But Is It Open?The online edition
Type & theme

Make yourself comfortable

Page color
Text size
20 px
Reading width
Ask about this chapter ↗

Your settings and place stay in this browser.

Chapter 18 · Timothy O’Brien

What Should Come Next?

6 minute read16 source notesFree online edition

The profile describes a release’s gaps. Closing them takes money, people, and an institution willing to keep publishing after the first announcement.

The change I’d most like to see is an open-source model whose release an independent organization controls. Contributors and users would have a route into decisions, and no single sponsor could close the project when its strategy changed. I also think that may be unrealistic at frontier scale for a while.

The next two or three years

My near-term view is pessimistic. From this November 2026 snapshot, I don’t expect a fully open model to match the leading general-purpose systems of its own release year in the next two to three years. Matching today’s capabilities later would still matter, but it is a different claim.

Compute is one obstacle. Published estimates for final frontier training runs already reach tens or hundreds of millions of dollars. Epoch’s researchers estimate roughly 2.4-fold annual growth in frontier training costs since 2016 and project the largest runs passing $1 billion around 2027.1 These are estimates, not audited bills, and a final run excludes much of the research and failed experimentation that precedes it.

A training run, a research program, a purchased cluster, and a company’s annual infrastructure budget are different quantities. They cannot be substituted for one another to establish what an open effort must spend. Existing public infrastructure and donated compute can change the calculation. Even so, a competitive program needs sustained access to hardware, engineers, data work, and repeated experiments. I don’t know of a dependable funding model for all of that under independent release governance.

Competition is the other obstacle. Companies may benefit from publishing some artifacts while keeping the training material or most capable service private. The question is whether an institution can fund the complete release without depending on one company’s continuing preference to open it.

What foundation governance could buy

Software foundations offer useful designs, though they do not all govern projects the same way.

At the Apache Software Foundation, Project Management Committee members serve as individuals. Each release requires a formal vote, and the foundation pays attention to employer dominance. Corporate contributions are welcome, but companies do not acquire project votes by sponsoring the foundation.2 The project has an institutional home beyond its largest employer.

The PyTorch example separates several kinds of authority. Its foundation has corporate participation and a Technical Advisory Council. The PyTorch project’s technical governance is a distinct maintainer structure: membership belongs to individuals, and a lead core maintainer can resolve decisions the core group cannot settle.3 A multivendor foundation board is therefore not, by itself, proof that no company or person controls a release. To evaluate independence, read the actual release, appointment, and removal rules.

Both arrangements can accommodate substantial corporate support. Their attraction is durability and a defined path for participation. Their limits matter too: governance rules do not supply engineers or compute, and a project can remain practically dependent on one sponsor even when its formal structure is independent.

I would like to see a model organization solve both problems, sustained resources and shared authority. Naming that preference is easier than financing it. A foundation’s existence would be the beginning of the test.

More light on training

A nearer-term change is more complete disclosure of what went into a model and how it was trained.

MLCommons contributes infrastructure for that work. Croissant describes machine-learning datasets in a common metadata format so people can find and reuse them. AILuminate supplies safety benchmarks. These are different activities from ranking developer transparency; Stanford’s Foundation Model Transparency Index does the latter.45 A common format helps people ask comparable questions, but the information still needs checking against evidence.

Regulators are requiring disclosure too. The EU AI Act requires general-purpose model providers to publish a sufficiently detailed summary of training content using an AI Office template, subject to the regulation’s application and transition rules.6 A summary is not the corpus, but it creates a public account that can be compared with other releases and challenged when it is incomplete.

Five to ten years out

Further out, I’m more optimistic about open research and shared training material than about any one model.

Consider looped transformers. A conventional transformer passes inputs through its stack of layers; a recurrent-depth design reuses a block of weights, allowing more computation without proportionally increasing parameters. Universal Transformers established an early version of the idea. Geiping and colleagues trained the 3.5-billion-parameter Huginn on Oak Ridge’s Frontier supercomputer and found that additional inference-time loops improved results up to a point. Mixture-of-Recursions and Ouro explore adaptive numbers of passes.7 These results do not establish that a small open project can replace a frontier lab. They show a useful research direction that other groups can inspect and extend.

Training resources are expanding as well. Pleias’s Common Corpus contains about two trillion tokens of public-domain and permissively licensed multilingual text, with document-level provenance. Common Pile assembles about eight terabytes of public-domain and openly licensed text, primarily English. The Open License Corpus and KL3M pursue related approaches to documented, legally usable training resources.8910

Public institutions can help with access to hardware and collaboration. The U.S. National AI Research Resource pilot, launched in January 2024, connects researchers with compute, datasets, and models.11 OpenEuroLLM is working toward multilingual foundation models with open development resources in Europe.12 These projects have different scales and release commitments. They should be judged by what they make available, not grouped together as a completed alternative to closed labs.

Over five to ten years, I expect governments, nonprofits, universities, and smaller companies to build a more useful shared baseline. Clear reuse permissions and provenance can reduce uncertainty; they cannot make every copyright, privacy, or attribution question disappear. That slower accumulation seems more likely to me than an independent foundation suddenly releasing a frontier system.

The case for a closed frontier

There is a coherent alternative to my preference. Fund the corpora, standards, benchmarks, and tools, while accepting a closed commercial frontier above them. Companies earn the revenue to operate large clusters; the public supports resources that many teams can reuse. On this account, trying to keep a public model at the frontier could consume funding better spent elsewhere.

Company behavior supports part of the argument. Mark Zuckerberg explained Meta’s open releases partly through its business model: selling model access was not its core business.13 Mistral publishes Apache models alongside products with custom commercial terms.14 DeepSeek’s reported $5.576 million for V3 covered the official training run, not prior research and experiments, and the work was financed commercially.15 Publishing useful artifacts does not make the economics disappear. Some capabilities also create dual-use risks that cannot be withdrawn after a public weights release.

Where I remain doubtful is the durability of that division. A company’s willingness to release can change. Research institutions and downstream users still need tools and knowledge they can retain, study, and develop when it does. Independent governance can help preserve that work, provided the resources and decision rights are real.

Where I come down

I would invest in open research, shared training material, and open-source technology now, while continuing to work toward independently governed models. That position does not require assuming an open effort will lead every benchmark. It requires believing that people need a practical way to investigate and improve systems they depend on.

Trade secrecy and contracts can keep much of a model’s development beyond inspection.16 The test for the commons is what someone outside the original sponsor can still build, verify, and change. Funding that ability is the part of the future we can act on today.

Notes

  1. Ben Cottier et al., “The rising costs of training frontier AI models,” arXiv:2405.21015, https://arxiv.org/abs/2405.21015. The paper estimates final-run training costs and reports approximately 2.4-fold annual growth since 2016, with billion-dollar runs possible by 2027 if the trend persists. These estimates are not a measure of a company’s total development program or annual capital spending. ↩

  2. Apache Software Foundation, “Project Independence,” https://community.apache.org/projectIndependence.html; “Becoming a PMC member,” https://community.apache.org/committers/becomingapmcmember.html. ↩

  3. Soumith Chintala, “PyTorch strengthens its governance by joining the Linux Foundation,” September 12, 2022, https://pytorch.org/blog/PyTorchfoundation/; PyTorch Foundation Technical Advisory Council, https://pytorch.org/tac/; PyTorch, “Governance | Mechanics,” updated October 2, 2026, https://docs.pytorch.org/docs/2.15/community/governance.html. The project document separates technical from business governance, assigns maintainer roles to individuals rather than company seats, and describes the lead core maintainer’s authority. ↩

  4. MLCommons, “Croissant,” https://mlcommons.org/working-groups/data/croissant/; Mubashara Akhtar et al., “Croissant: A Metadata Format for ML-Ready Datasets,” arXiv:2403.19546, https://arxiv.org/abs/2403.19546; MLCommons, “AILuminate,” https://mlcommons.org/ailuminate/. ↩

  5. Stanford CRFM, Foundation Model Transparency Index, https://crfm.stanford.edu/fmti; 2025 edition paper, arXiv:2512.10169, https://arxiv.org/html/2512.10169v1. ↩

  6. Regulation (EU) 2024/1689 (AI Act), Article 53(1)(d), which requires providers of general-purpose AI models to “draw up and make publicly available a sufficiently detailed summary about the content used for training” according to an AI Office template, https://artificialintelligenceact.eu/article/53/; European Commission, “Explanatory Notice and Template for the Public Summary of Training Content for general-purpose AI models,” July 24, 2025, https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models. ↩

  7. Mostafa Dehghani et al., “Universal Transformers,” ICLR 2019, arXiv:1807.03819, https://arxiv.org/abs/1807.03819; Jonas Geiping et al., “Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach,” February 2025, arXiv:2502.05171, https://arxiv.org/abs/2502.05171 (3.5 billion parameters, trained on Oak Ridge National Laboratory’s Frontier supercomputer); Sangmin Bae et al., “Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation,” July 2025, arXiv:2507.10524, https://arxiv.org/abs/2507.10524; Rui-Jie Zhu et al., “Scaling Latent Reasoning via Looped Language Models” (Ouro), October 2025, arXiv:2510.25741, https://arxiv.org/abs/2510.25741. Overviews: Sebastian Raschka, LLM Architecture Gallery, https://sebastianraschka.com/llm-architecture-gallery/looped-depth-sharing; “What are looped transformers,” https://saulius.io/blog/what-are-looped-transformers-recurrent-depth. ↩

  8. Pleias et al., “Common Corpus,” arXiv:2506.01732, https://arxiv.org/html/2506.01732v3; AI Alliance, “Pleias releases Common Corpus,” https://thealliance.ai/blog/pleias-releases-common-corpus-open-multilingual-dataset-for-llm-training. ↩

  9. Nikhil Kandpal et al., “The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text,” June 2025, arXiv:2506.05209, https://arxiv.org/abs/2506.05209. The paper says the collection focuses primarily on English; its Comma models were trained on 1 and 2 trillion tokens drawn from it. ↩

  10. Sewon Min et al., “SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore,” arXiv:2308.04430, https://arxiv.org/abs/2308.04430, which introduces the 228-billion-token Open License Corpus; “The KL3M Data Project: Copyright-Clean Training Resources for Large Language Models,” April 2025, arXiv:2504.07854, https://arxiv.org/abs/2504.07854. ↩

  11. National Science Foundation, “National AI Research Resource (NAIRR) Pilot seeks datasets,” https://www.nsf.gov/funding/information/dcl-national-ai-research-resource-nairr-pilot-seeks-datasets (pilot launched January 2024); NSF, “National Artificial Intelligence Research Resource,” https://www.nsf.gov/focus-areas/si/nairr; pilot resources, https://nairrpilot.org. Live project counters change and are not used here as a durable measure of scale. ↩

  12. OpenEuroLLM, https://www.openeurollm.eu. ↩

  13. Mark Zuckerberg, “Open Source AI Is the Path Forward,” Meta, July 23, 2024, https://about.fb.com/news/2024/07/open-source-ai-is-the-path-forward/. The letter says “selling access to AI models isn’t our business model” and that Meta wants to avoid being locked into a competitor’s closed ecosystem. ↩

  14. Mistral AI, “Mistral 7B,” September 27, 2023, https://mistral.ai/news/announcing-mistral-7b (Apache 2.0); Mistral-Small-4-119B-2603 model card, https://huggingface.co/mistralai/Mistral-Small-4-119B-2603 (Apache 2.0); “Mistral Large 4 (Le Chonk) is now open weight,” October 2026, https://mistral.ai/news/mistral-large-4-oss (Mistral Ecosystem Open-Weight license, with a separate-license condition for commercial Model-as-a-Service offerings). ↩

  15. DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437, https://arxiv.org/abs/2412.19437. The report gives $5.576 million at an assumed $2 per H800 GPU hour and notes the figure covers “only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments.” ↩

  16. Camilla A. Hrdy, “Trade Secrecy Meets Generative AI,” 100 Chicago-Kent Law Review 317 (2025), https://scholarship.kentlaw.iit.edu/cklawreview/vol100/iss1/14. Hrdy argues that algorithms, source code, training data, and details of how a model was trained can all qualify as trade secrets, and that companies use terms of use and EULAs (confidentiality and anti-reverse-engineering clauses) to get more protection than trade secret law alone gives them. See also Hrdy, “Keeping ChatGPT a Trade Secret While Selling It Too,” 40 Berkeley Technology Law Journal (2025), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4879849. For frontier models, the main locks are trade secrecy and contract rather than patents or trademarks. ↩

From But Is It Open? by Timothy O’Brien (Discursive, 2026). © 2026 Timothy O’Brien. All rights reserved.

Chapter 186 min leftNext →

The book, chapter by chapter.

Source note

Go to the notes & sources ↓

The vocabulary.

A short guide to the book’s recurring terms. Follow each chapter for the argument and its sources.

Open weights
A release makes a trained model’s learned parameters available. The accompanying terms decide what you can do with them. Access to weights alone does not provide the training data, code or recipe. Read the chapter →
Copyleft
A way of using license conditions to preserve recipients’ freedoms when covered software is passed on. The obligations depend on the license and on what you distribute or make available. Read the chapter →
Permissive licenses
Licenses such as MIT and Apache 2.0 grant broad reuse rights with comparatively few conditions. Notice obligations still apply, and a software license does not automatically cover a release’s data or other components. Read the chapter →
Reproducibility
The ability to reconstruct and examine how an artifact was produced using the resources that are published. Running an existing checkpoint is a different achievement from rebuilding its training. Read the chapter →
Governance
Who can participate, who makes decisions, and how authority over a project or release is shared. A right to use the code does not by itself create a voice in upstream decisions. Read the chapter →
Training data
The examples used to train a model. Publishing the data, documenting its provenance and describing how it was filtered are distinct kinds of disclosure; privacy and other rights may limit what can be shared. Read the chapter →

A passage worth sharing.

Select a passage in the chapter, or share the one you’re reading.