining Multimodal Reasoning and Beating OpenAI’s o1,

    Ashley , “Kimi k1.5: How China’s New AI Powerhouse is Redefining Multimodal Reasoning and Beating OpenAI’s o1,” Medium, 2025, accessed: Jan. 28, 2025. [Online]. Avail- able: https://medium .com/@ashinno43/kimi-k1-5-how-this-next-gen-ai-model-is-revolutionizing- multimodal-reas...

    2025

  15. [23]

    Qwen2.5-VL

    Qwen, “Qwen2.5-VL.” [Online]. Available: https://github .com/QwenLM/Qwen2.5-VL/blob/main/ README.md

  16. [24]

    Introducing Deep Research

    OpenAI, “Introducing Deep Research.” [Online]. Available: https://openai .com/index/introducing- deep-research/

  17. [25]

    OpenAI ’reviewing’ allegations that its AI models were used to make DeepSeek,

    M. Sweney and D. Milmo, “OpenAI ’reviewing’ allegations that its AI models were used to make DeepSeek,” The Guardian , Jan. 29 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www.theguardian.com/technology/2025/jan/29/openai-chatgpt-deepseek-china-us-ai-models

    2025

  18. [26]

    OpenAI o3-mini,

    OpenAI, “OpenAI o3-mini,” Available at: https://openai .com/index/openai-o3-mini/, Jan. 31 2025

    2025

  19. [27]

    OpenAI Releases Its New o3-mini Reasoning Model for Free,

    S. J. Mulligan, “OpenAI Releases Its New o3-mini Reasoning Model for Free,” MIT Technology Review, Jan. 31 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www .technologyreview.com/2025/ 01/31/1110757/openai-makes-its-reasoning-model-for-free/

    2025

  20. [28]

    China’s DeepSeek AI Shakes Industry and Dents America’s Swagger,

    L. Jamali, “China’s DeepSeek AI Shakes Industry and Dents America’s Swagger,”BBC News, Jan. 28 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www.bbc.co.uk/news/articles/cd643wx888qo

    2025

  21. [29]

    CHIPS and Science Act,

    Wikipedia, “CHIPS and Science Act,” Available at: https://en .wikipedia.org/wiki/ CHIPS_and_Science_Act, 2025, accessed: Jan. 28, 2025

    2025

  22. [30]

    DeepSeek: The Chinese AI App That Has the World Talking,

    N. Ng, B. Drenon, T. Gerken, and M. Cieslak, “DeepSeek: The Chinese AI App That Has the World Talking,” BBC News , Jan. 27 2025, accessed: Jan. 27, 2025. [Online]. Available: https://www.bbc.co.uk/news/articles/c5yv5976z9po

    2025

  23. [31]

    DeepSeek,

    DeepSeek-AI, “DeepSeek,” Hugging Face, 2025, accessed: Jan. 27, 2025. [Online]. Available: https://huggingface.co/deepseek-ai

    2025

  24. [32]

    deepseek-r1,

    Ollama, “deepseek-r1,” Available at: https://ollama .com/library/deepseek-r1, 2025, accessed: Jan. 27, 2025

    2025

  25. [33]

    Someone on X Claims to Have Jailbroken R1 by Invoking the Name of Pliny , a Renowned LLM Jailbreaker,

    T. Kellog, “Someone on X Claims to Have Jailbroken R1 by Invoking the Name of Pliny , a Renowned LLM Jailbreaker,” BlueSky , Jan. 24 2025, accessed: Jan. 27, 2025. [Online]. Available: https://bsky.app/profile/timkellogg.me/post/3lgj25q42w22h

    2025

  26. [34]

    DeepSh*t: Exposing the Security Risks of DeepSeek-r1,

    Martin et al., “DeepSh*t: Exposing the Security Risks of DeepSeek-r1,” Hidden Layer , Jan. 30 2025, accessed: Feb. 1, 2025. [Online]. Available: https://hiddenlayer .com/innovation-hub/deepsht- exposing-the-security-risks-of-deepseek-r1/

    2025

  27. [35]

    Recent Jailbreaks Demonstrate Emerging Threat to DeepSeek,

    K. Wilhoit, “Recent Jailbreaks Demonstrate Emerging Threat to DeepSeek,” Palo Alto Networks, Jan. 30 2025, accessed: Feb. 1, 2025. [Online]. Available: https://unit42 .paloaltonetworks.com/jailbreaking- deepseek-three-techniques/

    2025

  28. [36]

    DeepSeek FAQ,

    B. Thompson, “DeepSeek FAQ,” Stratechery, Jan. 27 2025, accessed: Jan. 28, 2025. [Online]. Available: https://stratechery.com/2025/deepseek-faq/ Page 8 of 9

    2025

  29. [37]

    Billionaire and Scale AI CEO Alexandr Wang: DeepSeek Has About 50,000 NVIDIA H100s That They Can’t Talk About Because of the US Export Controls That Are in Place,

    @kimmonismus, “Billionaire and Scale AI CEO Alexandr Wang: DeepSeek Has About 50,000 NVIDIA H100s That They Can’t Talk About Because of the US Export Controls That Are in Place,” X (formerly Twitter), Jan. 24 2025, accessed: Feb. 3, 2025. [Online]. Available: https://x.com/kim...

    2025

  30. [38]

    With R1, a Lot of People Have Been Asking How Come We Didn’t Discover This 2 Years Ago?

    @its_dibya, “With R1, a Lot of People Have Been Asking How Come We Didn’t Discover This 2 Years Ago?” X (formerly Twitter), Jan. 26 2025, accessed: Feb. 3, 2025. [Online]. Available: https://x.com/its_dibya/status/1883595705736163727

    2025

  31. [39]

    The Specific RL Alg Doesn’t Matter Much. . . ,

    @jiayi_pirate, “The Specific RL Alg Doesn’t Matter Much. . . ,” X (formerly Twitter), Jan. 24 2025, accessed: Feb. 3, 2025. [Online]. Available: https://x.com/jiayi_pirate/status/1882839504899420517

    2025

  32. [40]

    All About DeepSeek – The Chinese AI Startup Challenging US Big Tech,

    J. MSV, “All About DeepSeek – The Chinese AI Startup Challenging US Big Tech,” Forbes, Jan. 26 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www.forbes.com/sites/janakirammsv/2025/01/26/ all-about-deepseekthe-chinese-ai-startup-challenging-the-us-big-tech

    2025

  33. [41]

    Announcing The Stargate Project,

    OpenAI, “Announcing The Stargate Project,” OpenAI Blog , Jan. 21 2025, accessed: Feb. 3 2025. [Online]. Available: https://openai.com/index/announcing-the-stargate-project/

    2025

  34. [42]

    OpenAI Says Chinese Rivals Using Its Work for Their AI Apps,

    J. de Silva and G. Fraser, “OpenAI Says Chinese Rivals Using Its Work for Their AI Apps,”BBC News, 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www.bbc.co.uk/news/articles/c9vm1m8wpr9o

    2025

  35. [43]

    Be Careful with DeepSeek Australia Says – So Is It Safe to Use?

    T. Gerken, “Be Careful with DeepSeek Australia Says – So Is It Safe to Use?” BBC News, 2025, accessed: Jan. 28, 2025. [Online]. Available: https://www.bbc.co.uk/news/articles/cx2k7r5nrvpo

    2025

  36. [44]

    New DeepSeek Warning — Do You Need To Delete Your iPhone, Android App?

    Z. Doffman, “New DeepSeek Warning — Do You Need To Delete Your iPhone, Android App?” Forbes, Jan. 30 2025, accessed: Feb. 3, 2025. [Online]. Available: https://www .forbes.com/sites/zakdoffman/ 2025/01/30/new-deepseek-warning-do-you-need-to-delete-your-iphone-android-app/

    2025

  37. [45]

    DeepSeek blocked on Apple and Google app stores in Italy,

    E. Pollina, “DeepSeek blocked on Apple and Google app stores in Italy,” Reuters, Jan. 29 2025, accessed: Feb 3, 2025. [Online]. Available: https://www .reuters.com/technology/deepseek-app-unavailable- apple-google-app-stores-italy-2025-01-29/

    2025

  38. [46]

    Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information, Including Chat History,

    G. Nagli, “Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information, Including Chat History,” Jan. 29 2025, accessed: Feb 3, 2025. [Online]. Available: https: //www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak

    2025

  39. [47]

    European AI alliance unveils LLM alternative to Silicon Valley and DeepSeek,

    T. Macaulay , “European AI alliance unveils LLM alternative to Silicon Valley and DeepSeek,” The Next Web, Feb. 3 2025, accessed: Feb. 3, 2025. [Online]. Available: https://thenextweb.com/news/european- ai-alliance-openeurollm-challenges-us-china Page 9 of 9

    2025

  40. [2025]

    Available: https://www.nature.com/articles/d41586-025-00229-6

    [Online]. Available: https://www.nature.com/articles/d41586-025-00229-6

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.