Latest / USA Daily News / AI Rivalry Forces a New U.S. Strategy as Bessent Pushes Open-Source Models to Compete With China

Transcript

AI Rivalry Forces a New U.S. Strategy as Bessent Pushes Open-Source Models to Compete With China · USA Daily News

• 0:00
  Speaker 1: The artificial intelligence race between the United States and China
• 0:04
  Speaker 1: is moving into a new phase, and Treasury Secretary Scott
• 0:08
  Speaker 1: Besant is arguing that America cannot afford to rely exclusively
• 0:12
  Speaker 1: on a small number of powerful private laboratories if it
• 0:16
  Speaker 1: wants to maintain its technological advantage. Speaking before the House
• 0:20
  Speaker 1: Financial Services Committee, Besant focused attention on a part of
• 0:24
  Speaker 1: the AI industry that has become increasingly important in the
• 0:28
  Speaker 1: global competition— the open source and open models that can be examined,
• 0:33
  Speaker 1: adapted and deployed by a much wider community of developers.
• 0:37
  Speaker 1: His argument reflects a growing debate in Washington and Silicon
• 0:41
  Speaker 1: Valley over how the United States should balance innovation, security,
• 0:46
  Speaker 1: commercial incentives, and national competitiveness. Rather than viewing AI only
• 0:51
  Speaker 1: as a race between a handful of corporations developing proprietary systems,
• 0:56
  Speaker 1: Besson presented broader access to advanced models as a potential
• 1:00
  Speaker 1: strategic advantage. The issue has become especially significant because Chinese
• 1:05
  Speaker 1: companies have increasingly developed capable AI systems that are available
• 1:09
  Speaker 1: more openly and can be adapted at comparatively lower cost.
• 1:14
  Speaker 1: According to the Fox Business Report, Besson specifically pointed to
• 1:18
  Speaker 1: Anthropic's mythos model as an important development in the progression
• 1:22
  Speaker 1: toward artificial general intelligence, while arguing that the United States
• 1:27
  Speaker 1: needs more open-source.
• 1:28
  Speaker 2: Models of its own.
• 1:30
  Speaker 1: His comments came as lawmakers were examining economic and international issues,
• 1:35
  Speaker 1: but the discussion quickly expanded into one of the most
• 1:38
  Speaker 1: consequential technology debates in the world. The underlying question is
• 1:43
  Speaker 1: simple to state but extraordinarily complicated to answer. Should America
• 1:47
  Speaker 1: encourage wider access to powerful AI systems in order to
• 1:51
  Speaker 1: accelerate domestic innovation— even when that openness can create security,
• 1:57
  Speaker 1: intellectual property, and misuse concerns? Besson's answer leaned toward greater
• 2:02
  Speaker 1: development of American open models, accompanied by opposition to excessive
• 2:06
  Speaker 1: regulation that could leave major laboratories with too much control
• 2:11
  Speaker 1: over the direction of the technology. he warned about what
• 2:14
  Speaker 1: he described as regulatory capture, arguing that rules designed in
• 2:19
  Speaker 1: ways that favor large incumbent companies could slow innovation and
• 2:24
  Speaker 1: make it harder for newer American developers to compete. That
• 2:28
  Speaker 1: concern is significant because AI development has already become concentrated
• 2:32
  Speaker 1: around companies with enormous computing resources, specialized talent, vast datasets,
• 2:39
  Speaker 1: and access to advanced chips. smaller laboratories, universities, startups, and
• 2:44
  Speaker 1: independent developers often cannot reproduce that infrastructure from scratch. Open
• 2:50
  Speaker 1: models can change the equation by giving those groups a
• 2:53
  Speaker 1: starting point they can inspect, modify, and build upon. The
• 2:57
  Speaker 1: strategic value is therefore not limited to the models themselves.
• 3:02
  Speaker 1: It extends to an ecosystem of businesses, researchers, software developers,
• 3:06
  Speaker 1: and institutions that can innovate around those models. In theory,
• 3:11
  Speaker 1: A broader open model ecosystem could make the United States
• 3:15
  Speaker 1: less dependent on a handful of companies and could allow
• 3:18
  Speaker 1: American technology to spread more quickly into industries ranging from
• 3:23
  Speaker 1: finance and manufacturing to education, logistics, cybersecurity, and government services.
• 3:30
  Speaker 1: But openness carries a second dimension that makes the issue
• 3:33
  Speaker 1: far more complicated. The same accessibility that allows legitimate innovators
• 3:38
  Speaker 1: to experiment can also make advanced capabilities available to actors
• 3:43
  Speaker 1: who may use them irresponsibly. That is where the debate
• 3:46
  Speaker 1: over safety enters the picture. Some policymakers and technology leaders
• 3:51
  Speaker 1: argue that highly capable systems require stronger safeguards before being
• 3:56
  Speaker 1: broadly released. while others maintain that restricting open development could
• 4:01
  Speaker 1: unintentionally hand an advantage to foreign competitors or cement the
• 4:05
  Speaker 1: dominance of the largest existing companies. Besson's remarks place him
• 4:10
  Speaker 1: firmly in the part of that debate that sees broader
• 4:13
  Speaker 1: domestic development as an important tool for maintaining competitiveness. His
• 4:18
  Speaker 1: argument also comes against a backdrop of rising concern that
• 4:21
  Speaker 1: China is closing technological gaps in AI and is using
• 4:25
  Speaker 1: open approaches to increase adoption. the geopolitical dimension is impossible
• 4:30
  Speaker 1: to ignore. Artificial intelligence is increasingly viewed not only as
• 4:35
  Speaker 1: a commercial technology, but also as a strategic capability linked
• 4:40
  Speaker 1: to economic productivity, national security, military applications, scientific research, and
• 4:47
  Speaker 1: technological influence. For Washington, falling behind in AI could therefore
• 4:53
  Speaker 1: carry consequences far beyond the technology sector. If American companies
• 4:58
  Speaker 1: remain at the frontier, but Chinese models become more widely available,
• 5:03
  Speaker 1: cheaper and easier for organizations around the world to adopt,
• 5:07
  Speaker 1: the United States could face a different kind of competitive challenge.
• 5:11
  Speaker 1: It would not necessarily be a question of which country
• 5:14
  Speaker 1: has the single most capable model. The debate over open
• 5:18
  Speaker 1: source AI becomes even more consequential when the technology itself
• 5:22
  Speaker 1: is examined closely. In traditional software development, Open-source systems are
• 5:28
  Speaker 1: generally associated with code that can be inspected, modified, and
• 5:32
  Speaker 1: redistributed under specified licensing conditions. In AI, the terminology can
• 5:38
  Speaker 1: become more complicated because a model may expose some combination
• 5:42
  Speaker 1: of source code, model weights, training information, documentation, or tools
• 5:47
  Speaker 1: without making every part of the development process public. The
• 5:51
  Speaker 1: Fox Business Report describes open-source models as systems that provide
• 5:55
  Speaker 1: the public with meaningful access to components such as code
• 5:59
  Speaker 1: or training parameters, in contrast to closed systems whose developers
• 6:03
  Speaker 1: retain control over access and use. That distinction has become
• 6:08
  Speaker 1: central to the current AI debate because access determines who
• 6:11
  Speaker 1: can experiment, who can customize a system, and how quickly
• 6:15
  Speaker 1: new applications can emerge. A closed model may be extraordinarily powerful,
• 6:21
  Speaker 1: but developers typically depend on the company operating the system.
• 6:25
  Speaker 1: They may access it through an application programming interface, a
• 6:29
  Speaker 1: hosted service, or a controlled platform. An open model can
• 6:33
  Speaker 1: give developers a different relationship with the technology because they
• 6:37
  Speaker 1: can sometimes download, modify, fine-tune, and.
• 6:41
  Speaker 2: Deploy the system themselves.
• 6:43
  Speaker 1: That flexibility can be valuable to companies that need specialized performance,
• 6:48
  Speaker 1: researchers studying model behavior and organizations that cannot send sensitive
• 6:53
  Speaker 1: information to an external provider. From a national competitiveness perspective,
• 6:58
  Speaker 1: those capabilities can create a broad base of innovation rather
• 7:02
  Speaker 1: than concentrating experimentation inside a few corporations. Besson's argument fits
• 7:08
  Speaker 1: into that theory. He is not simply advocating openness. because
• 7:12
  Speaker 1: it is philosophically attractive. He is presenting it as a
• 7:16
  Speaker 1: way for the United States to expand the number of
• 7:18
  Speaker 1: people and organizations capable of building.
• 7:21
  Speaker 2: On advanced AI.
• 7:22
  Speaker 1: The more developers that can work with American models, the
• 7:26
  Speaker 1: argument goes, the more applications may emerge, the more domestic
• 7:30
  Speaker 1: companies may form, and the more difficult it becomes for
• 7:34
  Speaker 1: foreign competitors to capture the surrounding ecosystem. That brings the
• 7:38
  Speaker 1: conversation directly to China. Besson criticized what he described as
• 7:43
  Speaker 1: Chinese AI companies using a technique called distillation to learn
• 7:47
  Speaker 1: from American closed models. Distillation is a legitimate technical concept
• 7:52
  Speaker 1: in machine learning. A smaller model can be trained to
• 7:56
  Speaker 1: imitate or approximate the outputs of a larger model, allowing
• 8:00
  Speaker 1: knowledge to be transferred into a system that is cheaper
• 8:03
  Speaker 1: or more efficient to run. The controversy arises when the
• 8:07
  Speaker 1: source model is proprietary and the transfer is alleged to
• 8:10
  Speaker 1: involve unauthorized extraction of value. American technology companies and officials
• 8:16
  Speaker 1: have accused Chinese firms of using this process to reproduce
• 8:20
  Speaker 1: capabilities from U.S. systems, while Chinese representatives have disputed broader
• 8:25
  Speaker 1: American accusations of technological wrongdoing. The distinction between legal model
• 8:31
  Speaker 1: training practices and unlawful appropriation therefore matters enormously. Open models
• 8:37
  Speaker 1: may provide one way to change the competitive environment because
• 8:40
  Speaker 1: they can make certain American capabilities available legally and deliberately
• 8:45
  Speaker 1: rather than forcing developers to depend on closed systems. That
• 8:49
  Speaker 1: could reduce the incentive to reproduce proprietary models through questionable methods,
• 8:55
  Speaker 1: although it cannot eliminate the problem by itself. Open access
• 8:59
  Speaker 1: also creates a different challenge because once a model is
• 9:02
  Speaker 1: broadly distributed, the developer has less control over what other
• 9:06
  Speaker 1: users do with it. That raises questions about cybersecurity, privacy,
• 9:11
  Speaker 1: dual-use capabilities, and the possibility of malicious adaptation. A model
• 9:16
  Speaker 1: that helps a legitimate business automate customer service could potentially
• 9:21
  Speaker 1: be adapted for harmful purposes under different circumstances. Policymakers therefore
• 9:26
  Speaker 1: face a balancing act. Excessive restrictions may slow legitimate development,
• 9:32
  Speaker 1: reduce competition, and strengthen incumbent firms, but insufficient safeguards may
• 9:37
  Speaker 1: create risks that are difficult to reverse once models are
• 9:40
  Speaker 1: widely distributed. The debate becomes even more complicated when national
• 9:45
  Speaker 1: security enters the picture. A country cannot easily tell its
• 9:49
  Speaker 1: researchers and companies to move more slowly while assuming that
• 9:53
  Speaker 1: foreign competitors will make the same choice. If Chinese developers
• 9:58
  Speaker 1: continue improving capable models while American companies face increasingly restrictive rules,
• 10:04
  Speaker 1: the relative balance could change. That concern is one reason
• 10:08
  Speaker 1: the issue has attracted the attention of Treasury officials, technology executives,
• 10:13
  Speaker 1: and lawmakers whose normal responsibilities do not always include detailed
• 10:18
  Speaker 1: AI policy. Artificial intelligence now intersects with trade, finance, national security,
• 10:25
  Speaker 1: intellectual property, and industrial policy. Behind the argument for more
• 10:30
  Speaker 1: open-source AI is another conflict that could shape the industry
• 10:34
  Speaker 1: for years. The question of who should write the rules
• 10:37
  Speaker 1: governing advanced artificial intelligence and how those rules should affect
• 10:41
  Speaker 1: companies at different stages of development. Besant specifically warned lawmakers
• 10:46
  Speaker 1: about the possibility of regulatory capture of a term used
• 10:50
  Speaker 1: to describe circumstances in which an industry or dominant companies
• 10:54
  Speaker 1: can exert excessive influence over the roles intended to govern them.
• 10:58
  Speaker 1: In the AI market, that concern carries unusual weight because
• 11:02
  Speaker 1: the largest laboratories already possess advantages that smaller competitors cannot
• 11:07
  Speaker 1: easily reproduce. They have immense computing infrastructure, highly specialized employees,
• 11:13
  Speaker 1: access to huge quantities of data, strong financial backing, and
• 11:18
  Speaker 1: established relationships with cloud providers and chip manufacturers. If regulations
• 11:24
  Speaker 1: become expensive or technically demanding, the largest companies may be
• 11:29
  Speaker 1: able to absorb those costs while startups and independent developers cannot.
• 11:34
  Speaker 1: A rule intended to improve safety could therefore have an
• 11:37
  Speaker 1: unintended market effect by making competition harder. Besson's warning suggests
• 11:42
  Speaker 1: that policymakers should examine not only what a proposed AI
• 11:46
  Speaker 1: rule is supposed to accomplish, but also which companies can
• 11:49
  Speaker 1: realistically comply with it. The distinction becomes especially important when
• 11:54
  Speaker 1: considering pre-release testing requirements, security reviews, and other obligations imposed
• 12:00
  Speaker 1: on models before they can be distributed. According to the
• 12:04
  Speaker 1: Fox Business Report, The Trump administration's voluntary AI framework released
• 12:09
  Speaker 1: in August took a different approach by exempting open source
• 12:13
  Speaker 1: and open-weight models from government pre-release security reviews that would
• 12:17
  Speaker 1: instead focus on proprietary closed systems. Supporters of this philosophy
• 12:22
  Speaker 1: could argue that open development creates more transparency and wider
• 12:27
  Speaker 1: scrutiny because outside researchers can inspect and test models instead
• 12:31
  Speaker 1: of relying entirely on the developer's own assurances. Critics could
• 12:36
  Speaker 1: respond that openness does not automatically make a powerful model
• 12:40
  Speaker 1: safe and that public availability could increase the number of
• 12:43
  Speaker 1: actors capable of exploiting vulnerabilities or misusing advanced capabilities. Both
• 12:49
  Speaker 1: concerns have to be considered because AI systems do not
• 12:53
  Speaker 1: fit neatly into traditional categories of software or physical products.
• 12:57
  Speaker 1: A model can be copied, modified, and distributed at extraordinary speed.
• 13:02
  Speaker 1: meaning a single release can influence thousands of downstream applications.
• 13:07
  Speaker 1: The regulatory challenge is therefore not simply deciding whether a
• 13:11
  Speaker 1: model is safe when it leaves a company's servers. It
• 13:14
  Speaker 1: involves understanding how that model might evolve after release, who
• 13:18
  Speaker 1: might modify it, and what protections will exist around derivative systems.
• 13:23
  Speaker 1: This is one reason the open-source debate has become such
• 13:26
  Speaker 1: a powerful proxy for larger questions about innovation and control.
• 13:31
  Speaker 1: Supporters of open models often argue that decentralized development reduces
• 13:36
  Speaker 1: dependence on a few technology giants and gives researchers, startups,
• 13:41
  Speaker 1: and institutions more freedom. Advocates of stricter controls often emphasize
• 13:46
  Speaker 1: the possibility that powerful capabilities could be misused or that
• 13:50
  Speaker 1: public access might make oversight more difficult. Besson's comments do
• 13:55
  Speaker 1: not eliminate those concerns. but they emphasize a particular economic argument.
• 14:01
  Speaker 1: America should not construct rules that unintentionally cause its own
• 14:05
  Speaker 1: AI ecosystem to become less innovative while competing countries continue
• 14:10
  Speaker 1: developing rapidly. That concern has been echoed by technology companies.
• 14:15
  Speaker 1: The Fox Business Report notes that a group including NVIDIA, Microsoft, Meta,
• 14:21
  Speaker 1: Dell Technologies, Palantir, Hugging Face, Mozilla, and Mistral signed a
• 14:26
  Speaker 1: joint letter in July arguing against restrictions they believed could
• 14:29
  Speaker 1: stifle open model development. The company supported addressing illegal attempts
• 14:34
  Speaker 1: to extract value from closed models, but argued that those
• 14:38
  Speaker 1: concerns should be handled through targeted legal and commercial measures
• 14:42
  Speaker 1: rather than broad restrictions on open model techniques. The argument
• 14:47
  Speaker 1: reveals a subtle but important distinction. Supporting open models does
• 14:52
  Speaker 1: not necessarily mean opposing all regulation. Many companies are effectively
• 14:57
  Speaker 1: arguing for rules that target specific harmful conduct without imposing
• 15:01
  Speaker 1: sweeping obligations that make legitimate development harder. That distinction could
• 15:06
  Speaker 1: become central to future policy discussions. Regulators may be asked
• 15:11
  Speaker 1: to distinguish between a developer releasing an American model under
• 15:15
  Speaker 1: transparent conditions and an entity unlawfully extracting capabilities from another
• 15:20
  Speaker 1: company's proprietary system. They may also have to determine how
• 15:25
  Speaker 1: responsibilities should be divided when an open model is modified
• 15:28
  Speaker 1: downstream or deployed by an independent organization. China's position in
• 15:33
  Speaker 1: the AI race is one of the reasons the open
• 15:36
  Speaker 1: model argument has become so urgent. The competition between Washington
• 15:41
  Speaker 1: and Beijing increasingly involves not only the performance of individual
• 15:46
  Speaker 1: AI models, but also the methods used to build an
• 15:49
  Speaker 1: ecosystem around them. Chinese technology companies have invested heavily in
• 15:54
  Speaker 1: artificial intelligence, and some Chinese models have gained attention for
• 15:58
  Speaker 1: being capable while also being offered with relatively open licensing
• 16:03
  Speaker 1: or access arrangements. That creates a different competitive environment from
• 16:07
  Speaker 1: one in which only a few proprietary American systems dominate
• 16:11
  Speaker 1: the market. Besson's concern is that the United States could
• 16:15
  Speaker 1: inadvertently create a weakness if it concentrates too heavily on
• 16:19
  Speaker 1: closed models while Chinese developers gain advantages from wider dissemination.
• 16:25
  Speaker 1: His argument is not that every Chinese model is superior
• 16:28
  Speaker 1: or that open development automatically guarantees victory. Rather, he is
• 16:33
  Speaker 1: emphasizing the possibility that a widely available American model could
• 16:38
  Speaker 1: have strategic value by giving developers around the world a
• 16:41
  Speaker 1: reason to build their applications around American technology. In the
• 16:46
  Speaker 1: technology industry, network effects can be extraordinarily powerful. Once a
• 16:51
  Speaker 1: platform becomes widespread, Developers produce tools for it, businesses adopt
• 16:57
  Speaker 1: those tools, universities teach related skills, cloud providers optimize for it,
• 17:02
  Speaker 1: and investors fund companies based on its continued relevance. That
• 17:07
  Speaker 1: ecosystem can become self-reinforcing. The question is whether the United
• 17:12
  Speaker 1: States wants American AI to influence that ecosystem primarily through
• 17:17
  Speaker 1: a small collection of closed platforms or through a broader
• 17:20
  Speaker 1: mixture of proprietary and open systems. China's approach has made
• 17:25
  Speaker 1: the question more urgent because Chinese researchers and companies have
• 17:29
  Speaker 1: demonstrated that capable models can increasingly be developed under tighter
• 17:33
  Speaker 1: resource constraints and with different distribution strategies. Besson's comments about
• 17:38
  Speaker 1: distillation add another layer to the competition. If a Chinese
• 17:43
  Speaker 1: company can observe outputs from a leading American closed model
• 17:47
  Speaker 1: and use those outputs to improve another model, then some
• 17:50
  Speaker 1: of the advantage generated by expensive American research could effectively
• 17:55
  Speaker 1: spread beyond the companies that created it. American officials argue
• 17:59
  Speaker 1: that such practices can amount to unauthorized appropriation when the
• 18:03
  Speaker 1: process crosses legal boundaries. China has rejected some of the
• 18:08
  Speaker 1: strongest U.S. accusations and has itself emphasized the importance of
• 18:13
  Speaker 1: open and accessible AI development. That disagreement is likely to
• 18:17
  Speaker 1: continue because the technology is evolving faster than many traditional
• 18:22
  Speaker 1: intellectual property frameworks were designed to handle. Copyright, trade secrets,
• 18:27
  Speaker 1: model weights, training data, and generated outputs do not always
• 18:32
  Speaker 1: fit neatly into existing categories. A company may invest billions
• 18:36
  Speaker 1: of dollars in training infrastructure while another organization attempts to
• 18:41
  Speaker 1: reproduce the resulting capabilities indirectly. Determining where legitimate research ends
• 18:46
  Speaker 1: and unlawful copying begins can become extremely difficult. The strategic
• 18:51
  Speaker 1: challenge for the United States is therefore to protect valuable
• 18:55
  Speaker 1: intellectual property without restricting itself so heavily that American developers
• 19:00
  Speaker 1: lose the ability to compete. Open models may provide part
• 19:04
  Speaker 1: of that strategy by deliberately distributing American technology on terms
• 19:09
  Speaker 1: that encourage adoption rather than leaving foreign competitors to determine
• 19:13
  Speaker 1: whether which systems become globally popular. If developers in other
• 19:17
  Speaker 1: countries can legally access a powerful American model, customize it,
• 19:22
  Speaker 1: and build businesses around it, the United States may retain
• 19:26
  Speaker 1: influence even when users operate outside American borders. This is
• 19:31
  Speaker 1: particularly relevant for emerging markets and organizations that cannot afford
• 19:35
  Speaker 1: the high costs associated with premium proprietary AI services. An
• 19:40
  Speaker 1: open model can be adapted to local languages, specialized industries,
• 19:45
  Speaker 1: and unique technical requirements. That flexibility may increase adoption. and
• 19:50
  Speaker 1: create new demand for American hardware, cloud infrastructure, software tools,
• 19:56
  Speaker 1: and expertise. In that sense, open-source AI could become a
• 20:00
  Speaker 1: form of economic diplomacy even without being officially treated as such.
• 20:05
  Speaker 1: Technology standards and developer ecosystems often create long-lasting relationships. Countries
• 20:11
  Speaker 1: that build their digital infrastructure around one ecosystem may continue
• 20:16
  Speaker 1: using compatible services for years. Washington, therefore, has an incentive
• 20:20
  Speaker 1: to think about AI competition in terms broader than benchmark scores.
• 20:25
  Speaker 1: The question is also about influence, accessibility, and scale. A
• 20:30
  Speaker 1: highly capable American model used by only a limited number
• 20:33
  Speaker 1: of companies may generate enormous revenue for its owner, but
• 20:37
  Speaker 1: an open model used by thousands of developers could generate
• 20:41
  Speaker 1: a much wider ecosystem effect. For American businesses, the push
• 20:46
  Speaker 1: toward open AI models could have consequences that extend far
• 20:50
  Speaker 1: beyond the rivalry with China. The most immediate effect could
• 20:53
  Speaker 1: be greater competition in the AI marketplace. When advanced models
• 20:58
  Speaker 1: remain entirely controlled by a handful of companies, businesses often
• 21:03
  Speaker 1: have limited choices about pricing, technical specifications, access policies, and
• 21:09
  Speaker 1: the direction of future upgrades. An open model can create
• 21:13
  Speaker 1: another option. A company might deploy the system internally, customize
• 21:18
  Speaker 1: it for a specific industry, or combine it with its
• 21:20
  Speaker 1: own data and software without depending entirely on a single
• 21:24
  Speaker 1: external provider. That flexibility can be particularly attractive in sectors
• 21:29
  Speaker 1: where privacy, latency, cost, or regulatory requirements make cloud-based proprietary
• 21:35
  Speaker 1: services less practical. Banks, manufacturers, hospitals, government agencies, research institutions,
• 21:43
  Speaker 1: and logistics companies all have different requirements. A general-purpose model
• 21:48
  Speaker 1: may be impressive, but a specialized model trained or adapted
• 21:52
  Speaker 1: for a specific environment can sometimes be more valuable. Open
• 21:56
  Speaker 1: development makes that kind of specialization easier because users can
• 22:00
  Speaker 1: gain more control over the model itself. That can accelerate experimentation.
• 22:06
  Speaker 1: One developer may optimize an open model for medical terminology.
• 22:11
  Speaker 1: Another may adapt it to legal documents. Another may make
• 22:14
  Speaker 1: it work more efficiently on relatively inexpensive hardware. Those improvements
• 22:19
  Speaker 1: can then inspire additional developments across the ecosystem. The result
• 22:24
  Speaker 1: is a network of innovation that can move in many
• 22:26
  Speaker 1: directions simultaneously rather than following a single corporate roadmap. Besson's
• 22:32
  Speaker 1: call for more open-source models can therefore be seen as
• 22:35
  Speaker 1: an argument for increasing the number of paths through which
• 22:38
  Speaker 1: American AI can advance.
• 22:40
  Speaker 2: The economic implications could be significant.
• 22:44
  Speaker 1: A larger ecosystem may create more startups, more demand for
• 22:48
  Speaker 1: computing infrastructure, and more opportunities for developers with specialized expertise.
• 22:54
  Speaker 1: It could also encourage competition over efficiency. If several companies
• 22:58
  Speaker 1: are working from powerful open models, they may compete on
• 23:02
  Speaker 1: deployment costs, tools, and hardware optimization, and specialized applications instead
• 23:08
  Speaker 1: of each spending enormous sums to recreate the entire foundation
• 23:12
  Speaker 1: model from scratch. That could reduce barriers to entry. At
• 23:16
  Speaker 1: the same time, open models could challenge the business models
• 23:20
  Speaker 1: of established AI providers. Companies that rely on charging for
• 23:25
  Speaker 1: access to proprietary systems may face pressure if equivalent or
• 23:29
  Speaker 1: near-equivalent capabilities become available through models that customers can deploy themselves.
• 23:35
  Speaker 1: That pressure could encourage providers to innovate faster, lower prices,
• 23:40
  Speaker 1: offer better services, or differentiate themselves through proprietary features that
• 23:45
  Speaker 1: are difficult to reproduce. In that sense, Open-source competition could
• 23:50
  Speaker 1: strengthen the entire market even if it reduces the dominance
• 23:54
  Speaker 1: of individual companies.
• 23:56
  Speaker 2: But there are risks.
• 23:57
  Speaker 1: The economics of large AI systems remain heavily dependent on
• 24:01
  Speaker 1: computing costs. Training advanced models can require vast amounts of
• 24:06
  Speaker 1: specialized hardware, electricity, and data center capacity. An open model
• 24:11
  Speaker 1: does not eliminate those costs. It changes who can access
• 24:15
  Speaker 1: the resulting technology. companies may still need substantial resources to
• 24:20
  Speaker 1: fine-tune or operate the models at scale. That creates another
• 24:24
  Speaker 1: strategic issue for the United States. If America wants an
• 24:28
  Speaker 1: open AI ecosystem to flourish, it may also need a
• 24:31
  Speaker 1: strong domestic infrastructure ecosystem capable of providing affordable computing. This
• 24:37
  Speaker 1: links AI policy to energy policy, chip manufacturing, cloud computing,
• 24:42
  Speaker 1: and data center investment. The technology cannot be separated from
• 24:47
  Speaker 1: the physical infrastructure supporting it. Besant's economic perspective is therefore
• 24:52
  Speaker 1: relevant because the AI race intersects directly with the broader
• 24:56
  Speaker 1: American economy. Open models can contribute to productivity, but productivity
• 25:01
  Speaker 1: gains depend on adoption. A powerful model that remains inaccessible
• 25:06
  Speaker 1: to most businesses will have a smaller economic effect than
• 25:10
  Speaker 1: one that can be integrated into thousands of workflows. That
• 25:13
  Speaker 1: makes distribution a strategic question. There is also a workforce dimension.
• 25:19
  Speaker 1: Wider model availability could allow smaller companies to build sophisticated
• 25:23
  Speaker 1: AI applications without hiring enormous teams of machine learning specialists.
• 25:29
  Speaker 1: Developers may increasingly focus on adapting existing models rather than
• 25:33
  Speaker 1: training new ones from scratch. That could shift where value
• 25:37
  Speaker 1: is created in the technology industry. Instead of concentrating almost
• 25:42
  Speaker 1: all value in foundation model companies, more value could migrate
• 25:46
  Speaker 1: toward application developers, data specialists, industry experts, and organizations with
• 25:52
  Speaker 1: unique domain knowledge. The open model economy could therefore resemble
• 25:57
  Speaker 1: earlier technology waves in which foundational software became widely available
• 26:01
  Speaker 1: and companies built new businesses on top of it. Yet
• 26:05
  Speaker 1: policymakers must still decide where safeguards belong. The disagreement over
• 26:10
  Speaker 1: open and closed AI also reflects a deeper economic question
• 26:14
  Speaker 1: about concentration of power in the technology industry. Artificial intelligence
• 26:20
  Speaker 1: requires extraordinary resources at the frontier, and that naturally gives
• 26:24
  Speaker 1: an advantage to companies with the capital to build massive
• 26:27
  Speaker 1: data centers, acquire advanced chips, train large teams, and absorb
• 26:32
  Speaker 1: years of development costs before a product becomes profitable. those
• 26:37
  Speaker 1: barriers can reinforce the position of existing giants. If the
• 26:41
  Speaker 1: frontier keeps moving upward in cost, the number of organizations
• 26:45
  Speaker 1: capable of independently building state-of-the-art models could shrink. Besson's concern
• 26:50
  Speaker 1: about regulatory capture can therefore be interpreted within a larger
• 26:55
  Speaker 1: conversation about market concentration. Regulations have consequences not only for
• 27:00
  Speaker 1: safety but also for competition. A requirement that sounds manageable
• 27:05
  Speaker 1: to a company with billions of dollars in annual revenue
• 27:08
  Speaker 1: could be extremely difficult for a startup with a small
• 27:11
  Speaker 1: engineering team. The result could be an industry in which
• 27:15
  Speaker 1: the largest players become more powerful precisely because they are
• 27:19
  Speaker 1: best positioned to satisfy increasingly expensive compliance requirements. Open-source development
• 27:25
  Speaker 1: offers one possible counterweight because it allows many organizations to
• 27:29
  Speaker 1: begin from a common technological foundation instead of each attempting
• 27:33
  Speaker 1: to recreate the enormous research and infrastructure effort required to
• 27:37
  Speaker 1: build a frontier model. This does not eliminate concentration entirely.
• 27:42
  Speaker 1: Cloud providers, chip manufacturers, and major AI laboratories may still
• 27:47
  Speaker 1: dominate key layers of the ecosystem, but it can distribute
• 27:51
  Speaker 1: innovation more widely. It could also alter the relationship between
• 27:55
  Speaker 1: corporations and researchers. Universities and independent laboratories often cannot afford
• 28:01
  Speaker 1: to train frontier models themselves, but access to strong open
• 28:05
  Speaker 1: systems may allow them to conduct meaningful research on model behavior, efficiency, safety,
• 28:11
  Speaker 1: and specialized applications. That research can feed back into the
• 28:15
  Speaker 1: wider industry. Open models can also make it easier to
• 28:19
  Speaker 1: compare systems. Researchers can inspect performance, test weaknesses, and develop
• 28:24
  Speaker 1: independent benchmarks.
• 28:26
  Speaker 2: Closed models can.
• 28:27
  Speaker 1: Make some of that work more difficult because users may
• 28:30
  Speaker 1: have limited information about internal architecture, training data, or updates.
• 28:36
  Speaker 1: Transparency is therefore one of the strongest arguments made by
• 28:40
  Speaker 1: supporters of openness. Yet transparency has limits. An open model
• 28:45
  Speaker 1: does not automatically reveal everything about how it was created.
• 28:49
  Speaker 1: Training data can remain partly undisclosed. Licensing terms can restrict
• 28:54
  Speaker 1: certain uses and model behavior may still be difficult to interpret.
• 28:59
  Speaker 1: The technology itself remains complicated. A neural network can contain
• 29:04
  Speaker 1: billions of parameters whose relationships are not easily translated into
• 29:08
  Speaker 1: simple explanations. That means openness should not be confused with
• 29:13
  Speaker 1: complete understandability. Policymakers need to recognize that difference. They also
• 29:19
  Speaker 1: need to consider the possibility that some capabilities are safer
• 29:22
  Speaker 1: when kept under controlled access. A model with strong biological, cyber,
• 29:27
  Speaker 1: or autonomous agent capabilities might present different risks from a
• 29:31
  Speaker 1: basic language model or coding assistant. A single policy for
• 29:36
  Speaker 1: every AI system may therefore be inappropriate. Besson's testimony points
• 29:41
  Speaker 1: toward a model in which open systems are encouraged while
• 29:44
  Speaker 1: specific risks are addressed through targeted rules. That approach could
• 29:48
  Speaker 1: involve stronger restrictions on particular capabilities rather than broad limits
• 29:53
  Speaker 1: on the entire category of open models. The challenge is
• 29:57
  Speaker 1: defining those thresholds in a way that remains workable as
• 30:00
  Speaker 1: technology evolves. AI development can move quickly enough that a
• 30:05
  Speaker 1: rule designed around one model generation may become outdated within
• 30:09
  Speaker 1: a year. Regulation that is too specific may quickly become obsolete,
• 30:14
  Speaker 1: while regulation that is too broad could inhibit legitimate development.
• 30:18
  Speaker 1: The administration's voluntary framework, as described by Fox Business, indicates
• 30:24
  Speaker 1: an effort to distinguish open or open-weight systems from proprietary
• 30:28
  Speaker 1: models when considering pre-release security review. That policy approach is
• 30:34
  Speaker 1: itself part of the broader experiment. The United States is
• 30:37
  Speaker 1: effectively testing whether market incentives, voluntary standards, and targeted restrictions
• 30:43
  Speaker 1: can preserve rapid development without leaving major risks unaddressed. The
• 30:48
  Speaker 1: outcome will matter far beyond one administration. AI regulation is
• 30:53
  Speaker 1: likely to evolve over multiple presidential terms and through future
• 30:57
  Speaker 1: congressional debates. Businesses need predictable rules, but predictable does not
• 31:02
  Speaker 1: necessarily mean strict or permissive. It means companies should have
• 31:07
  Speaker 1: some understanding of what conduct will be allowed, what obligations
• 31:10
  Speaker 1: will apply, and how those obligations will change. Uncertainty can
• 31:15
  Speaker 1: discourage investment as effectively as heavy regulations. The United States
• 31:20
  Speaker 1: has another advantage in the AI competition that is easy
• 31:23
  Speaker 1: to overlook when discussions focus entirely on which laboratory has
• 31:28
  Speaker 1: the most advanced model, the strength of its broader technology ecosystem.
• 31:33
  Speaker 1: American universities, software companies, semiconductor designers, cloud providers, venture capital firms,
• 31:41
  Speaker 1: research institutions, and global technology platforms form an interconnected network
• 31:47
  Speaker 1: that has supported rapid innovation for decades. The question raised
• 31:52
  Speaker 1: by Besant is whether that ecosystem can be extended further
• 31:55
  Speaker 1: through more open AI development. Open models could provide a
• 31:59
  Speaker 1: common foundation on which companies and researchers build specialized systems,
• 32:05
  Speaker 1: potentially increasing the speed at which ideas move from laboratories
• 32:09
  Speaker 1: into commercial applications. This could be especially important as AI
• 32:14
  Speaker 1: shifts from demonstrations toward widespread deployment. The first phase of
• 32:18
  Speaker 1: the AI boom was dominated by impressive models that generated text, images, code,
• 32:24
  Speaker 1: and other content. The next phase is likely to involve
• 32:28
  Speaker 1: integration into ordinary business operations, where the practical questions are
• 32:32
  Speaker 1: less glamorous, but far more consequential. Can an AI system
• 32:37
  Speaker 1: reduce manufacturing waste? Can it help engineers identify design problems?
• 32:42
  Speaker 1: Can it automate financial analysis? Can it assist scientists in
• 32:46
  Speaker 1: processing large datasets? Can it make government services more efficient?
• 32:51
  Speaker 1: The answer to those questions will depend heavily on cost, reliability, security,
• 32:57
  Speaker 1: and adaptability. Open models may help on the adaptability side
• 33:01
  Speaker 1: because organizations can modify them for specific purposes. A company
• 33:06
  Speaker 1: may decide that it does not need the broadest possible model,
• 33:09
  Speaker 1: but instead needs one that performs exceptionally well in a
• 33:13
  Speaker 1: narrow technical field. An open architecture can make that specialization
• 33:17
  Speaker 1: easier to pursue. It also gives businesses greater leverage when
• 33:21
  Speaker 1: negotiating with large AI providers. If a customer has no
• 33:26
  Speaker 1: alternative to a proprietary model, the provider holds substantial power
• 33:30
  Speaker 1: over price and access. If the customer can deploy an
• 33:34
  Speaker 1: open system or switch between several models, competition increases. That
• 33:39
  Speaker 1: dynamic can encourage better pricing and service. The effect could
• 33:44
  Speaker 1: be particularly important for small and mid-sized businesses, which may
• 33:48
  Speaker 1: not have the resources to negotiate large enterprise agreements with
• 33:52
  Speaker 1: technology giants.
• 33:54
  Speaker 2: In economic terms.
• 33:56
  Speaker 1: The availability of strong open models could lower the cost
• 33:59
  Speaker 1: of experimentation. A startup could test an idea with an
• 34:03
  Speaker 1: existing model before raising enough capital to develop a proprietary system.
• 34:09
  Speaker 1: A regional manufacturer could experiment with AI without building a
• 34:13
  Speaker 1: huge internal research organization. A university could create applications tailored
• 34:19
  Speaker 1: to its own researchers or students. Government agencies could evaluate
• 34:23
  Speaker 1: different systems without being locked immediately.
• 34:26
  Speaker 2: Into one vendor.
• 34:27
  Speaker 1: This diversity could produce an ecosystem in which many small
• 34:31
  Speaker 1: improvements accumulate. At the frontier, progress can sometimes come from
• 34:36
  Speaker 1: dramatic breakthroughs, but in deployment, progress often comes from thousands
• 34:42
  Speaker 1: of incremental adaptations. Open systems are well suited to that pattern.
• 34:47
  Speaker 1: Developers can identify weaknesses, create improvements, and publish modifications that
• 34:53
  Speaker 1: others can use. Over time, That process can turn a
• 34:57
  Speaker 1: general model into a foundation for a wide range of
• 35:00
  Speaker 1: specialized technologies.
• 35:02
  Speaker 2: Yet the U.S.
• 35:03
  Speaker 1: Advantage cannot be assumed to last forever. China has demonstrated
• 35:07
  Speaker 1: that it can mobilize significant resources around strategic technologies, and
• 35:12
  Speaker 1: Chinese firms can benefit from a large domestic market that
• 35:16
  Speaker 1: encourages rapid scaling.
• 35:18
  Speaker 2: The global availability of.
• 35:20
  Speaker 1: AI models means technical advances can spread faster than they
• 35:24
  Speaker 1: did in previous technological eras. That makes ecosystem speed important.
• 35:30
  Speaker 1: If an American company creates a powerful model but restricts
• 35:33
  Speaker 1: it tightly while another country provides a capable open alternative,
• 35:38
  Speaker 1: developers may gradually build around the alternative. Once an ecosystem
• 35:43
  Speaker 1: reaches critical mass, shifting away from it can be difficult.
• 35:47
  Speaker 1: Software compatibility, developer training, data pipelines, and institutional knowledge create
• 35:53
  Speaker 1: switching costs. Besson's emphasis on open-source models can therefore be
• 35:58
  Speaker 1: understood as an effort to influence the direction of those
• 36:01
  Speaker 1: network effects before they become entrenched. The strategy does not
• 36:05
  Speaker 1: require every American model to be open. Closed systems can
• 36:09
  Speaker 1: remain highly valuable, especially for companies that need strong control
• 36:14
  Speaker 1: over sensitive capabilities or proprietary customer data. The argument is
• 36:19
  Speaker 1: that the market should contain enough open alternatives to prevent
• 36:22
  Speaker 1: openness from becoming an afterthought. A healthy AI ecosystem may
• 36:27
  Speaker 1: ultimately require both approaches. Proprietary models pushing the frontier of
• 36:32
  Speaker 1: performance and open models distributing capabilities broadly enough to stimulate
• 36:37
  Speaker 1: experimentation and competition. The challenge for policymakers is to avoid
• 36:42
  Speaker 1: regulations that unintentionally eliminate one.
• 36:45
  Speaker 2: Side of that ecosystem.
• 36:47
  Speaker 1: The strongest argument against a purely expansionary approach to open
• 36:51
  Speaker 1: AI is also the hardest problem for policymakers to solve security.
• 36:57
  Speaker 1: Artificial intelligence can be used for productive purposes, but advanced
• 37:01
  Speaker 1: systems may also provide capabilities that create risks in cybersecurity, disinformation, fraud, surveillance,
• 37:09
  Speaker 1: or the development of dangerous technical applications. When access is
• 37:14
  Speaker 1: tightly controlled, the developer can monitor usage, impose restrictions, and
• 37:19
  Speaker 1: intervene when suspicious behavior occurs. When a model is distributed openly,
• 37:24
  Speaker 1: that control becomes weaker. The model may be copied, modified,
• 37:29
  Speaker 1: or integrated into systems that the original developer never sees.
• 37:33
  Speaker 1: Supporters of openness answer that centralized control has.
• 37:37
  Speaker 2: Risks of its own.
• 37:39
  Speaker 1: If only a small number of private companies control powerful
• 37:42
  Speaker 1: AI systems, the public, researchers, and smaller businesses become dependent
• 37:48
  Speaker 1: on those companies. Errors, biases, or security weaknesses may remain
• 37:53
  Speaker 1: hidden behind proprietary barriers. Open models can create opportunities for
• 37:58
  Speaker 1: independent researchers to test systems and identify vulnerabilities. They can
• 38:03
  Speaker 1: also make it easier for organizations to run AI locally
• 38:06
  Speaker 1: rather than sending sensitive information to external servers. Both arguments
• 38:11
  Speaker 1: are legitimate, which is why the policy debate cannot be
• 38:14
  Speaker 1: reduced to the simple question of whether openness is good
• 38:17
  Speaker 1: or bad. The relevant question is what type of model
• 38:20
  Speaker 1: is being opened, what capabilities it possesses, what safeguards are
• 38:25
  Speaker 1: attached to it, and how much control remains possible after release.
• 38:30
  Speaker 1: A smaller language model designed for ordinary business tasks may
• 38:34
  Speaker 1: present very different risks from a highly advanced autonomous system
• 38:38
  Speaker 1: capable of operating tools, writing sophisticated code, or assisting with
• 38:42
  Speaker 1: complex scientific research. The technology is not uniform. Regulation therefore
• 38:48
  Speaker 1: has to account for capability differences. This is one reason
• 38:52
  Speaker 1: the distinction between open-source and open-weight models also matters. Some
• 38:57
  Speaker 1: systems may make their trained weights available without providing the
• 39:01
  Speaker 1: full training process or dataset. Others may provide more extensive components.
• 39:06
  Speaker 1: Policymakers have to understand those differences before designing roles. The
• 39:11
  Speaker 1: Trump administration's approach, as reported by Fox Business. has been
• 39:16
  Speaker 1: to exempt open source and open-weight models from certain pre-release
• 39:20
  Speaker 1: government security reviews under its voluntary framework while placing greater
• 39:25
  Speaker 1: scrutiny on proprietary systems. That policy reflects a judgment that
• 39:30
  Speaker 1: broader openness can support innovation and that not every model
• 39:34
  Speaker 1: should be subjected to the same process critics could argue
• 39:37
  Speaker 1: that releasing a capable model without formal review creates avoidable
• 39:42
  Speaker 1: risks Supporters could argue that forcing every model through the
• 39:46
  Speaker 1: same review could drive development toward the largest companies that
• 39:50
  Speaker 1: can afford the compliance burden. Besson's testimony essentially strengthens the
• 39:55
  Speaker 1: second concern. His warning about regulatory capture suggests that safety
• 40:00
  Speaker 1: rules should not be designed in ways that unintentionally suppress competition.
• 40:05
  Speaker 1: Yet competition itself is not a substitute for safety. The
• 40:09
  Speaker 1: challenge is determining how to make competitive development responsible. One
• 40:13
  Speaker 1: possible path involves capability thresholds, where models that meet certain
• 40:18
  Speaker 1: risk levels face additional requirements while lower-risk systems remain more accessible.
• 40:24
  Speaker 1: Another possibility involves targeted restrictions on particular uses rather than
• 40:28
  Speaker 1: broad restrictions on the underlying model. Export controls, intellectual property enforcement,
• 40:35
  Speaker 1: Cybersecurity standards and laws against malicious deployment can also operate
• 40:40
  Speaker 1: alongside open development. The objective would be to address concrete
• 40:45
  Speaker 1: harms without blocking legitimate research. The United States is not
• 40:49
  Speaker 1: making this choice in a vacuum. China is building its
• 40:53
  Speaker 1: own AI industry under a different regulatory and strategic environment,
• 40:57
  Speaker 1: and American policymakers cannot assume that restrictive measures imposed domestically
• 41:03
  Speaker 1: will have equivalent effects abroad. If advanced AI becomes easier
• 41:07
  Speaker 1: to develop and distribute globally, the balance between openness and
• 41:12
  Speaker 1: control will become increasingly international. Developers can collaborate across borders,
• 41:18
  Speaker 1: models can be reproduced, and technical knowledge can spread quickly.
• 41:23
  Speaker 1: That means national policy can influence, but not fully determine
• 41:27
  Speaker 1: how AI evolves. It also means international standards may eventually
• 41:31
  Speaker 1: become necessary. The challenge is finding areas of agreement without
• 41:36
  Speaker 1: creating rules that freeze competition or prevent beneficial experimentation. The
• 41:41
  Speaker 1: upcoming discussions between American and Chinese leaders are likely to
• 41:45
  Speaker 1: include AI precisely because both countries recognize how significant the
• 41:50
  Speaker 1: technology has become. Yet geopolitical rivalry makes cooperation difficult. Besson's
• 41:57
  Speaker 1: testimony also sends a message to the technology industry about
• 42:01
  Speaker 1: the direction policymakers may want to see in the coming years.
• 42:05
  Speaker 1: Washington is increasingly treating AI development as an issue of
• 42:09
  Speaker 1: national strategy rather than a matter that can be left
• 42:12
  Speaker 1: entirely to private companies. The reason is straightforward. Artificial intelligence
• 42:18
  Speaker 1: now affects economic competitiveness, defense capabilities, productivity, research, infrastructure.
• 42:25
  Speaker 2: And global influence.
• 42:27
  Speaker 1: Once a technology reaches that level of importance, governments begin
• 42:31
  Speaker 1: looking for ways to shape its development. The difficulty is
• 42:35
  Speaker 1: deciding where government involvement helps and where it starts to
• 42:38
  Speaker 1: create obstacles. Besson's argument suggests that the administration sees government
• 42:43
  Speaker 1: primarily as an enabler of a broader American AI ecosystem
• 42:47
  Speaker 1: rather than as the manager of individual companies. Encouraging open
• 42:52
  Speaker 1: models is Reducing unnecessary barriers and protecting domestic innovation could
• 42:58
  Speaker 1: create conditions in which more firms compete. At the same time,
• 43:03
  Speaker 1: Washington will likely continue to address national security concerns related
• 43:07
  Speaker 1: to advanced AI and foreign access to American technology. The
• 43:12
  Speaker 1: result may be a policy environment in which openness is
• 43:16
  Speaker 1: encouraged in some areas while restrictions are tightened around other areas.
• 43:21
  Speaker 1: For Silicon Valley, that could create both opportunities and uncertainties.
• 43:26
  Speaker 1: Companies developing open systems may benefit from a friendlier regulatory environment,
• 43:31
  Speaker 1: while firms that depend on closed models could face questions
• 43:35
  Speaker 1: about competition, intellectual property, and market concentration. Large laboratories may
• 43:41
  Speaker 1: also discover that their role is changing. Instead of assuming
• 43:45
  Speaker 1: that the future belongs entirely to proprietary models, they may
• 43:49
  Speaker 1: need to compete in a market where open systems are
• 43:52
  Speaker 1: good enough for a growing share of applications. That could
• 43:55
  Speaker 1: push companies toward more efficient architectures, better tools, specialized enterprise services,
• 44:01
  Speaker 1: and differentiated products built around models rather than relying exclusively
• 44:06
  Speaker 1: on model access itself. The most valuable layer of the
• 44:10
  Speaker 1: AI economy may not always be the model. It could
• 44:14
  Speaker 1: be the infrastructure surrounding the model. Data management, security, agent frameworks,
• 44:20
  Speaker 1: workflow integration, evaluation tools, custom deployment, and specialized applications are
• 44:26
  Speaker 1: all potential areas of growth. Open models may actually increase
• 44:31
  Speaker 1: demand for those services because companies need help adapting and
• 44:35
  Speaker 1: operating them. In that scenario, openness would not destroy the
• 44:39
  Speaker 1: commercial AI market. It would redistribute value within it. Cloud
• 44:44
  Speaker 1: providers could sell compute, Software firms could provide deployment platforms
• 44:49
  Speaker 1: and consultants could specialize in AI integration. Hardware companies could
• 44:55
  Speaker 1: benefit from the increased number of organizations experimenting with models.
• 44:59
  Speaker 1: Universities could train more developers using accessible systems. The entire
• 45:04
  Speaker 1: ecosystem could expand. This is one reason the July letter
• 45:09
  Speaker 1: signed by a broad group of major technology companies is significant.
• 45:13
  Speaker 1: the companies did not argue that all AI should be unrestricted.
• 45:18
  Speaker 1: Their stated position was that illegal attempts to extract value
• 45:21
  Speaker 1: from proprietary models should be addressed through targeted legal mechanisms
• 45:26
  Speaker 1: rather than broad restrictions on open development. That suggests an
• 45:30
  Speaker 1: emerging industry consensus around a middle path. Companies want protection
• 45:35
  Speaker 1: from intellectual property theft and dangerous misuse. but they also
• 45:40
  Speaker 1: fear rules that make open innovation so expensive that only
• 45:44
  Speaker 1: established giants can participate. The economic argument is closely connected
• 45:49
  Speaker 1: to the geopolitical argument. A dynamic domestic ecosystem is more
• 45:54
  Speaker 1: likely to produce the next generation of AI breakthroughs than
• 45:58
  Speaker 1: a highly concentrated market where only a few firms can
• 46:01
  Speaker 1: afford to experiment. Competition creates redundancy. If one approach fails,
• 46:07
  Speaker 1: another can continue. If one company falls behind, others can
• 46:12
  Speaker 1: move forward. If developers discover a new technique, they can
• 46:16
  Speaker 1: adapt it across multiple models. That diversity may help the
• 46:20
  Speaker 1: United States respond to technological shifts more quickly. China's growing
• 46:25
  Speaker 1: AI capabilities increase the urgency because American dominance cannot be
• 46:30
  Speaker 1: assumed to continue automatically. Reuters has reported that the AI
• 46:35
  Speaker 1: competition is becoming an important part of broader U.S.-China relations,
• 46:40
  Speaker 1: with both countries recognizing the connection between advanced AI and
• 46:44
  Speaker 1: national power. That environment creates pressure for American policymakers to
• 46:49
  Speaker 1: think about technology policy in strategic terms. Besson's testimony reflects
• 46:55
  Speaker 1: that pressure. He is effectively arguing that the United States
• 46:59
  Speaker 1: should not let concerns about control and regulation produce a
• 47:03
  Speaker 1: market structure that slows American innovation while other nations continue advancing.
• 47:08
  Speaker 1: But there is another side to the equation. Open development
• 47:12
  Speaker 1: must remain compatible with responsible security practices, intellectual property enforcement,
• 47:18
  Speaker 1: and legitimate protections for users. The most important lesson from
• 47:22
  Speaker 1: the debate may be that the global AI race cannot
• 47:25
  Speaker 1: be reduced to a scoreboard showing which company has the
• 47:28
  Speaker 1: highest benchmark result. Models matter, but so do the ecosystems
• 47:33
  Speaker 1: built around them. An AI system becomes strategically influential when
• 47:38
  Speaker 1: developers use it, businesses build products around it, universities teach it,
• 47:43
  Speaker 1: governments integrate it, and researchers continue improving it. That is
• 47:48
  Speaker 1: why open models can have a significance that exceeds their
• 47:51
  Speaker 1: immediate technical performance. They can become building blocks for other technologies.
• 47:57
  Speaker 1: A widely available model may be less powerful than the
• 48:00
  Speaker 1: absolute frontier model in some tests and still have a
• 48:03
  Speaker 1: greater impact because far more people can access it. That
• 48:08
  Speaker 1: creates a different form of technological leverage. The United States
• 48:12
  Speaker 1: and China are competing not only to create powerful AI,
• 48:16
  Speaker 1: but also to shape the standards, platforms, and developer communities
• 48:20
  Speaker 1: that will define the next generation of digital services. Besson's
• 48:24
  Speaker 1: support for more American open-source models can therefore be viewed
• 48:28
  Speaker 1: as an ecosystem strategy. The objective is to ensure that
• 48:32
  Speaker 1: American technology remains present wherever AI development occurs, including among
• 48:37
  Speaker 1: organizations that cannot afford access to the most expensive proprietary systems.
• 48:43
  Speaker 1: That global distribution could have economic and strategic benefits. When
• 48:48
  Speaker 1: businesses around the world use American technology, they often become
• 48:52
  Speaker 1: connected to American suppliers, cloud platforms, software tools, and professional services.
• 48:59
  Speaker 1: Those relationships can reinforce U.S.
• 49:01
  Speaker 2: Influence.
• 49:03
  Speaker 1: At the same time, widespread use creates incentives for continued
• 49:07
  Speaker 1: American investment because companies have access to larger markets. Open
• 49:12
  Speaker 1: models could be especially important in countries where budgets are
• 49:16
  Speaker 1: smaller and local adaptation is essential. An openly available model
• 49:21
  Speaker 1: can be translated, optimized, and modified by local developers in
• 49:26
  Speaker 1: ways that a closed service may not permit. This could
• 49:29
  Speaker 1: give American AI a broader geographic footprint. But the strategy
• 49:34
  Speaker 1: also depends on trust. Developers need confidence that open models
• 49:38
  Speaker 1: are legally usable, technically reliable and supported by a healthy ecosystem.
• 49:44
  Speaker 1: Companies need confidence that building on an American model will
• 49:48
  Speaker 1: not expose them to unpredictable policy changes. Governments need confidence
• 49:53
  Speaker 1: that the models can be deployed safely. If those conditions
• 49:56
  Speaker 1: are missing, developers may choose alternatives regardless of model quality.
• 50:02
  Speaker 1: China understands that dynamic as well. Chinese technology companies have
• 50:07
  Speaker 1: increasingly promoted the idea that AI should be broadly accessible
• 50:11
  Speaker 1: and useful rather than controlled exclusively by a few Western firms.
• 50:16
  Speaker 1: Chinese officials have also argued that AI should not become
• 50:19
  Speaker 1: the monopoly of major powers and have called for wider
• 50:23
  Speaker 1: participation in its development. American policymakers view some of those
• 50:28
  Speaker 1: statements with skepticism because of the broader strategic rivalry and
• 50:32
  Speaker 1: allegations that surrounding intellectual property. Yet the competition over openness
• 50:38
  Speaker 1: itself is revealing. Both sides understand that distribution matters. Neither
• 50:43
  Speaker 1: wants the other to dominate the global AI ecosystem. The
• 50:47
  Speaker 1: United States has historically benefited from open standards and developer
• 50:51
  Speaker 1: communities across the technology industry, from operating systems to programming
• 50:56
  Speaker 1: languages and web infrastructure. AI may follow a similar pattern,
• 51:01
  Speaker 1: although its economic and security consequences are much greater. The
• 51:05
  Speaker 1: country that establishes the most widely used foundations could influence
• 51:10
  Speaker 1: how businesses operate, how software is written, and how future
• 51:14
  Speaker 1: AI capabilities are developed. That influence is not necessarily determined
• 51:19
  Speaker 1: by government policy alone. It depends on whether developers find
• 51:23
  Speaker 1: the technology useful, affordable, and adaptable. Besson's message is therefore
• 51:29
  Speaker 1: ultimately directed toward creating conditions in which American developers can
• 51:33
  Speaker 1: compete on those dimensions. The emphasis on open models also
• 51:37
  Speaker 1: challenges the idea that national AI leadership should be measured
• 51:41
  Speaker 1: only by the size of a few research laboratories. Leadership
• 51:45
  Speaker 1: may instead depend on the resilience of the entire ecosystem.
• 51:50
  Speaker 1: A country with 10 strong AI companies, a large developer community,
• 51:55
  Speaker 1: a powerful semiconductor industry. Extensive cloud infrastructure and universities producing
• 52:01
  Speaker 1: skilled researchers may be better positioned for sustained competition than
• 52:05
  Speaker 1: a country with one or two extraordinary models, but a
• 52:09
  Speaker 1: narrower supporting base. That broader perspective is increasingly important as
• 52:14
  Speaker 1: the technology matures. Frontier breakthroughs will remain critical, but deployment
• 52:19
  Speaker 1: will determine how much economic value those breakthroughs create. Open
• 52:23
  Speaker 1: models could accelerate deployment by reducing costs and allowing customization.
• 52:29
  Speaker 1: Closed models could retain advantages in performance, reliability, and controlled access.
• 52:35
  Speaker 2: A mixed ecosystem may.
• 52:37
  Speaker 1: Therefore be the most realistic future. Besson's testimony does not
• 52:41
  Speaker 1: suggest abandoning proprietary AI.