If open source is to win, it must go public

David Pham; Jenia Jitsev; Joseph Low; Joshua Tan; Katherine Elkins; Magnus Sahlgren; Nicholas Vincent; Sampo Pyysalo

REVIEW 3 major objections 4 minor 12 references

If open source is to win, it must go public

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Open source AI will not democratize access on its own; it must be embedded in public AI infrastructure that provides public support, public access, public accountability, and private commitments.

desk verdict A clear, honest position paper that overstates the certainty of its central claim but deserves peer review for its taxonomy and fair treatment of counterarguments. read the letter →

arxiv 2507.09296 v2 pith:GGS2LJ2J submitted 2025-07-12 cs.CY

Joshua Tan , Nicholas Vincent , Katherine Elkins , Magnus Sahlgren , Joseph Low , David Pham , Sampo Pyysalo , Jenia Jitsev This is my paper · ORCID

classification cs.CY

keywords opensourceAIpublicweightsgovernancegoodsactivationgapinfrastructuredigital

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

Open source AI, the paper argues, is not like earlier open source software. Releasing model weights does not put a working system in anyone's hands, because training, post-training, inference, and upkeep demand capital, compute, and engineering that only a few actors can provide. The paper's claim is that open source AI, as currently practiced, will not by itself democratize access to AI or provision public goods the way comparable open source efforts did in other software categories. To close that gap, open source must be nested inside public AI, meaning institutions and infrastructure that fund, host, maintain, and govern models in the public interest, with public support, public access, public accountability, and private commitments. If the paper is right, the democratizing promise of open source will not transfer to AI without deliberate public investment.

What carries the argument

The load-bearing mechanism is the activation gap between released open weights and usable AI systems. Open weights are inert artifacts; without inference, post-training, localization, tooling, interfaces, uptime, and governance, only actors with substantial capital, compute, and engineering can deploy them. The paper also formalizes the situation through the economics of impure public goods and club goods: model weights are non-rival and openly licensed, but compute and energy are private complements, so access is mediated by a toll good, much like a library whose catalog grows so large that typical users must hire a private guide to find anything. The four principles of public AI are the proposed institutional answer to that gap.

What would settle it

Measure, over several years, the cost and time required to turn released open-weight models into production-grade services using only commodity hardware and open post-training and tooling stacks; if that cost steadily falls to parity with hosted frontier assistants, the structural-gap claim is undercut.

Watch

Extended reading notes

Core claim

The paper asserts that open source AI, as currently practiced, will not by itself democratize access to AI or provision public goods as comparable open source efforts have done in other software categories. The reason is an activation gap: open weights are the finished artifact without the source, and everything that makes a model actually useful, namely inference, fine-tuning, alignment, tool integration, interfaces, and uptime, is private, expensive, and controlled by a few actors. The paper proposes to embed open source AI within a broader vision of public AI, defined by public support, public access, public accountability, and private commitments. In this vision, models are treated as public infrastructure, like libraries or highways, rather than as artifacts released into an unmanaged commons.

Load-bearing premise

The claim stands on two premises: the activation gap between released weights and usable AI systems is structural and persistent rather than temporary, and public institutions can actually build and govern AI infrastructure efficiently without being captured.

Editorial extensions

If this is right

Without public infrastructure, open-weight releases remain usable mainly by actors that already have compute and engineering capacity, so openness in AI will not produce the same public-goods outcomes as earlier open source software.
Publicly funded inference, post-training, and data capacity become necessary for open models to compete with closed frontier systems.
Shared public infrastructure gives researchers, civic technologists, and local communities the ability to inspect model internals, adapt models to local languages and needs, and conduct independent audits.
Governments can treat AI as an infrastructure asset, using public investment to complement regulation rather than relying only on constraints on private firms.
Private license changes or service shutdowns become less catastrophic when public institutions underwrite the commons.

Reading between the lines

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

If the activation gap is structural, the decisive locus of control in AI shifts from who trains models to who owns inference, post-training, and serving infrastructure; open-weight releases could become less important than open deployment infrastructure.
A concrete test of the paper's position would be an activation-gap index measuring the cost, time, and engineering effort needed to turn a released open model into a production-grade service; a persistently high index would confirm the structural claim, while a steady decline would challenge it.
The paper's logic implies that even fully open licenses are insufficient for AI; what determines whether openness yields public goods is the governance of the serving layer, not just the artifact.
If public inference funding emerges, startup value would migrate toward applications and localized adaptation, which may be why some large firms would support such public investment even as it narrows their control.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

This is a useful position paper. The novelty is not the idea that open source AI needs public infrastructure—that is in the literature it cites—but the way it structures the argument: a three-part taxonomy of resource, licensing, and governance challenges, a concrete set of public AI principles, and a good-faith engagement with five alternative views. The authors do not dodge the strongest objections. They admit that current public models like OpenEuroLLM are "unremarkable", that local inference has made "massive progress", and that public institutions can be slow and capture-prone. That honesty is the paper's best quality.

The paper's central claim is the soft spot. Section 4 asserts that open source AI "will not by itself democratize access to AI", but the evidence presented supports a weaker claim: that it probably won't, absent intervention. Section 3.3 concedes the activation gap may narrow through llama.cpp and Ollama, and Section 6.2 concedes open source may or may not be beaten by closed source. If the gap closes, the necessity argument weakens to a "it wouldn't hurt" claim. The paper does not need to retract its position—it is a position paper, after all—but it should frame the prediction as a plausible forecast rather than a settled fact.

The second issue is self-reference. Several authors are affiliated with Public AI Network, and the paper cites that network's own reports, plus its own Airbus for AI proposal, as affirmative evidence. That is not disqualifying, but it should be disclosed. The third issue is feasibility. The paper answers the "public AI will be inefficient" objection with examples like CERN and GPS, which is fair as illustration, but it is not systematic evidence. That is a limitation, not a flaw, for this genre.

Who is this for? ML researchers and policymakers thinking about open models and AI infrastructure. The paper gives them a clear framework and a set of concrete initiatives to debate. It deserves a serious referee. I would accept it with minor revisions: soften the certainty, add a disclosure note, and acknowledge more directly that the persistence of the activation gap is an empirical question.

Recommendation: engage with it, send it to review. It is not a new result, but it is a well-built argument that the community should confront.

Referee Report

3 major / 4 minor

Summary. This position paper argues that open source and open-weight AI models, despite their successes, will not on their own democratize access to AI or provision public goods as prior open source software categories have. The authors contend that models are 'inert' without expensive activation (inference, fine-tuning, tooling, hosting), that licensing and transparency are fragile, and that governance gaps allow private capture. The paper proposes 'public AI' as a necessary complement, defined by public support, public access, public accountability, and private commitments, and illustrates the idea with examples such as BigScience, LAION, OpenEuroLLM, and NDIF. It then addresses five alternative views, including the claims that the market is working, that open source will eventually win, that hosting already works, that regulation suffices, and that public AI would be inefficient or capture-prone.

Significance. If the central claim is accepted, the paper has significant implications for AI policy, research funding, and the open source community. Its strength is a coherent, well-cited diagnosis of resource, licensing, and governance failures, and it engages seriously with counterarguments rather than ignoring them. The paper also provides concrete examples of existing public initiatives. However, the central claim is a strong prediction about the future trajectory of open source AI, and the evidence offered is largely qualitative and static. The usefulness of the paper depends on whether the 'activation gap' is structural or transient; the manuscript does not yet demonstrate that. For a position paper, this is a load-bearing gap that needs either more systematic evidence or a more cautious formulation.

major comments (3)

[Section 4, Sections 3.3 and 6.2] The central assertion that open source AI 'will not by itself democratize access to AI' is stated as a fact, but the manuscript's own evidence suggests the opposite trajectory may be possible. Section 3.3 ('Expanding Gaps') acknowledges 'massive progress' in running models locally via llama.cpp and Ollama, then lists only contingent deficiencies (post-training alignment, tool integration, uptime guarantees) without arguing that these deficiencies are immune to the same open-tooling dynamics that closed earlier gaps. Section 6.2 concedes that 'Open source may or may not be beat by closed source.' Download-count comparisons in Section 6.2 are a static snapshot and do not establish that the activation gap is widening over time. Because the necessity of public AI rests on the gap being structural and persistent, the paper should provide longitudinal evidence (e.g., trends in cost-per-token, local model capabilities, availability of open post-training stacks) or weaken the claim to a conditional statement. As written, the strongest assertion of the paper is supported by an unproven persistence assumption.
[Section 6.5] The paper explicitly concedes that public institutions may be inefficient and capture-prone ('This is a valid concern'), but the rebuttal relies on historical analogies (GPS, internet, CERN, W3C) and a suggestion to redirect existing public AI spending. It does not address the specific mechanisms by which AI infrastructure would avoid capture by large incumbents or by political interests, despite the paper's own diagnosis in Sections 3.2 and 3.3 of how private actors capture open-source contributions. Since the entire prescription depends on the feasibility of capture-resistant public AI, this point needs a more concrete governance analysis, or the proposal should be framed as an experimental agenda rather than a settled solution.
[Section 5] The examples in Section 5 are presented as evidence that 'public AI is not a theoretical aspiration,' but most are instances of public funding of training compute (BigScience, LAION, OpenEuroLLM) rather than full implementations of the four principles in Section 4 (public support, public access, public accountability, private commitments). Several of the cited projects lack explicit public governance or accountability mechanisms. The paper should clarify how public funding of model training maps onto the broader definition of public AI; otherwise the examples overstate the degree to which the proposed framework is already realized.

minor comments (4)

[Abstract and Introduction] The paper uses 'open source' to include open-weight models throughout the abstract and introduction, although Section 3.2 later makes an important distinction. Consider using 'open-weight' or 'open model' consistently to avoid the known conflation the paper itself criticizes.
[Section 3.3] The claim that open model harnesses like OpenCode or OpenHands are 'several times more expensive' than direct subscriptions is unquantified. Please provide a citation or a calculation for this comparison.
[Section 6.2] The download-count comparison for LLaMA, Pythia, and OLMo would be more convincing if it specified which model variants are counted and whether the counts are aggregated across all versions of each model family. The current text appears to compare one repository per model, which may undercount or misrepresent adoption.
[Title and Section 1] The title's phrase 'win' is never defined. Please define success criteria in the introduction, for example in terms of democratized access, public-good provision, or long-term sustainability, so that the central assertion can be evaluated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central position is argued from independent evidence and does not reduce to self-citations or fitted inputs.

full rationale

This is a position paper rather than an empirical derivation, so the circularity patterns based on fitted parameters, constructor-defined quantities, or imported uniqueness theorems do not apply. The load-bearing assertion in Section 4 ('open source AI, as currently practiced, will not by itself democratize access to AI') is supported by external literature and structural arguments about pretraining cost, post-training opacity, licensing fragility, and the activation gap, not by equations that define the conclusion into the premises. The activation-gap discussion in Section 3.3 even concedes 'massive progress' in local execution via llama.cpp and Ollama, which shows the claim is not a tautology; it is a contestable empirical judgment. The paper's self-citations, including Public AI Network (2024), Tan et al. (2025), Vincent (2026), and the Public AI Inference Utility website, appear in illustrative or policy-oriented passages, but the central argument does not reduce to those citations: the same claim is independently supported by cited work such as Choksi et al. (2025), Widder et al. (2024), Bommasani et al. (2024), and the OSI opinion on LLaMA. The paper does not fit a parameter and then call it a prediction, and it does not invoke a prior uniqueness theorem from its own authors to rule out alternatives. Section 6.5 explicitly admits the feasibility concern about public institutions is 'a valid concern,' further indicating the authors do not present their prescription as a forced mathematical consequence. Therefore, while some cited sources are by the authors and vary in quality, no load-bearing derivation is circular.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

This is a position paper with no fitted parameters. The ledger records asserted premises the argument depends on: the persistence of the activation gap (Section 3.3), the persistence of resource concentration (Section 3.1), the capability of public institutions (conceded as uncertain in Section 6.5), and the economic significance of implicit data labour (supported by a co-author's non-peer-reviewed post, Vincent 2026). The central concept, public AI, is adopted from prior work including the authors' own Public AI Network (2024) report.

assumptions (4)

domain assumption Open weights without deployment infrastructure are inert, and the activation gap between released weights and usable systems is structural.
Section 3 and Section 4 ground the central claim on this premise; the paper supports it with trend examples, and Section 3.3 concedes rapid local-inference progress.
domain assumption Resource concentration in frontier AI (compute, data, post-training data) will persist.
Section 3.1 argues pretraining, post-training, and inference costs are prohibitive; whether this persists is a prediction about model economics.
domain assumption Public institutions can provision and govern AI infrastructure effectively and resist capture.
Section 6.5 admits this is a valid concern; the paper relies on analogies (GPS, internet, Hubble) rather than evidence of successful AI-specific public institutions.
domain assumption Usage and RLHF data from proprietary coding-agent harnesses constitutes a significant, uncompensated implicit data labour flow.
Section 3.3's coding-agent argument cites Vincent (2026), a non-peer-reviewed substack post by a co-author.

invented entities (1)

public AI as an institutional framework independent evidence
purpose: The paper's proposed solution: public funding, access, accountability, and private commitments to activate and sustain open models.
Not invented by this paper (adopted from the authors' Public AI Network 2024 report and Sieker et al. 2025), but central to the argument. Independent evidence comes from the concrete initiatives cited: BLOOM on Jean Zay, OpenEuroLLM, NDIF, and SEA-HELM, which have checkable outputs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of If open source is to win, it must go public." pith.science (2026). https://pith.science/paper/GGS2LJ2J

@misc{pith2026250709296,
  author       = {Pith},
  title        = {Pith review of: If open source is to win, it must go public},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGS2LJ2J}},
  note         = {Machine review of arXiv:2507.09296}
}

read the original abstract

Open source projects have made incredible progress in producing widely usable machine learning models and systems, but open source alone will face challenges in fully democratizing access to AI. Unlike previous generations of open source software, open source and open weight AI models require substantial resources to activate and maintain -- e.g., data and compute for pre-training, post-training, and deployment -- which only a few actors can currently provide. This position paper argues that open source AI must be complemented by public AI: infrastructure and institutions that ensure models are accessible, sustainable, and governed in the public interest. To achieve the full promise of AI models as prosocial public goods, we need to build public infrastructure to power and deliver open source software and models.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 4 canonical work pages

[3]

8 Position: If open source is to win, it must go public Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J

URL https://doi.org/10.5281/zeno do.15403103. 8 Position: If open source is to win, it must go public Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for con- trastive language-image learning. InProceedings of the IEEE/CVF Conference on Computer Vision and ...

doi:10.5281/zeno
[5]

URL https://aclanthology.org/2025.findin gs-acl.1267/

doi: 10.18653/v1/2025.findings-acl.1267. URL https://aclanthology.org/2025.findin gs-acl.1267/. Paszke, A., Gross, S., Massa, F., and et al. Pytorch: An imperative style, high-performance deep learning library. InAdvances in Neural Information Processing Systems 32, pp. 8024–8035, 2019. URL https://arxiv.or g/abs/1912.01703. Peng, B., Alcaide, E., Anthony...

arXiv 2025
[11]

Tan, J., Jackson, B., Berjon, R., and Coyle, D

URL https://www.bertelsmann-sti ftung.de/en/publications/publication /did/public-ai-white-paper-a-public-a lternative-to-private-ai-dominance. Tan, J., Jackson, B., Berjon, R., and Coyle, D. Airbus for AI: A global strategy for public value creation. Report, Bennett School of Public Policy, September 2025. URL https://www.repository.cam.ac.uk/han dle/1810...

work page 2025
[12]

Vaughan-Nichols, S

URL https://elpais.com/tecnologi a/2024-04-11/es-la-hora-de-aibus-por -que-europa-debe-crear-una-gran-empre sa-de-ai.html. Vaughan-Nichols, S. J. Meta abandons open-source Llama for proprietary Muse Spark, April 2026. URL https: //thenewstack.io/meta-abandons-llama -spark/. Vincent, N. The Coding Agent Data Deal, January 2026. URL https://dataleverage.sub...

arXiv 2024
[45]

Bommasani, R., Hudson, D

arXiv:2204.06745 [cs]. Bommasani, R., Hudson, D. A., Adeli, E., and et al. On the opportunities and risks of foundation models.arXiv preprint, 2021. URL https://arxiv.org/abs/ 2108.07258. Bommasani, R., Kapoor, S., Klyman, K., Longpre, S., Ramaswami, A., Zhang, D., Schaake, M., Ho, D. E., Narayanan, A., and Liang, P. Considerations for govern- ing open fo...

arXiv 2021
[48]

Public AI Inference Utility

arXiv:2305.13048 [cs]. Public AI Inference Utility. Public ai inference utility, 2025. URL https://publicai.co . Accessed: January 2026. Public AI Network. Public AI: Infrastructure for the Common Good. Technical report, August 2024. URL https://zenodo.org/doi/10.5281/zenodo. 13914560. Raymond, E. S.The Cathedral & the Bazaar: Musings on Linux and Open So...

arXiv 2025
[2016]

GPT-SW3: An Autoregressive Language Model for the Nordic Languages

URL https://fordfoundation.org/m edia/2976/roads-and-bridges-the-unsee n-labor-behind-our-digital-infrastru cture.pdf. Ekgren, A., Gyllensten, A. C., Stollenwerk, F., ¨Ohman, J., Isbister, T., Gogoulou, E., Carlsson, F., Heiman, A., Casademont, J., and Sahlgren, M. GPT-SW3: An Au- toregressive Language Model for the Nordic Languages, October 2024. URL htt...

work page Pith review arXiv 2024
[2022]

URL http://arxiv.org/abs/2204.067

work page

Show all 12 references

[2023]

URL http://arxiv.org/abs/2305.130
[2024]

doi: 10.18653/v1/2024.acl- long.841

Association for Computational Linguistics. doi: 10.18653/v1/2024.acl- long.841. URL https: //aclanthology.org/2024.acl-long.841. Hagar, N. Coding Agents for Investigative Journalism, Jan- uary 2026. URL https://generative-ai-new sroom.com/coding-agents-for-investiga tive-journ...

2024 doi
[2025]

Collective Intelligence Project

doi: 10.1109/ICSE-SEIP66354.2025.00043. Collective Intelligence Project. A Roadmap to Democratic AI. Technical report, Collective Intelligence Project, March 2024. URL https://static1.square space.com/static/631d02b2dfa9482a32d b47ec/t/65f9a1296f1a357e918f7a58/172 2968408230/C...

2025
[2026]

Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Car- lini, N., Taori, R., Dave, A., Shankar, V ., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., and Schmidt, L

Accessed Jan 29, 2026. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Car- lini, N., Taori, R., Dave, A., Shankar, V ., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., and Schmidt, L. Openclip, July 2021. URL https://doi.org/10 .5281/zenodo.5143773. Institute for ...

2026 arXiv

Pith tools

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

[1] [3]

8 Position: If open source is to win, it must go public Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J

URL https://doi.org/10.5281/zeno do.15403103. 8 Position: If open source is to win, it must go public Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., and Jitsev, J. Reproducible scaling laws for con- trastive language-image learning. InProceedings of the IEEE/CVF Conference on Computer Vision and ...

doi:10.5281/zeno

[2] [5]

URL https://aclanthology.org/2025.findin gs-acl.1267/

doi: 10.18653/v1/2025.findings-acl.1267. URL https://aclanthology.org/2025.findin gs-acl.1267/. Paszke, A., Gross, S., Massa, F., and et al. Pytorch: An imperative style, high-performance deep learning library. InAdvances in Neural Information Processing Systems 32, pp. 8024–8035, 2019. URL https://arxiv.or g/abs/1912.01703. Peng, B., Alcaide, E., Anthony...

arXiv 2025

[3] [11]

Tan, J., Jackson, B., Berjon, R., and Coyle, D

URL https://www.bertelsmann-sti ftung.de/en/publications/publication /did/public-ai-white-paper-a-public-a lternative-to-private-ai-dominance. Tan, J., Jackson, B., Berjon, R., and Coyle, D. Airbus for AI: A global strategy for public value creation. Report, Bennett School of Public Policy, September 2025. URL https://www.repository.cam.ac.uk/han dle/1810...

work page 2025

[4] [12]

Vaughan-Nichols, S

URL https://elpais.com/tecnologi a/2024-04-11/es-la-hora-de-aibus-por -que-europa-debe-crear-una-gran-empre sa-de-ai.html. Vaughan-Nichols, S. J. Meta abandons open-source Llama for proprietary Muse Spark, April 2026. URL https: //thenewstack.io/meta-abandons-llama -spark/. Vincent, N. The Coding Agent Data Deal, January 2026. URL https://dataleverage.sub...

arXiv 2024

[5] [45]

Bommasani, R., Hudson, D

arXiv:2204.06745 [cs]. Bommasani, R., Hudson, D. A., Adeli, E., and et al. On the opportunities and risks of foundation models.arXiv preprint, 2021. URL https://arxiv.org/abs/ 2108.07258. Bommasani, R., Kapoor, S., Klyman, K., Longpre, S., Ramaswami, A., Zhang, D., Schaake, M., Ho, D. E., Narayanan, A., and Liang, P. Considerations for govern- ing open fo...

arXiv 2021

[6] [48]

Public AI Inference Utility

arXiv:2305.13048 [cs]. Public AI Inference Utility. Public ai inference utility, 2025. URL https://publicai.co . Accessed: January 2026. Public AI Network. Public AI: Infrastructure for the Common Good. Technical report, August 2024. URL https://zenodo.org/doi/10.5281/zenodo. 13914560. Raymond, E. S.The Cathedral & the Bazaar: Musings on Linux and Open So...

arXiv 2025

[7] [2016]

GPT-SW3: An Autoregressive Language Model for the Nordic Languages

URL https://fordfoundation.org/m edia/2976/roads-and-bridges-the-unsee n-labor-behind-our-digital-infrastru cture.pdf. Ekgren, A., Gyllensten, A. C., Stollenwerk, F., ¨Ohman, J., Isbister, T., Gogoulou, E., Carlsson, F., Heiman, A., Casademont, J., and Sahlgren, M. GPT-SW3: An Au- toregressive Language Model for the Nordic Languages, October 2024. URL htt...

work page Pith review arXiv 2024

[8] [2022]

URL http://arxiv.org/abs/2204.067

work page