Competition, Innovation, and Research

¹⁵⁴ Other sections in this Report also contain references to these topics.

¹⁵⁵ OpenAI’s GPT-4, for example, cost around $100 million USD to train. See Knight, W. (2023, April 17). OpenAI’s CEO Says the Age of Giant AI Models Is Already Over.

¹⁵⁶ Birchler, U., & Bütler, M. (2007). Information Economics (Routledge Advanced Texts in Economics and Finance) (1st Edition)

¹⁵⁷ Jones, R., & Mendelson, H. (2011). Information Goods vs. Industrial Goods: Cost Structure and Competition. Management Science, 57(1), 164–176.

¹⁵⁸ Nix, N., & et al. (2024, March 10). Silicon Valley is pricing academics out of AI research.

¹⁵⁹ Lee, K., & et al. (2024, March 12). Building Meta’s GenAI Infrastructure. See also, Clark, J. (2024, March 25). Import AI 366: 500bn text tokens; Facebook vs Princeton; why small government types hate the Biden EO.

¹⁶⁰ Hays, K. (2024, January 19). Zuck’s GPU Flex Will Cost Meta as Much as 18 Billion by the end of 2024.

¹⁶¹ Meta aims to have 350,000 NVIDIA H100 GPUs by the end of the year. If each one costs $20,000 (a modest estimate according to the Business Insider article above), the total cost will be $7B. This does not include Meta’s other computing resources, or the money required for datasets, human resources, or other requirements like energy.

¹⁶²  See, e.g., Vipra, J., & Korinek, A. (2023). Market concentration implications of foundation models: The Invisible Hand of ChatGPT. Brookings at 9-24 (analyzing the economics of foundation models). See also CDT Comment at 5. Cf. Economic Report of the President. p.280 (2024). The White House. (“In other cases, however, some combination of high entry costs, data availability, and network effects may drive markets toward having only a small number of players. Markets for generative AI products, which require huge amounts of data and computing power to train, may be particularly prone to this issue, with some even suggesting that such markets may naturally trend toward monopoly[. . .].” (internal citation omitted).

¹⁶³ See, Kapoor, S.et al., (2024). On the Societal Impact of Open Foundation Models. at 5. ArXiv.

¹⁶⁴ See generally Solaiman, I. (2023). The Gradient of Generative AI Release: Methods and Considerations. Hugging Face. See also Hugging Face Comment at 10-15.

¹⁶⁵ See, e.g., Alliance for Trust in AI Comment at 5 (“While available model weights may make it easier to develop advanced AI, there are still significant barriers to run and modify large or advanced models. It is not clear whether the model weights themselves provide enough information to end users to significantly change what they can do or develop themselves.”); Intel Comment at 8 (“[A]lmost all innovation in AI to-date has been due to openly available infrastructure [. . .]” (beyond just model weights to include “architecture and dataset transparency.”)). See also RAND Comment at 2 (“Whether [access to open foundation models] will be enough to maintain a competitive market for foundation model based products or services in general will depend on the price to develop and the performance of open models compared with closed models and on how the economics of fine-tuning, adapting, and serving foundation models differs in a particular business application between large and small companies.”) (internal citation omitted).

¹⁶⁶  See, e.g., Engine Comment at 3 (“Moreover, the extent of openness matters. Whether open source AI resources, for example, include detailed documentation, have publicly available model weights, or license-based restrictions can impact how useful those resources are for startups. Policymakers should be very clear-eyed about consequences for startups and innovation of adding policy-related barriers to these resources.”); Public Knowledge Comment at 11 (“Open source model weights, commercially available data warehouses, and public compute resources would enable many new model developers to use the data to develop and train new models. In addition, foundation models and APIs could also be opened, so that developers have reliable access to these resources.”) (internal citation omitted); ACLU et al. Comment at 9 (“The potential promise of ‘open’ AI is that it may allow increased competition and customization of AI models, disrupting the potential concentration developing in the advanced AI market. However, this competition will only exist if ‘open’ AI models are able to be hosted at the scale necessary for success.”) (quotation marks in original). Cf. IBM Comment at 5 (“The most obvious benefit of an open ecosystem is that it lowers the barrier to entry for competition and innovation. By making many of the technical resources necessary to develop and deploy AI more readily available, including model weights, open ecosystems enable small and large firms alike, as well as research institutions, to develop new and competitive products and services without steep, and potentially prohibitive, upfront costs.”). See also Widder, D., & et al. (2023). Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. at 7. (“Access to compute presents a significant barrier to reusability for even the most maximally ‘open’ AI systems, because of the high cost involved in both training and running inferences on large-scale AI models at scale (i.e. instrumenting them in a product or API for widespread public use).”) (quotation marks in original); Strengthening and Democratizing the U.S. Artificial Intelligence Innovation Ecosystem: An Implementation Plan for a National Artificial Intelligence Research Resource. at v. (2023). (“The [National AI Research Resource] should comprise a federated set of computational, data, testbed, and software resources from a variety of providers, along with technical support and training, to meet the needs of [its] target user base.”). Cf. Economic Report of the President. at 281. (2024). The White House. (“Similarly, freely available and portable data may encourage a competitive landscape and ensure that gains from data are widely distributed.”).

¹⁶⁷  See, e.g., ACLU et al. Comment at 9 (“Currently, the major commercial cloud computing vendors allow other AI models, including ‘open’ AI models, to be hosted on their cloud computing services. But there is no requirement for any major commercial cloud computing vendors to allow ‘open’ AI models to be hosted on their services, and the potential for self-preferencing may make the use of non-native AI models more difficult or expensive.”) (internal citation omitted) (quotation marks in original). Open models also benefit the cloud computing market, dominated by Amazon, Google, and Microsoft, which also shows anticompetitive and cumulative advantage features. See generally, e.g., Narechania, T., & Sitaraman, G. (2023). Working Paper Number 24-8. See also Paul, K. (2023, July 18). Meta opens AI model to commercial use, throwing nascent market into flux. (“Asked why Microsoft would support an offering that might degrade OpenAI’s value, a Microsoft spokesperson said giving developers choice in the types of models they use would help extend its position as the go-to cloud platform for AI work.”).

¹⁶⁸ The Open Source Definition. (2024, February 16).

¹⁶⁹ Hoffmann, M., & et al. (2024). The Value of Open Source Software (Harvard Business School Strategy Unit Working Paper No. 24-038). Harvard Business School.

¹⁷⁰ See, e.g., CDT Comment at 2-4.

¹⁷¹ Blind, K., & Schubert, T. (2023). Estimating the GDP effect of Open Source Software and its complementarities with R&D and patents: Evidence and policy implications. The Journal of Technology Transfer, 49:466–491.

¹⁷² West, J., & Gallagher, S. (2006). Challenges of open innovation: The paradox of firm investment in open-source software. 36(3), 319–331.

¹⁷³ See, e.g., Mozilla Comment at 12 (“As Widder, West, and Whittaker have argued, promoting openness in AI alone is not sufficient for creating a more competitive ecosystem. There are also risks of openness being co-opted by big industry players, and a long track record of companies drawing significant benefits from open source technology without re-investing into the communities that have developed those technologies.”), referencing Widder, D., & et al. (2023). Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. at 6. See also ACLU et al. Comment at 6 (“Further compounding the complexity around ‘open’ AI is the fact that it is not always easy to separate ‘openness’ from the business interests of large AI developers, who may benefit from open innovation on their platforms and may later withdraw commitments to openness after the benefits have reached a critical mass, knowing that smaller developers are unlikely to have the resources necessary to independently compete.”) (quotation marks in original) (internal citations omitted).

¹⁷⁴ Paul, K. (2023, July 18). Meta opens AI model to commercial use, throwing nascent market into flux.

¹⁷⁵ Yao, D. (2023, July 27). Meta to Charge for Llama 2 After All – If You’re a Hyperscaler.

¹⁷⁶ Widder, D., & et al. (2023). Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. at 13.

¹⁷⁷  See Engler, A. (2021, August 10). How open-source software shapes AI policy. (“In fact, for Google and Facebook, the open sourcing of their deep learning tools (Tensorflow and PyTorch, respectively), may have the exact opposite effect, further entrenching them in their already fortified positions. While [open source software] is often associated with community involvement and more distributed influence, Google and Facebook appear to be holding on tightly to their software. [. . .] [T] these companies are gaining influence over the AI market through OSS, while the OSS AI tools not backed by companies, such as Caffe and Theano, seem to be losing significance in both AI research and industry. By making their tools the most common in industry and academia, Google and Facebook benefit from the public research conducted with those tools, and, further, they manifest a pipeline of data scientists and machine learning engineers trained in their systems.”) (internal hyperlink omitted).

¹⁷⁸ Staff in the Federal Trade Commission Office of Technology. (2024, July 10). On Open-Weights Foundation Models.

¹⁷⁹ Scharre, P. (2024, March 13). Future-Proofing Frontier AI Regulation.

¹⁸⁰ See, Melton, M. (2024, February). Generative AI startup Latimer, known as the “BlackGPT”, will launch a new bias detection tool and API. Business Insider.

¹⁸¹  See, Rishi Bommasani et al. Comment at 5 (“Open foundation models promote competition in some layers of the AI stack. Given the significant capital costs of developing foundation models, broad access to model weights and greater customizability can also reduce market concentration by enabling greater competition in downstream markets. However, open foundation models are unlikely to reduce market concentration in the highly concentrated upstream markets of computing and specialized hardware providers.”) (internal citation omitted); RAND Comment at 2 (“Whether [access to open foundation models] will be enough to maintain a competitive market for foundation model based products or services in general will depend on the price to develop and the performance of open models compared with closed models and on how the economics of fine-tuning, adapting, and serving foundation models differs in a particular business application between large and small companies.”) (internal citation omitted). See also, Hugging Face Comment at 9 (“Additionally, open-weight models have been customized and adapted to run on a greater variety of infrastructure, including individual GPUs and even CPUs, reducing points of market concentration with cloud providers and reducing costs for procurement.”) (referencing Ggerganov / llama. cpp. (n.d.) and Hood, S. (2023, December 14). llamafile: Bringing LLMs to the people, and to your own computer); CDT Comment at 9 (“Importantly, the innovation in developing smaller and more powerful models, often based directly on much larger models, is not just important in terms of competition and innovation. It is also important because some models such as Mistral 7B are now small enough to run locally on an end-user’s laptop or even a phone, mitigating the need for a cloud-based provider at all.”) (internal citation omitted).

¹⁸²  See, e.g., RAND Comment at at 2 (“Open foundation models may reduce market concentration. When smaller actors can access open foundation models, they can avoid the large expense of developing their own models and can therefore compete with large tech companies in adapting the foundation model to a particular business context.”) (internal citation omitted), Engine Comment at 3 (“Openness in AI helps alleviate costs associated with the expensive parts of building models, leaving startups to focus their limited resources on their core and differentiating innovation.”); The Abundance Institute Comment at 4 (“In particular, the high costs of compiling data and purchasing compute to train foundational models are a significant barrier to model training. Sharing model weights eliminates this cost barrier, broadening access and enabling users that would otherwise simply be priced out of building their own AI stack.”).

¹⁸³  See, e.g., a16z Comment at 11 (“Open Models increase competition in the development and improvement of foundation models because they do not restrict the use of AI to gatekeeper companies with the most market power or resources. This accessibility increases the prospect of competition and allows for participation by developers who may otherwise have been boxed out of working with AI due to their lacking the requisite access or resources that are necessary components to working within a closed ecosystem.”); CSET Comment at 4 (“Downloadable weights [. . .] may reduce the concentration of power in the AI industry because the original developers do not control access to the models.”); Intel Comment at 9 (“Open model weights also spur more startups and innovations, enabling startups to quickly prototype without access to immense capital, fostering a more competitive landscape.”). Cf. Public Knowledge Comment at 10 (“Dominant companies often utilize gatekeeper power to further their own market power and cut off new entrants from the chance to compete. Open technologies may serve to counteract this exclusionary conduct and lower barriers to entry for innovative, up-and-coming rivals. Historically, we’ve already seen how open access to technology patents had competitive benefits, leading to a wellspring of innovative products.”).

¹⁸⁴ See, e.g., Mozilla Comment at 12 (“The increased availability of ‘open’ alternatives in the AI market can support competition by reducing switching costs as relying on specific proprietary model APIs or platform ecosystems (like those offered by the leading cloud service providers) can create lock-in effects for customers, both in the private and public sector.”) (quotation marks in original). Open models allow companies to switch seamlessly between baseline models without added costs. Proprietary models introduce the threat of “lock-in effects,” where a company has built a product around a certain API or provider and then cannot transfer to a new model without rebuilding their entire product. Many technology corporations have established verticals with certain cloud service providers and data collection infrastructures, and a company cannot easily exit this vertical. Conversely, when companies build on open models, they have access to those model weights forever and can switch between cloud providers and other vendors with ease.

¹⁸⁵ See Kapoor, S. (2024). On the Societal Impact of Open Foundation Models. at 5. ArXiv.

¹⁸⁶ United Nations General Assembly, Seizing the opportunities of safe, secure and trustworthy artificial intelligence systems for sustainable development. March 11, 2024. A/78/L.49.

¹⁸⁷  See ACLU et al. Comment at 10 (“As seen in other technological contexts, diffusing market concentration, especially over gateway or bottleneck facilities, can increase the diversity of voices, including for marginalized communities.”) (internal citation omitted); GitHub Comment at 10 (“An expanded developer base, particularly outside of a small set of companies located in a few major tech hubs, supports diversity of identity and perspective in the ecosystem.”).

¹⁸⁸  See CDT Comment at 7 (“[Open Foundation Models] are already driving innovation across the ecosystem as tens or hundreds of thousands of businesses begin adapting model capabilities to their own use cases and customer needs in a wide variety of contexts.”); Phase 2 at 3 (“Making foundation model weights widely available lowers barriers to entry and enables a broader range of companies to develop AI applications. This is particularly beneficial for startups and small businesses that lack the resources to develop foundation models from scratch. Open models level the playing field and ensure the economic gains from AI are widely distributed. We expect this to drive competition and innovation in sectors like healthcare, education, and marketing as more players are able to leverage AI to build groundbreaking products and services.”).

¹⁸⁹ Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. ArXiv.

¹⁹⁰  See, e.g., CTA Comment at 5 “Because [open weight models] have lower barriers to entry (e.g., cost, expertise), they are more accessible to the general public. Leveraging input and feedback from the broader AI community of researchers and users can help identify and mitigate bugs, biases, and safety issues that may otherwise go unnoticed, ultimately leading to better performing and safer AI products. This lower barrier to entry can help to drive AI research and development by academics or other subject matter experts, enabling communities with bespoke datasets and unique needs to form around specific platforms or industry sectors.” (citing Elizabeth Seger, et al. (Sept. 29, 2023). Open-Sourcing Highly Capable Foundation Models; Ly, J. (2024, March 12). Open Foundation Models: Implications of Contemporary Artificial Intelligence. Center for Security and Emerging Technology.

¹⁹¹ There may also be ways to achieve similar benefits to research and development through methods other than making model weights widely available. See, e.g., RAND Comment at 4 (“Structured access is an alternative approach that can provide users with many of the benefits of making foundation model weights widely accessible while reducing some of the risks.”); Anthony Barrett Comment at 15 (“[B]enefits of broad independent evaluation for improving the safety, security, and trustworthiness of AI are not necessarily best supported by making model weights widely available. Those benefits can also be achieved by facilitating safe and protected independent researcher access.”).

¹⁹² See, e.g., Hugging Face Comment at 9 (“Robust innovation on both performance and safety questions requires scientific rigor and scrutiny, which is enabled by openness and external reproducibility. Supporting that research requires sharing models to validate findings and lower the barrier to entry for participation given the growing resource gap between researchers in different institutions.”) (internal citations omitted); Center for Democracy & Technology, & et al. (March, 25, 2024). RE: Openness and Transparency in AI Provide Significant Benefits for Society at 2. (“Open models also help accelerate scientific research because they can be less expensive, easier to fine-tune, and supportive of reproducible research.”). See Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. at 19. ArXiv. (listing examples of research done using open foundation models).

¹⁹³  See, e.g., RAND Comment at 3 (“Making foundation model weights accessible helps uncover vulnerabilities, biases, and potentially dangerous capabilities. With a wider set of eyes examining these models, there is a higher likelihood of identifying and addressing issues that might have been overlooked by the original developers, as is the case with open-source software broadly. This scrutiny is useful for developing AI systems that are secure, fair, and aligned with societal values. The detection and mitigation of biases in AI models, for instance, are critical steps toward ensuring that AI technologies do not perpetuate or exacerbate social inequalities.”); IBM Comment at 6 (‘In some contexts, AI safety can also depend on the ability for diverse stakeholders to scrutinize and evaluate models to identify any vulnerabilities, identify undesirable behaviors, and ensure they are functioning properly. However, without ‘deep access,’ which includes access to model weights, these evaluations will be severely limited in their effectiveness.”) (internal citation omitted); CDT Comment at 10-14 (explaining the “black-box” methods of auditing for closed foundation models versus “white-box” methods of auditing for open foundation models). Cf. Engler, A. (2021, August 10). How open-source software shapes AI policy. (“Similarly, open-source AI tools can enable the broader and better use of ethical AI. Open-source tools like OSS like IBM’s AI Fairness 360, Microsoft’s Fairlearn, and the University of Chicago’s Aequitas ease technical barriers to detecting and mitigating AI bias. There are also opensource tools for interpretable and explainable AI, such as IBM’s AI Explainability 360 or Chris Molnar’s interpretable machine learning tool and book, which make it easier for data scientists to interrogate the inner workings of their models.”).

¹⁹⁴ See GitHub Comment at 9 (“To-date, researchers have credited [open source and widely available AI] models with supporting work to [. . .] advance the efficiency of AI models enabling them to use less resources and run on more accessible hardware.”) (citing Tim Dettmers, et al., “QLoRA: Efficient Finetuning of Quantized LLMs,” ArXiv, May 23, 2023, and its associated repository at Artidoro / qlora. (n.d.)). 
Quantifying Memorization Across Neural Language Models).

¹⁹⁵  See, e.g., Intel Comment at 8-9 (“Open model weights are likely going to aid researchers to find impactful and beneficial use cases of AI that will be overlooked by narrow and immediate commercial interests of proprietary model vendors. An example of this is applying leading-edge AI principles to open scientific problems.”); OTI Comment at 16 (“One of the key benefits of a healthy ecosystem characterized by a prevalence of open models is that many people can learn how the technology works. This enables technologists and community leaders to partner in ways that are tailored to address specific community needs and implement community-driven solutions. Relatedly, open-source projects can also be used to fill technological gaps that aren’t being met in the private sector.”) (internal citations omitted).

¹⁹⁶ See MLCommons Comment at 3 (“Models with open weights have played a central role in developing widely trusted benchmarks that have been used to evaluate and measure AI models, and in doing so have helped drive progress in AI. GLUE, BigBench, Harness, HELM and openCLIP Benchmark are all examples of widely used benchmarks that have helped researchers and developers measure progress in the development of AI models.”) (internal citations omitted).

¹⁹⁷ See EleutherAI Comment at 24 (“Even for researchers in industrial labs such as Google, open models can enable research on model safety that would not otherwise be possible: in an earlier revision of Quantifying Memorization Across Neural Language Models, Carlini et al. state that their research on harmful memorization in language models “would not have been possible without EleutherAI’s complete public release of The Pile dataset and their GPTNeo family of models.”) (internal citations omitted). See also Zou, A. (2023). Universal and Transferable Adversarial Attacks on Aligned Language Models.

¹⁹⁸ See Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. at 4. ArXiv. (referencing research that See Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. at 4. ArXiv. (referencing research that requires no safety filters). See also CDT Comment at 9 (“Furthermore, [open foundation models] enable a variety of AI research not enabled by closed foundation models, including research around AI interpretability methods, security, model training and inference efficiency, and the public development of robust watermarking techniques.”) (listing examples) (internal citations omitted). See also Kapoor, S., & Narayanan, A. (2023, March 22). OpenAI’s policies hinder reproducible research on language models. To be clear, the benefits in this space are not a zero-sum game. One may need access to both open weight models and “closed” foundation models. See, e.g. MLCommons Comment at 3 (describing the limitations of relying solely on models with open weights or models with closed weights to evaluate models and urging simultaneous use).

¹⁹⁹ See, e.g., CSET Comment at 11 (“Most current [Biological Design Tools] are open models developed by academic labs. The life sciences community places a high value on scientific transparency and openness, and tends to favor open sharing of resources [. . .] Shifting away from the open sharing of model weights would also require additional resources, as many academic researchers do not have the time, funding, and infrastructure to set up and maintain an API.”) (internal hyperlink omitted).

²⁰⁰ See, e.g., Miller, K. (2024, March 12). Open Foundation Models: Implications of Contemporary Artificial Intelligence. (“Actors may opt to use open models instead of paying for access to closed models, which may reduce the revenue of developers and disincentivize investments in capital-intensive R&D.”).

²⁰¹ See, e.g., CSET Comment at 10 (“More research is needed to determine what types of research are enabled by open weights, and how that may allow more entrants into the market. Many prospective entrants may lack resources, and it is unclear the extent to which resource constraints may limit the benefits of open models to R&D. Actors may lack the data to fine-tune open models, or lack the compute to use or experiment with open models rigorously and at scale (although resources provided through the NAIRR pilot may help alleviate resource constraints).”) (internal hyperlinks omitted).

²⁰² 1 See, e.g., CSET Comment at 10 (“More research is needed to determine what types of research are enabled by open weights, and how that may allow more entrants into the market. Many prospective entrants may lack resources, and it is unclear the extent to which resource constraints may limit the benefits of open models to R&D. Actors may lack the data to fine-tune open models, or lack the compute to use or experiment with open models rigorously and at scale (although resources provided through the NAIRR pilot may help alleviate resource constraints).”) (internal hyperlinks omitted).

²⁰³ Other sections in this Report also contain references to these topics.

²⁰⁴ See Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. at 4. ArXiv. These researchers also note that “new research directions such as merging models might allow open foundation model developers to reap some of these benefits (akin to open source software).” (internal citation omitted).

²⁰⁵ See generally id. Kleinberg and Raghavan highlight several concerns with algorithmic monoculture: 1) risk of severe harm in monoculture systems due to unexpected shocks, and 2) decrease in decision-making quality across the board. See also, e.g., a16z Comment at 20 (“Algorithmic monocultures resulting from reliance on a few Closed Models can create resilience problems and generate systemic risk. If those models are compromised, the impacts could be widespread and pervasive.”). Cf. Vipra, J., & Korinek, A. (2023). Market concentration implications of foundation models: The Invisible Hand of ChatGPT. at 25. Brookings. (“Foundation models will likely be integrated into production and delivery processes for goods and services across many sectors of the economy. We can imagine one foundation model in its fine-tuned versions powering decision-making processes in search, market research, customer service, advertising, design, manufacturing, and many more. If foundation models are integrated into a growing number of economic activities, then widespread, cross-industrial applications mean that any errors, vulnerabilities, or failures in a foundation model can threaten a significant amount of economic activity, producing the risk of systemic economic effects.”).

²⁰⁶ See, e.g., CDT Comment at 5 (“[W]hen many different decisionmakers and service providers rely on the same systems, there can be a trend toward ‘algorithmic monoculture’ whereby systemic exclusion of individuals or groups in AI-driven decisionmaking occurs across the ecosystem”) (citing Rishi Bommasani et al., “Picking on the Same Person: Does Algorithmic Monoculture Lead to Outcome Homogenization?,” ArXiv, November 25, 2022. [perma.cc/F7JB-3AK3]).

²⁰⁷ See, e.g., Mozilla Comment at 12 (“Additionally, concentrating cutting-edge research in ever-fewer research labs may also exacerbate phenomena like algorithmic monoculture and entrench (or increase the ‘stickiness’ of) existing technological paradigms at the expense of pursuing new research directions) (quotation marks in original), citing Fishman, N., & Hancox-Li, L. (2022). Should attention be all we need? The epistemic and ethical implications of unification in machine learning. and Hooker, S. (2020). The Hardware Lottery. ArXiv. See also Fishman, N., & Hancox-Li, L. (2022). Should attention be all we need? The epistemic and ethical implications of unification in machine learning. at 14.