Latest / USA Daily News / Nvidia CEO Jensen Huang Says “No Slowdown” for AI — The Global Race Is Just Beginning

Transcript

Nvidia CEO Jensen Huang Says “No Slowdown” for AI — The Global Race Is Just Beginning · USA Daily News

• 0:00
  Speaker 1: Artificial intelligence has reached a point where the debate is
• 0:03
  Speaker 1: no longer simply about what the technology can do, but
• 0:06
  Speaker 1: about how quickly it should be allowed to advance. For years,
• 0:10
  Speaker 1: the dominant story surrounding AI was one of acceleration, increasingly
• 0:15
  Speaker 1: capable models, rapidly expanding computing infrastructure, enormous private investment, and
• 0:22
  Speaker 1: a race among technology companies to introduce systems that could
• 0:25
  Speaker 1: write software, analyze information, generate images and video, assist researchers,
• 0:32
  Speaker 1: operate business processes, and increasingly perform tasks with limited human intervention.
• 0:38
  Speaker 1: But by September 2026, a significant disagreement had emerged inside
• 0:43
  Speaker 1: the technology industry over whether that momentum should continue at
• 0:47
  Speaker 1: full speed. On one side are executives and researchers who
• 0:51
  Speaker 1: argue that increasingly powerful AI systems require a slower and
• 0:56
  Speaker 1: more deliberate development process, giving companies and governments more time
• 1:01
  Speaker 1: to understand risks and establish safeguards. On the other side
• 1:05
  Speaker 1: are leaders who argue that slowing the technology could sacrifice
• 1:09
  Speaker 1: economic opportunities, scientific progress, national competitiveness, and potentially strategic advantages.
• 1:17
  Speaker 1: NVIDIA Chief Executive Jensen Huang has become one of the
• 1:20
  Speaker 1: most prominent voices on the acceleration side of that argument.
• 1:24
  Speaker 1: making his position particularly significant because NVIDIA sits at the
• 1:28
  Speaker 1: center of the computing infrastructure that powers much of the
• 1:32
  Speaker 1: modern AI industry. Huang's comments come at a moment when
• 1:36
  Speaker 1: the meaning of an AI slowdown itself is being heavily debated.
• 1:40
  Speaker 1: The issue is not necessarily whether companies should stop building
• 1:44
  Speaker 1: artificial intelligence altogether. Instead, the disagreement concerns the pace at
• 1:49
  Speaker 1: which the capabilities of the most advanced systems should continue
• 1:53
  Speaker 1: increasing and what safety measures should accompany that progress. Anthropic
• 1:58
  Speaker 1: chief executive Dario Amadei has publicly argued that the frontier
• 2:02
  Speaker 1: of AI development should be paced more carefully, while other
• 2:06
  Speaker 1: prominent technology leaders have expressed support for additional caution, testing,
• 2:11
  Speaker 1: security measures, and oversight. The calls gained attention after warnings
• 2:16
  Speaker 1: from people working within the AI industry raised concerns about
• 2:20
  Speaker 1: systems becoming more capable in ways that developers may not
• 2:24
  Speaker 1: fully anticipate or control. Reuters reported that several major AI
• 2:29
  Speaker 1: leaders publicly supported efforts to slow development in September, reflecting
• 2:33
  Speaker 1: a degree of concern that had become increasingly visible across
• 2:37
  Speaker 1: the industry. At the same time, other executives have challenged
• 2:41
  Speaker 1: that approach, arguing that technological development and.
• 2:45
  Speaker 2: Safety do not have to be mutually exclusive.
• 2:49
  Speaker 1: The resulting disagreement has turned what was once largely a
• 2:52
  Speaker 1: technical discussion among researchers into a much broader argument involving
• 2:57
  Speaker 1: business strategy, public policy, national security, and the future structure
• 3:03
  Speaker 1: of the global economy. Jensen Huang's position matters partly because
• 3:07
  Speaker 1: his company occupies an unusual place in this transformation. NVIDIA
• 3:12
  Speaker 1: does not primarily build the consumer-facing AI assistance that most
• 3:16
  Speaker 1: people interact with directly. Instead, the company has become a
• 3:21
  Speaker 1: critical supplier of the processors and accelerated computing systems used
• 3:25
  Speaker 1: to train and operate many of the most advanced AI models.
• 3:29
  Speaker 1: That means the expansion of artificial intelligence is closely connected
• 3:34
  Speaker 1: to demand for NVIDIA's technology. As companies build larger data
• 3:38
  Speaker 1: centers and invest in more computing capacity, they require enormous
• 3:43
  Speaker 1: quantities of advanced processors, memory, networking equipment, and supporting infrastructure.
• 3:49
  Speaker 1: In practical terms, the faster the AI industry expands its
• 3:54
  Speaker 1: computational ambitions, the greater the potential demand for the hardware
• 3:58
  Speaker 1: ecosystem NVIDIA helps provide. That relationship does not by itself
• 4:03
  Speaker 1: establish why Huang holds a particular position and the economic
• 4:07
  Speaker 1: interests of a company should not be treated as proof
• 4:10
  Speaker 1: of the motivations of an individual executive. But it does
• 4:14
  Speaker 1: explain why Huang is an important participant in any discussion
• 4:18
  Speaker 1: about the speed of AI development. His company is deeply
• 4:21
  Speaker 1: embedded in the infrastructure required to turn increasingly ambitious AI
• 4:26
  Speaker 1: plans into functioning systems. The significance of Huang's position became
• 4:31
  Speaker 1: especially visible in September 2026, when he was asked directly
• 4:35
  Speaker 1: whether artificial intelligence development should slow down. In remarks reported
• 4:40
  Speaker 1: by CBS News, Huang said companies should go as fast
• 4:44
  Speaker 1: as we can, while also distinguishing speed from the release
• 4:48
  Speaker 1: of technology that has not been adequately prepared for use.
• 4:52
  Speaker 1: His position has therefore not been presented simply as an
• 4:55
  Speaker 1: argument that safety should be ignored. Instead, Huang has emphasized
• 5:00
  Speaker 1: the idea that companies can continue moving quickly while taking
• 5:04
  Speaker 1: responsibility for the.
• 5:05
  Speaker 2: Products they release.
• 5:07
  Speaker 1: The argument for slowing artificial intelligence development begins with a
• 5:11
  Speaker 1: concern that is different from the usual criticism of technology.
• 5:15
  Speaker 1: the possibility that the systems being built are becoming capable
• 5:18
  Speaker 1: faster than the people responsible for developing them can reliably understand, test,
• 5:24
  Speaker 1: and control them. In September 2026, that concern moved into
• 5:30
  Speaker 1: the center of the public debate after anthropic CEO Dario
• 5:33
  Speaker 1: Amadei called for companies developing the most advanced AI systems
• 5:38
  Speaker 1: to deliberately moderate the rate at which model capabilities improve.
• 5:42
  Speaker 1: His argument was not that artificial intelligence should be abandoned,
• 5:46
  Speaker 1: nor that research should come to an absolute stop. Instead,
• 5:51
  Speaker 1: he argued that the industry should create more time between
• 5:54
  Speaker 1: major increases in capability, allowing safety measures, independent evaluations, and
• 6:00
  Speaker 1: governance systems to catch up. The distinction is important because
• 6:04
  Speaker 1: the debate is often portrayed as a simple conflict between.
• 6:08
  Speaker 2: Innovation and fear.
• 6:10
  Speaker 1: In reality, The people advocating caution generally describe their position
• 6:14
  Speaker 1: as an attempt to manage the risks created by increasingly
• 6:18
  Speaker 1: powerful systems while allowing useful development to continue. Their concern
• 6:23
  Speaker 1: is that the traditional technology model of building first, releasing quickly,
• 6:28
  Speaker 1: and improving through competition may become less appropriate as AI
• 6:32
  Speaker 1: systems gain the ability to perform more complex and autonomous tasks.
• 6:38
  Speaker 1: One of the central issues is the changing nature of
• 6:41
  Speaker 1: AI capabilities. Earlier generations of software typically followed clearly defined
• 6:46
  Speaker 1: instructions and operated within relatively narrow boundaries. Modern AI systems
• 6:52
  Speaker 1: can interpret natural language, generate code, reason through complex problems,
• 6:57
  Speaker 1: use external tools, interact with computer environments, and in some
• 7:02
  Speaker 1: cases carry out sequences of actions with limited supervision. As
• 7:06
  Speaker 1: these systems become more capable, the traditional distinction between a
• 7:11
  Speaker 1: tool that responds to a human and an agent that
• 7:14
  Speaker 1: can pursue a goal becomes less clear. Safety researchers are
• 7:18
  Speaker 1: particularly interested in what happens when AI is given access
• 7:22
  Speaker 1: to networks, files, software systems, financial tools, or other digital environments.
• 7:28
  Speaker 1: A chatbot producing an incorrect answer presents one category of problem.
• 7:33
  Speaker 1: An AI system capable of independently taking actions, finding vulnerabilities,
• 7:39
  Speaker 1: writing and executing code, or interacting with other systems creates
• 7:43
  Speaker 1: a different category of risk. Recent discussions within the AI
• 7:47
  Speaker 1: industry have focused on incidents and tests in which advanced
• 7:51
  Speaker 1: models demonstrated behaviors associated with hacking, surveillance, fraud, or unauthorized
• 7:58
  Speaker 1: interaction with systems. These events do not establish that AI
• 8:02
  Speaker 1: has become independently dangerous or uncontrollable, but they have strengthened
• 8:07
  Speaker 1: arguments that capability testing needs to become more rigorous as
• 8:11
  Speaker 1: systems become more autonomous.
• 8:13
  Speaker 2: Another concern involves.
• 8:15
  Speaker 1: The possibility that AI systems could become useful in ways
• 8:19
  Speaker 1: their creators did not originally anticipate. A highly capable model
• 8:23
  Speaker 1: can be used for legitimate purposes such as software development,
• 8:27
  Speaker 1: scientific research, logistics, education and medical discovery. But the same
• 8:33
  Speaker 1: underlying capabilities can potentially be redirected toward harmful objectives. Researchers
• 8:39
  Speaker 1: have raised concerns about cybersecurity because advanced models may reduce
• 8:43
  Speaker 1: the expertise required to discover vulnerabilities, create malicious code, automate
• 8:49
  Speaker 1: phishing campaigns or scale digital attacks. There are also concerns
• 8:53
  Speaker 1: about misinformation, automated fraud, manipulation, and the possibility of highly
• 8:59
  Speaker 1: personalized influence operations. These risks are not entirely new. Malicious
• 9:05
  Speaker 1: actors have used automation and digital tools for years. What
• 9:09
  Speaker 1: has changed is the possibility that increasingly capable AI could
• 9:13
  Speaker 1: make those activities cheaper, faster, more scalable, and accessible to
• 9:18
  Speaker 1: people with less technical expertise. That makes AI safety partly
• 9:23
  Speaker 1: a question of misuse. The argument for additional caution is
• 9:27
  Speaker 1: therefore not limited to a hypothetical future in which machines
• 9:31
  Speaker 1: somehow rebel against humans. Some of the most immediate concerns
• 9:35
  Speaker 1: involve humans using increasingly powerful AI systems against other humans.
• 9:40
  Speaker 1: Cybersecurity has become one of the most important areas in
• 9:44
  Speaker 1: this discussion because AI development is increasingly intertwined with the
• 9:48
  Speaker 1: digital infrastructure on which modern economies depend. Governments, banks, businesses, hospitals,
• 9:55
  Speaker 1: communications networks, and critical infrastructure all rely on computers. systems
• 10:00
  Speaker 1: that can contain valuable or sensitive information. If AI becomes
• 10:05
  Speaker 1: better at finding weaknesses than defenders are at protecting them,
• 10:08
  Speaker 1: the balance between offense and defense could change. There are
• 10:12
  Speaker 1: competing possibilities here. AI could strengthen cybersecurity by helping organizations
• 10:18
  Speaker 1: identify vulnerabilities, monitor networks, and respond to attacks more rapidly.
• 10:24
  Speaker 1: But it could also lower the cost of conducting attacks.
• 10:27
  Speaker 1: Jensen Huang's response to the growing calls for caution rests
• 10:31
  Speaker 1: on a different interpretation of the central AI problem. While
• 10:35
  Speaker 1: some researchers and technology executives have argued that the pace
• 10:39
  Speaker 1: of capability development itself should be reduced, Huang has consistently
• 10:43
  Speaker 1: drawn a line between moving quickly and releasing systems that
• 10:47
  Speaker 1: are not ready for real-world use. In his recent public comments,
• 10:52
  Speaker 1: he argued that artificial intelligence should advance as fast as
• 10:56
  Speaker 1: we can, while adding an important qualification, companies should never
• 11:01
  Speaker 1: move faster than they can responsibly manage. That distinction forms
• 11:06
  Speaker 1: the foundation of his position. Huang is not presenting safety
• 11:10
  Speaker 1: as an obstacle that should be ignored in pursuit of speed. Instead,
• 11:15
  Speaker 1: he describes safety as an engineering responsibility that should develop
• 11:19
  Speaker 1: alongside increasingly capable technology. At Driforce in San Francisco, he
• 11:25
  Speaker 1: emphasized that companies should not release products when they are
• 11:29
  Speaker 1: not confident in their functionality, capabilities, or safety. Later, in
• 11:34
  Speaker 1: an interview with CBS News, he again argued that rapid
• 11:38
  Speaker 1: progress and safety should not be treated as opposing objectives.
• 11:43
  Speaker 1: In his framing, the question is not whether the industry
• 11:47
  Speaker 1: should choose between innovation and safety, but whether it can
• 11:50
  Speaker 1: build the engineering processes needed to pursue both at the
• 11:53
  Speaker 1: same time. That argument begins with Hwang's rejection of what
• 11:57
  Speaker 1: he sees as an overly dramatic description of AI's future.
• 12:02
  Speaker 1: During his CBS interview, He pushed back against predictions that
• 12:06
  Speaker 1: humanity could be destroyed by AI within a decade, characterizing
• 12:11
  Speaker 1: those claims as unsupported by sufficient science or evidence. At
• 12:15
  Speaker 1: the same time, he acknowledged that the concerns behind the
• 12:19
  Speaker 1: warnings should not simply be dismissed. That distinction is significant.
• 12:25
  Speaker 1: Huang's position is not that AI safety concerns are imaginary. Rather,
• 12:30
  Speaker 1: he argues that they need to be evaluated through engineering, testing, measurement,
• 12:35
  Speaker 1: and practical safeguards rather than through what he regards as
• 12:39
  Speaker 1: highly dramatic predictions about an imminent collapse of human control.
• 12:43
  Speaker 1: For Huang, an industry dealing with complex computing systems should
• 12:48
  Speaker 1: be capable of testing those systems, identifying weaknesses, correcting problems,
• 12:53
  Speaker 1: and delaying release when a product fails those tests. This
• 12:57
  Speaker 1: reflects a broader technological philosophy in which difficult engineering problems
• 13:02
  Speaker 1: are approached through iterative improvement rather than by stopping the
• 13:05
  Speaker 1: underlying field. The assumption is that uncertainty creates a reason
• 13:10
  Speaker 1: to improve the process, not necessarily a reason to halt
• 13:13
  Speaker 1: the process itself.
• 13:15
  Speaker 2: The concept of safety as.
• 13:17
  Speaker 1: An engineering problem is particularly important to understanding Huang's argument.
• 13:22
  Speaker 1: Engineering disciplines routinely operate under conditions of uncertainty. Aircraft manufacturers
• 13:28
  Speaker 1: test components, software developers search for vulnerabilities, automobile companies conduct
• 13:34
  Speaker 1: crash tests, and industrial firms build redundant systems to reduce
• 13:39
  Speaker 1: the likelihood and impact of failures. Huang is effectively applying
• 13:44
  Speaker 1: a similar philosophy to artificial intelligence. Identify dangerous capabilities, test
• 13:50
  Speaker 1: systems before deployment, improve safeguards, and prevent products from being
• 13:55
  Speaker 1: released when the risks are not adequately understood. His comments
• 13:59
  Speaker 1: suggest that the industry should devote more resources to safety
• 14:03
  Speaker 1: as AI systems move from experimental laboratories into increasingly consequential
• 14:08
  Speaker 1: real-world applications. Nvidia itself has been promoting AI-based cybersecurity systems
• 14:14
  Speaker 1: that use automated agents for defense, illustrating Huang's broader view
• 14:19
  Speaker 1: that advanced AI capabilities can create new defensive tools as
• 14:23
  Speaker 1: well as new risks. At a September cybersecurity event, he
• 14:27
  Speaker 1: argued that attacks are increasingly automated and that defense therefore
• 14:32
  Speaker 1: also has to become more automated. That perspective reflects a
• 14:36
  Speaker 1: recurring theme in his argument. A rapidly changing technological environment
• 14:40
  Speaker 1: cannot necessarily be made safer simply by slowing the technology
• 14:45
  Speaker 1: because adversaries and competitors may continue progressing regardless. Huang's approach
• 14:51
  Speaker 1: also rests on the belief that companies have powerful incentives
• 14:55
  Speaker 1: to avoid deploying unsafe products. In his CBS interview, he
• 15:00
  Speaker 1: argued that businesses that release harmful technology can face legal consequences,
• 15:05
  Speaker 1: commercial losses, and reputational damage.
• 15:09
  Speaker 2: The logic is straightforward.
• 15:11
  Speaker 1: An AI system that creates serious harm could expose its
• 15:14
  Speaker 1: developer to lawsuits, regulatory action, customer departures, investor pressure, and
• 15:20
  Speaker 1: significant damage to its brand. From this perspective, Huang believes
• 15:25
  Speaker 1: the market itself creates an incentive for responsible development. Companies
• 15:30
  Speaker 1: do not merely have a technological obligation to build safe products,
• 15:34
  Speaker 1: they also have a business interest in doing so. He
• 15:37
  Speaker 1: therefore questions the argument that the only way to make
• 15:40
  Speaker 1: AI companies take safety seriously is through a broad slowdown
• 15:44
  Speaker 1: or an entirely new regulatory structure. Behind the argument over
• 15:49
  Speaker 1: how quickly artificial intelligence should advance is a powerful economic reality.
• 15:54
  Speaker 1: Building increasingly capable AI requires an extraordinary amount of computing infrastructure,
• 16:00
  Speaker 1: and NVIDIA has become one of the companies most closely associated.
• 16:04
  Speaker 2: With that infrastructure.
• 16:06
  Speaker 1: The modern AI economy depends on specialized processors that can
• 16:10
  Speaker 1: perform enormous numbers of calculations simultaneously, along with networking equipment, memory, storage,
• 16:18
  Speaker 1: data centers, cooling systems, electricity, and software. NVIDIA has built
• 16:24
  Speaker 1: a particularly strong position within that stack through its graphics
• 16:27
  Speaker 1: processing units, networking technologies, and software ecosystem. The scale of
• 16:33
  Speaker 1: that business has changed the economic meaning of the AI debate.
• 16:38
  Speaker 1: AI is no longer simply a research field in which
• 16:41
  Speaker 1: companies spend money to develop experimental models. It has become
• 16:45
  Speaker 1: an infrastructure industry requiring enormous and continuing capital investment. NVIDIA's
• 16:51
  Speaker 1: financial results illustrate just how quickly that transformation has taken place.
• 16:57
  Speaker 1: In its second quarter of fiscal 2027, covering the period
• 16:59
  Speaker 1: ended July 26, 2026, NVIDIA reported quarterly revenue of $ 96. 2 billion,
• 17:02
  Speaker 1: up 106% from a year earlier, while data center revenue
• 17:03
  Speaker 1: reached $ 89.
• 17:04
  Speaker 2: Billion, up 117% year over year.
• 17:20
  Speaker 1: Those numbers demonstrate that the commercial expansion of AI is
• 17:24
  Speaker 1: generating an exceptionally large demand for computing hardware. The importance
• 17:29
  Speaker 1: of those figures becomes clearer when NVIDIA's business is separated
• 17:34
  Speaker 1: into its different components. Although the company remains associated in
• 17:38
  Speaker 1: the public imagination with graphics technology and gaming, the center
• 17:43
  Speaker 1: of its financial transformation has become the data center. AI
• 17:47
  Speaker 1: developers need massive clusters of accelerators to train large models
• 17:51
  Speaker 1: and increasingly to run those models once they are deployed.
• 17:55
  Speaker 1: Cloud service providers purchase those systems and make computing capacity
• 18:00
  Speaker 1: available to AI companies, businesses, governments, and developers. Large technology
• 18:06
  Speaker 1: companies also build their own infrastructure, often operating enormous facilities
• 18:11
  Speaker 1: containing tens or hundreds of thousands of advanced processors. NVIDIA
• 18:16
  Speaker 1: therefore participates in the AI economy at an infrastructure level.
• 18:20
  Speaker 1: Its revenues can rise not only when a particular AI
• 18:23
  Speaker 1: application becomes popular with consumers, but when companies decide that
• 18:28
  Speaker 1: they need more computing capacity to support future applications. This
• 18:32
  Speaker 1: distinction is economically significant because it means the AI investment
• 18:37
  Speaker 1: cycle can continue even while the final uses of the
• 18:40
  Speaker 1: technology are still evolving. A company may not yet know
• 18:44
  Speaker 1: exactly which AI products will dominate the market, but it
• 18:48
  Speaker 1: can still spend billions of dollars preparing the computing infrastructure
• 18:51
  Speaker 1: that those products may eventually require. That infrastructure build-out has
• 18:56
  Speaker 1: expanded far beyond a handful of specialized AI laboratories. NVIDIA's
• 19:01
  Speaker 1: management said in its latest earnings call that demand was
• 19:05
  Speaker 1: coming from hyperscale cloud providers, AI laboratories, AI-native companies, enterprises,
• 19:12
  Speaker 1: industrial customers, and sovereign customers. The company also said cloud
• 19:17
  Speaker 1: industry backlog had exceeded $ 2 trillion, while capital expenditure by
• 19:22
  Speaker 1: the five largest hyperscalers was expected by NVIDIA to.
• 19:26
  Speaker 2: Approach $ 800 billion in 2026 and $ 1. 3 trillion in 2027.
• 19:27
  Speaker 1: These are company estimates and expectations rather than guarantees. but
• 19:38
  Speaker 1: they illustrate the scale of the investment cycle NVIDIA is
• 19:42
  Speaker 1: positioning itself to serve. The company's outlook is based on
• 19:46
  Speaker 1: the premise that organizations will continue adding computing capacity because
• 19:51
  Speaker 1: AI applications are creating new demand for processing power. Nvidia
• 19:55
  Speaker 1: has described the build-out as an infrastructure expansion involving multiple
• 20:00
  Speaker 1: generations of hardware and an increasingly diverse ecosystem of customers.
• 20:05
  Speaker 1: This creates an important economic connection between technological speed and
• 20:10
  Speaker 1: corporate spending. The more ambitious the AI industry becomes, the
• 20:14
  Speaker 1: more computational capacity it requires. Larger models can require larger
• 20:19
  Speaker 1: training runs. More users can require additional inference capacity. AI
• 20:25
  Speaker 1: agents that perform longer sequences of tasks can increase the
• 20:29
  Speaker 1: amount of computing consumed.
• 20:30
  Speaker 2: During each interaction.
• 20:32
  Speaker 1: Robotics and other forms of physical AI introduce another potential
• 20:36
  Speaker 1: source of demand because intelligent machines need computing both during
• 20:40
  Speaker 1: development and, in some cases, during operation. In that environment,
• 20:46
  Speaker 1: a slowdown in AI capability development could eventually affect the
• 20:50
  Speaker 1: infrastructure investment cycle. It would not necessarily cause an immediate
• 20:54
  Speaker 1: collapse in demand because companies have already committed money to
• 20:58
  Speaker 1: data centers and equipment and existing systems still need to
• 21:02
  Speaker 1: be expanded and replaced. But a sustained reduction in the
• 21:06
  Speaker 1: pace of AI development could alter how aggressively customers build
• 21:10
  Speaker 1: additional capacity. The debate over whether artificial intelligence should slow
• 21:15
  Speaker 1: down becomes considerably more complicated once it moves beyond technology
• 21:19
  Speaker 1: companies and into the realm of national strategy. Artificial intelligence
• 21:24
  Speaker 1: is increasingly treated by governments as a capability that can
• 21:28
  Speaker 1: influence economic productivity, scientific research, cybersecurity, military planning, intelligence analysis,
• 21:36
  Speaker 1: industrial automation, and the development of future technologies. That means
• 21:42
  Speaker 1: the speed at which advanced AI systems improve is not
• 21:45
  Speaker 1: viewed only through the lens of consumer products or corporate competition.
• 21:50
  Speaker 1: It is also being considered as part of a broader
• 21:52
  Speaker 1: contest over technological leadership. The United States and China are
• 21:57
  Speaker 1: at the center of that contest, with both countries investing
• 22:00
  Speaker 1: heavily in artificial intelligence while attempting to strengthen domestic computing capacity,
• 22:06
  Speaker 1: research expertise, semiconductor supply chains, and AI applications. Against that backdrop,
• 22:14
  Speaker 1: calls for slowing frontier AI development raise a difficult strategic question.
• 22:19
  Speaker 1: Could a deliberate reduction in the pace of American AI
• 22:22
  Speaker 1: development create an opening for competing countries to advance more quickly?
• 22:27
  Speaker 1: That question has been repeatedly raised by U.S. officials and
• 22:31
  Speaker 1: technology executives, although the scale of the actual strategic effect
• 22:36
  Speaker 1: remains uncertain. The issue is particularly relevant to Jensen Huang
• 22:40
  Speaker 1: because NVIDIA supplies much of the computing infrastructure used throughout
• 22:44
  Speaker 1: the AI ecosystem. while the company itself operates within a
• 22:48
  Speaker 1: global technology environment shaped by US-China competition, export restrictions, and
• 22:54
  Speaker 1: efforts to control.
• 22:55
  Speaker 2: The flow of advanced chips.
• 22:57
  Speaker 1: The basic geopolitical argument is built around the idea that
• 23:01
  Speaker 1: leadership and artificial intelligence can create advantages across multiple sectors
• 23:07
  Speaker 1: at once. A country that develops more capable AI systems
• 23:11
  Speaker 1: may be able to improve research productivity, automate portions of
• 23:15
  Speaker 1: industrial production, strengthen cyber defenses, accelerate drug discovery, optimize supply chains,
• 23:22
  Speaker 1: and increase the efficiency of government operations. Military organizations may
• 23:27
  Speaker 1: also seek AI for intelligence processing, autonomous systems, logistics, targeting support,
• 23:34
  Speaker 1: electronic warfare, and other applications. Not every anticipated use will materialize,
• 23:41
  Speaker 1: and many AI applications remain experimental. but governments have strong
• 23:46
  Speaker 1: incentives to avoid falling substantially behind in a technology that
• 23:50
  Speaker 1: could influence so many areas simultaneously. This creates a strategic
• 23:55
  Speaker 1: environment in which AI development resembles a competition for an
• 23:59
  Speaker 1: underlying platform rather than a single product category. Leadership can
• 24:03
  Speaker 1: potentially reinforce itself. Stronger research attracts investment. Investment expands computing infrastructure.
• 24:11
  Speaker 1: Infrastructure supports more experimentation. Successful applications generate commercial revenue, and
• 24:18
  Speaker 1: commercial success attracts more talent and capital. The United States
• 24:23
  Speaker 1: has historically held important advantages in frontier AI research and
• 24:27
  Speaker 1: private technology investment, but Chinese companies have made rapid progress,
• 24:32
  Speaker 1: making the question of relative speed increasingly significant. The Associated
• 24:38
  Speaker 1: Press reported in September that Chinese firms were closing parts
• 24:42
  Speaker 1: of the technology gap with U.S. developers, despite American restrictions
• 24:47
  Speaker 1: on advanced semiconductor technology. That competition is especially important because
• 24:52
  Speaker 1: artificial intelligence depends on more than software alone.
• 24:56
  Speaker 2: At the most.
• 24:57
  Speaker 1: Advanced level, AI capability is closely connected to access to
• 25:01
  Speaker 1: computing power, advanced processors, high-bandwidth memory, sophisticated networking, data center infrastructure, electricity,
• 25:11
  Speaker 1: and engineering talent. This creates an intersection between AI policy
• 25:16
  Speaker 1: and semiconductor policy.
• 25:18
  Speaker 2: The United States has attempted.
• 25:20
  Speaker 1: To restrict China's access to some of the most advanced
• 25:23
  Speaker 1: computing technologies on national security grounds, while China has simultaneously
• 25:28
  Speaker 1: pursued greater technological self-reliance. The objective on the American side
• 25:34
  Speaker 1: is not simply to prevent Chinese companies from obtaining a
• 25:37
  Speaker 1: particular generation of processor. It is part of a broader
• 25:41
  Speaker 1: effort to influence the technological ecosystem required to train and
• 25:45
  Speaker 1: operate highly capable AI systems. China, meanwhile, has invested heavily
• 25:50
  Speaker 1: in domestic semiconductor capabilities and AI infrastructure in an effort
• 25:55
  Speaker 1: to reduce dependence on foreign suppliers. The outcome of that
• 25:59
  Speaker 1: competition could affect not only the companies directly involved, but
• 26:03
  Speaker 1: also the countries and businesses that depend on global AI
• 26:07
  Speaker 1: supply chains. NVIDIA occupies a particularly sensitive position in that
• 26:11
  Speaker 1: system because its products are among the most important components
• 26:15
  Speaker 1: in the infrastructure used to train and run advanced AI.
• 26:19
  Speaker 1: The company has historically had major customers and commercial relationships
• 26:23
  Speaker 1: across the world, but U.S. export controls have limited the
• 26:27
  Speaker 1: sale of certain advanced AI processors to China. The question
• 26:31
  Speaker 1: of who should control artificial intelligence has become one of
• 26:35
  Speaker 1: the most consequential parts of the debate over how quickly
• 26:38
  Speaker 1: the technology should develop. As AI systems become more capable
• 26:42
  Speaker 1: and increasingly integrated into workplaces, financial services, education, communications, cybersecurity,
• 26:51
  Speaker 1: government operations, and critical infrastructure, Governments are being asked to
• 26:56
  Speaker 1: determine whether existing laws are sufficient or whether new rules
• 27:00
  Speaker 1: are required specifically for artificial intelligence. Technology companies, meanwhile, have
• 27:06
  Speaker 1: increasingly promoted internal safety programs, voluntary commitments, testing procedures, transparency measures,
• 27:14
  Speaker 1: and industry standards as alternatives or complements to government regulation.
• 27:19
  Speaker 1: The disagreement is closely connected to Jensen Huang's rejection of
• 27:23
  Speaker 1: calls for a general slowdown. Huang has argued that companies
• 27:27
  Speaker 1: should move rapidly but should not release systems they are
• 27:30
  Speaker 1: not confident are safe, while questioning whether broad new AI-specific
• 27:34
  Speaker 1: regulation is always necessary.
• 27:37
  Speaker 2: That approach places.
• 27:38
  Speaker 1: Substantial responsibility on the companies developing the technology. The alternative
• 27:44
  Speaker 1: approach gives governments a larger role in establishing mandatory standards
• 27:48
  Speaker 1: and determining which activities are acceptable before products reach the public.
• 27:53
  Speaker 1: The conflict between these models is not simply about regulation
• 27:57
  Speaker 1: versus no regulation. It concerns where responsibility should reside, how
• 28:02
  Speaker 1: risks should be measured, what powers regulators should have, and
• 28:06
  Speaker 1: whether rules can keep pace with a technology that changes
• 28:10
  Speaker 1: faster than traditional legislative processes. The United States has already
• 28:15
  Speaker 1: developed several different approaches to AI risk management, which illustrates
• 28:20
  Speaker 1: why the regulatory debate cannot be reduced to a single
• 28:23
  Speaker 1: national policy. One important example is the National Institute of
• 28:28
  Speaker 1: Standards and Technology's AI Risk Management Framework. NIST describes the
• 28:33
  Speaker 1: framework as a voluntary resource intended to help organizations manage
• 28:37
  Speaker 1: risks associated with artificial intelligence and improve trustworthiness across the design. development, deployment,
• 28:45
  Speaker 1: and evaluation of AI systems. The framework addresses areas including safety, security, resilience, transparency, explainability, privacy,
• 28:57
  Speaker 1: and harmful bias. NIST has continued developing the framework as
• 29:01
  Speaker 1: the technology evolves, and in April 2026, it released a
• 29:06
  Speaker 1: concept note for a critical infrastructure profile focused on trustworthy AI.
• 29:11
  Speaker 1: In August 2026, NIST also announced a public draft of
• 29:15
  Speaker 1: its TV Viathlon framework, designed to support testing, evaluation, verification,
• 29:22
  Speaker 1: and validation of AI systems, including generative and agentic systems.
• 29:27
  Speaker 1: These initiatives demonstrate one possible model of governance. The government
• 29:32
  Speaker 1: establishes technical guidance and evaluation tools, while leaving organizations considerable
• 29:38
  Speaker 1: flexibility in how they implement them. them. The attraction of
• 29:41
  Speaker 1: such a voluntary system is flexibility. AI companies operate across
• 29:46
  Speaker 1: many industries and deploy systems with very different risk profiles.
• 29:51
  Speaker 1: A model used to summarize low-stakes business documents does not
• 29:54
  Speaker 1: necessarily require the same level of scrutiny as an AI
• 29:58
  Speaker 1: system used in critical infrastructure, cybersecurity, financial decision-making, or highly
• 30:04
  Speaker 1: consequential government functions. A rigid set of rules applied equally
• 30:09
  Speaker 1: to every AI application could impose requirements that are unnecessary
• 30:13
  Speaker 1: in low-risk situations while potentially failing to address the specific
• 30:18
  Speaker 1: characteristics of high-risk systems. A framework such as NISTs allows
• 30:22
  Speaker 1: organizations to evaluate their own systems according to context. It
• 30:27
  Speaker 1: encourages companies to identify risks, measure them, document decisions, test systems,
• 30:33
  Speaker 1: and establish controls without requiring the government to prescribe every
• 30:38
  Speaker 1: technical detail. That model also allows standards to evolve more
• 30:42
  Speaker 1: quickly than legislation because technical frameworks can be revised by
• 30:47
  Speaker 1: agencies and experts without requiring a new act of Congress
• 30:51
  Speaker 1: each time a new capability emerges. But voluntary frameworks also
• 30:55
  Speaker 1: raise a fundamental question about enforcement. An organization can choose
• 31:00
  Speaker 1: to follow recommended practices, but a voluntary framework does not
• 31:04
  Speaker 1: automatically create the same legal obligations as a statute or regulation.
• 31:10
  Speaker 1: If a company decides that a particular safety procedure is
• 31:13
  Speaker 1: too expensive, too slow, or commercially inconvenient, a voluntary framework
• 31:19
  Speaker 1: may provide limited power to compel compliance. Supporters of stronger
• 31:24
  Speaker 1: regulation argue that this creates.
• 31:26
  Speaker 2: An uneven playing field.
• 31:28
  Speaker 1: A company that invests heavily in safety could face higher
• 31:31
  Speaker 1: costs and slower product development than a competitor that adopts
• 31:35
  Speaker 1: fewer safeguards. If customers primarily reward speed and functionality, the
• 31:41
  Speaker 1: more cautious company could be placed at a commercial disadvantage.
• 31:45
  Speaker 1: Mandatory rules can theoretically solve part of that problem by
• 31:49
  Speaker 1: establishing common minimum standards that apply to all participants. What
• 31:54
  Speaker 1: began as a disagreement over the appropriate pace of artificial
• 31:57
  Speaker 1: intelligence development has increasingly become a visible divide among some
• 32:02
  Speaker 1: of the technology industry's most influential companies and executives. The
• 32:07
  Speaker 1: remarkable feature of the debate is that many of the
• 32:09
  Speaker 1: people involved agree on the fundamental importance of AI and
• 32:13
  Speaker 1: acknowledge that increasingly capable systems present real risks. Their disagreement
• 32:19
  Speaker 1: is over what should happen in response. Anthropic CEO Dario
• 32:23
  Speaker 1: Amadei has argued that frontier AI development should be deliberately
• 32:27
  Speaker 1: paced so that safety systems.
• 32:29
  Speaker 2: And governance can keep up.
• 32:31
  Speaker 1: OpenAI CEO Sam Altman has expressed support for greater coordination
• 32:36
  Speaker 1: and stronger safeguards while stopping short of advocating a complete
• 32:40
  Speaker 1: halt to development. Elon Musk has also been associated with
• 32:44
  Speaker 1: calls for greater caution. On the other side, NVIDIA CEO
• 32:49
  Speaker 1: Jensen Huang has argued that AI should continue advancing as
• 32:53
  Speaker 1: rapidly as possible, provided companies do not release systems they
• 32:57
  Speaker 1: do not consider safe. What makes the dispute especially significant
• 33:01
  Speaker 1: is that these are not outsiders debating a technology they
• 33:05
  Speaker 1: barely understand. These are executives directing companies that are directly
• 33:09
  Speaker 1: involved in building, funding, deploying, or supplying the infrastructure for
• 33:14
  Speaker 1: advanced AI. Their differences therefore offer a window into the
• 33:17
  Speaker 1: strategic choices facing the industry itself. The current disagreement became
• 33:22
  Speaker 1: more prominent in September 2026, after Amadé published his proposal
• 33:28
  Speaker 1: to pace the frontier of AI development. His plan called
• 33:32
  Speaker 1: for third-party evaluators to receive permanent access to systems at
• 33:36
  Speaker 1: AI laboratories so they could assess compliance with safety measures,
• 33:41
  Speaker 1: investigate incidents, and evaluate model alignment during training. He also
• 33:46
  Speaker 1: proposed shared safety standards among leading AI companies and international
• 33:51
  Speaker 1: coordination around particularly serious risks. The significance of the proposal
• 33:56
  Speaker 1: was not simply that one executive was asking for caution.
• 34:00
  Speaker 1: It was that other major AI leaders broadly acknowledged the concern,
• 34:04
  Speaker 1: creating what the Associated Press described as a rare consensus
• 34:08
  Speaker 1: around the need to manage the speed of development more carefully.
• 34:12
  Speaker 1: Yet the consensus immediately encountered a problem. Agreeing that AI
• 34:17
  Speaker 1: safety matters is considerably easier than agreeing on how to
• 34:20
  Speaker 1: implement safeguards without undermining competition. Some executives favor internal or
• 34:26
  Speaker 1: industry-led mechanisms, while others see a larger role for government.
• 34:30
  Speaker 1: Some emphasize national competition, particularly with China, while others focus
• 34:36
  Speaker 1: more heavily on global coordination. The common ground, therefore, exists
• 34:40
  Speaker 1: mainly at the level of recognizing risk. The disagreements become
• 34:45
  Speaker 1: sharper when the conversation turns to rules, enforcement, timing, and
• 34:50
  Speaker 1: competitive strategy. Jensen Huang's position has made the division especially
• 34:55
  Speaker 1: clear because Nvidia represents a different part of the AI
• 34:58
  Speaker 1: ecosystem from Anthropic or OpenAI. Anthropic and OpenAI are primarily
• 35:04
  Speaker 1: known for developing and deploying advanced AI models and products,
• 35:09
  Speaker 1: while Nvidia supplies much of the computing infrastructure that allows
• 35:12
  Speaker 1: those models to be trained and operated. Huang's economic and
• 35:17
  Speaker 1: strategic perspective is therefore shaped by a different position in
• 35:20
  Speaker 1: the technology chain. He sees growing demand for computation as
• 35:25
  Speaker 1: a central part of the AI transformation and has argued
• 35:28
  Speaker 1: that the United States should continue advancing rapidly while maintaining safety.
• 35:33
  Speaker 2: In product deployment.
• 35:35
  Speaker 1: His message does not amount to an argument that AI
• 35:38
  Speaker 1: companies should ignore safety. Rather, it places the emphasis on
• 35:42
  Speaker 1: engineering and responsible release. At Dreeforce and in a subsequent
• 35:47
  Speaker 1: CBS interview, Huang rejected the idea that AI progress and
• 35:52
  Speaker 1: safety necessarily require a trade-off, saying the industry can move
• 35:56
  Speaker 1: quickly while ensuring products are ready before they are released.
• 36:01
  Speaker 1: That approach contrasts with Amadé' s argument that the development process
• 36:04
  Speaker 1: itself should be paced more deliberately. The difference is subtle
• 36:08
  Speaker 1: in one sense but fundamental in another. One side sees
• 36:12
  Speaker 1: rapid development combined with safeguards as workable, while the other
• 36:17
  Speaker 1: believes that at some stage speed itself becomes a risk factor.
• 36:20
  Speaker 2: That must be managed.
• 36:22
  Speaker 1: Sam Altman occupies an intermediate position that further demonstrates how
• 36:26
  Speaker 1: difficult it is to divide the industry into two simple camps.
• 36:31
  Speaker 1: OpenAI has continued to pursue increasingly capable systems while also
• 36:35
  Speaker 1: discussing safety, cybersecurity, and coordination among developers. Altman has backed
• 36:41
  Speaker 1: elements of Amadei's call for industry coordination, but has not
• 36:45
  Speaker 1: presented the objective as an indefinite halt to AI advancement.
• 36:49
  Speaker 1: The debate over how quickly artificial intelligence should advance ultimately
• 36:54
  Speaker 1: reaches far beyond technology companies, semiconductor manufacturers, and government regulators.
• 37:00
  Speaker 1: The most immediate question for billions of people is what
• 37:03
  Speaker 1: happens when increasingly capable AI systems become deeply embedded in
• 37:08
  Speaker 1: the ordinary functioning of the economy and society. The effects
• 37:12
  Speaker 1: are already beginning to appear in workplaces, education, customer service,
• 37:17
  Speaker 1: software development, media production, research, logistics, and countless other areas.
• 37:25
  Speaker 1: But continued acceleration could make those changes much broader and faster.
• 37:30
  Speaker 1: A future in which AI systems become significantly more capable
• 37:34
  Speaker 1: would not necessarily mean that human workers simply disappear. The
• 37:38
  Speaker 1: economic impact could take many different forms, including changes in
• 37:43
  Speaker 1: the tasks people perform, the skills employers value, the structure
• 37:47
  Speaker 1: of organizations, the speed at which businesses operate, and the
• 37:52
  Speaker 1: amount of output that can be produced by relatively small teams.
• 37:56
  Speaker 1: At the same time, The benefits of higher productivity could
• 37:59
  Speaker 1: be accompanied by difficult questions about wages, employment security, inequality,
• 38:05
  Speaker 1: access to opportunity, and who captures the financial gains created
• 38:10
  Speaker 1: by automation. This is one reason the argument over AI
• 38:14
  Speaker 1: acceleration has become a social issue as much as a
• 38:17
  Speaker 1: technical one. The faster the technology develops, the less time individuals, businesses,
• 38:24
  Speaker 1: educational institutions, and governments may have to adapt.
• 38:27
  Speaker 2: To the changes it creates.
• 38:30
  Speaker 1: One of the most important effects of accelerated AI development
• 38:33
  Speaker 1: could be a transformation in the nature of work itself.
• 38:37
  Speaker 1: Many jobs are composed of numerous individual tasks rather than
• 38:41
  Speaker 1: a single activity. A lawyer reviews documents, summarizes information, drafts correspondence,
• 38:48
  Speaker 1: analyzes evidence, and prepares arguments. A marketing professional researches audiences,
• 38:54
  Speaker 1: creates campaigns, analyzes performance, and produces content. A software engineer
• 39:01
  Speaker 1: writes code, tests it, investigates errors, documents systems, and communicates
• 39:06
  Speaker 1: with colleagues. An accountant processes financial information, prepares reports, checks records,
• 39:13
  Speaker 1: and identifies anomalies. An AI system does not necessarily have
• 39:18
  Speaker 1: to perform an entire occupation to have a major economic effect.
• 39:22
  Speaker 1: If it can reliably complete a significant portion of the
• 39:25
  Speaker 1: tasks that consume an employee's time, the role of the
• 39:29
  Speaker 1: worker can change even if the job title remains. Some
• 39:33
  Speaker 1: employees may become more productive because AI handles repetitive work
• 39:37
  Speaker 1: while humans focus on judgment, creativity, communication, or relationship building.
• 39:43
  Speaker 1: Other roles could see substantial reductions in demand if employers
• 39:47
  Speaker 1: discover that AI can perform enough of the underlying work
• 39:50
  Speaker 1: at a lower cost. The dividing line will not always
• 39:54
  Speaker 1: be obvious because technology can eliminate some tasks while simultaneously
• 39:59
  Speaker 1: creating new ones. This could produce a labor market that
• 40:03
  Speaker 1: looks very different from today's. In some industries, The first
• 40:07
  Speaker 1: response to AI adoption may be to give existing workers
• 40:10
  Speaker 1: access to better tools rather than reduce headcount. A small
• 40:14
  Speaker 1: team might suddenly be able to accomplish work that previously required.
• 40:19
  Speaker 2: A much larger organization.
• 40:21
  Speaker 1: A developer might use AI to produce and test software
• 40:24
  Speaker 1: more rapidly. A small business owner might generate marketing materials,
• 40:29
  Speaker 1: analyze sales data, answer routine customer inquiries, and automate administrative
• 40:35
  Speaker 1: tasks without hiring additional specialists. A researcher could use AI
• 40:40
  Speaker 1: to search through massive quantities of information or generate initial
• 40:44
  Speaker 1: hypotheses more efficiently. These changes could allow organizations to operate
• 40:49
  Speaker 1: with fewer layers of administration and potentially make entrepreneurship easier
• 40:54
  Speaker 1: by lowering the cost of launching a business. But the
• 40:57
  Speaker 1: same productivity improvement can create pressure on labor if companies
• 41:01
  Speaker 1: decide that fewer employees are necessary to produce the.
• 41:05
  Speaker 2: Same amount of output.
• 41:06
  Speaker 1: The economic outcome will depend partly on whether new demand
• 41:10
  Speaker 1: emerges as costs fall and whether businesses use productivity gains
• 41:14
  Speaker 1: to expand operations, reduce prices, increase investment, raise wages, or
• 41:20
  Speaker 1: reduce staffing. That question is closely connected to the classic
• 41:25
  Speaker 1: economic debate surrounding automation. Throughout history, technological advances have eliminated
• 41:31
  Speaker 1: certain forms of work while creating new industries and occupations.
• 41:36
  Speaker 1: Agricultural mechanization sharply reduced the amount of labor required in farming,
• 41:40
  Speaker 1: while industrialization transformed manufacturing and created entirely different categories of employment.
• 41:48
  Speaker 1: Computers eliminated many manual information processing tasks while generating vast
• 41:53
  Speaker 1: new industries in software, telecommunications, data services, and digital commerce.
• 41:59
  Speaker 1: Artificial intelligence could follow a similar pattern, but there is
• 42:03
  Speaker 1: a major difference in the range of work it can
• 42:06
  Speaker 1: potentially influence. Previous automation often focused on physical labor or
• 42:11
  Speaker 1: highly repetitive office tasks. At the center of the artificial
• 42:15
  Speaker 1: intelligence debate is a problem that neither rapid acceleration nor
• 42:19
  Speaker 1: deliberate restraint can completely eliminate. Nobody knows with certainty what
• 42:24
  Speaker 1: the most advanced AI systems will ultimately be capable of doing.
• 42:29
  Speaker 1: That uncertainty is what makes the disagreement between technology leaders
• 42:33
  Speaker 1: so difficult to resolve. Those calling for continued acceleration believe
• 42:38
  Speaker 1: the risks of slowing down may be underestimated, particularly if
• 42:42
  Speaker 1: advanced AI could produce major gains in science, productivity, cybersecurity, medicine, engineering,
• 42:50
  Speaker 1: and national competitiveness. Those urging greater caution fear that the
• 42:54
  Speaker 1: risks of moving too quickly may be even harder to
• 42:57
  Speaker 1: reverse once increasingly capable systems are deeply embedded in the
• 43:01
  Speaker 1: economy and digital infrastructure. The two sides are therefore contemplating
• 43:06
  Speaker 1: different forms of uncertainty. One worries about the consequences of
• 43:10
  Speaker 1: insufficient innovation and delayed progress. The other worries about the
• 43:15
  Speaker 1: consequences of insufficient control and delayed recognition of danger. Neither
• 43:20
  Speaker 1: argument can be settled simply by pointing to today's AI
• 43:23
  Speaker 1: systems because the debate is increasingly about what happens as
• 43:27
  Speaker 1: future systems become more capable, more autonomous, more widely deployed,
• 43:33
  Speaker 1: and potentially more difficult to evaluate. That leaves policymakers, businesses, researchers,
• 43:40
  Speaker 1: and the public facing a problem familiar in other high-impact technologies.
• 43:45
  Speaker 1: Decisions must be made before every consequence is known. The
• 43:49
  Speaker 1: first scenario is the one emphasized by the more cautious
• 43:53
  Speaker 1: voices in the industry, that AI capabilities could advance faster
• 43:57
  Speaker 1: than existing safeguards.
• 43:59
  Speaker 2: Under this scenario, the.
• 44:01
  Speaker 1: Danger would not necessarily come from a single dramatic event.
• 44:05
  Speaker 1: It could emerge through a series of smaller developments that
• 44:08
  Speaker 1: gradually increase the ability of AI systems to act independently,
• 44:13
  Speaker 1: access external tools, influence people, manipulate digital environments, or perform
• 44:18
  Speaker 1: sensitive tasks. A model that is initially used to generate
• 44:22
  Speaker 1: text or code may later gain the ability to browse
• 44:25
  Speaker 1: the internet, execute programs, interact with databases, manage software tools,
• 44:31
  Speaker 1: and coordinate multi-step operations. Each individual capability might appear manageable,
• 44:37
  Speaker 1: but the combination could create behaviors that are more difficult
• 44:41
  Speaker 1: to predict. The concern is therefore partly about complexity. As
• 44:45
  Speaker 1: systems become more capable, there may be more ways for
• 44:49
  Speaker 1: them to fail, more environments in which they can be deployed,
• 44:53
  Speaker 1: and more interactions between.
• 44:54
  Speaker 2: The model and the outside world.
• 44:57
  Speaker 1: A company can test a system extensively and still discover
• 45:00
  Speaker 1: unexpected behavior after deployment because real-world environments contain variables that
• 45:06
  Speaker 1: laboratory evaluations cannot reproduce perfectly. This problem becomes especially serious
• 45:12
  Speaker 1: when AI moves from answering questions to taking actions. Traditional
• 45:16
  Speaker 1: software often performs a defined function when instructed. An autonomous
• 45:21
  Speaker 1: or agentic AI system can potentially make decisions about how
• 45:25
  Speaker 1: to reach a goal, select tools, react to changing information,
• 45:29
  Speaker 1: and continue working through multiple steps. That creates a larger
• 45:34
  Speaker 1: safety surface. If the system misunderstands an objective, the consequences
• 45:39
  Speaker 1: may not remain confined to a single incorrect response. The
• 45:43
  Speaker 1: system could potentially continue acting on the misunderstanding until a
• 45:48
  Speaker 1: human intervenes. In a low-risk environment, that may mean an
• 45:52
  Speaker 1: incorrect spreadsheet or a wasted afternoon. In a high-risk environment,
• 45:57
  Speaker 1: it could mean unauthorized access to information, a flawed financial transaction,
• 46:03
  Speaker 1: a security failure, or an operational disruption. The issue is
• 46:07
  Speaker 1: not that current systems are secretly operating without controls. The
• 46:11
  Speaker 1: issue is that increasingly autonomous systems require new methods for
• 46:16
  Speaker 1: supervision because conventional software testing may not capture all of
• 46:20
  Speaker 1: the possible interactions between an AI model and the environment
• 46:24
  Speaker 1: around it. Cybersecurity offers one of the clearest examples of
• 46:28
  Speaker 1: this dual risk. Advanced AI can help defenders identify suspicious activity,
• 46:34
  Speaker 1: summarize enormous quantities of technical data, detect vulnerabilities, and respond
• 46:39
  Speaker 1: to threats rapidly. But a system capable of producing sophisticated
• 46:44
  Speaker 1: defensive analysis may also assist someone attempting to attack a network.
• 46:49
  Speaker 1: If future models become better at coding, reasoning, and discovering weaknesses,
• 46:54
  Speaker 1: the amount of specialized knowledge required to conduct certain cyber
• 46:58
  Speaker 1: operations could decline. This could increase the number of people
• 47:02
  Speaker 1: capable of launching sophisticated attacks. At the same time, defenders
• 47:07
  Speaker 1: may gain similar tools, creating an arms race between automated
• 47:12
  Speaker 1: offense and automated defense. The danger is that the speed
• 47:16
  Speaker 1: of machine-generated operations could eventually become much faster than human
• 47:20
  Speaker 1: organizations are accustomed to handling. The debate over Jensen Huang's
• 47:25
  Speaker 1: rejection of calls to slow artificial intelligence ultimately leads to
• 47:29
  Speaker 1: a question much larger than the disagreement between individual technology executives.
• 47:34
  Speaker 1: What kind of system will govern AI as its capabilities
• 47:37
  Speaker 1: continue to expand? The answer is unlikely to be determined
• 47:41
  Speaker 1: by one company, one government, or one technological breakthrough. Instead,
• 47:47
  Speaker 1: the future will probably emerge from a combination of corporate decisions,
• 47:51
  Speaker 1: national policies, technical standards, international negotiations, market incentives, and public pressure.
• 47:59
  Speaker 1: The competing arguments that have emerged across the AI industry
• 48:03
  Speaker 1: point towards several possible directions. One is continued acceleration, with
• 48:08
  Speaker 1: companies rapidly improving AI capabilities while relying heavily on engineering
• 48:14
  Speaker 1: safeguards and responsible deployment. Another is a more regulated model
• 48:19
  Speaker 1: in which governments establish increasingly detailed requirements for testing, transparency, cybersecurity,
• 48:27
  Speaker 1: risk management, and accountability. A third possibility combines the two,
• 48:32
  Speaker 1: rapid technological development accompanied by stronger controls around the most
• 48:37
  Speaker 1: powerful systems and the highest-risk applications.
• 48:41
  Speaker 2: The debate is.
• 48:42
  Speaker 1: Therefore moving away from the simple question of whether AI
• 48:45
  Speaker 1: should advance and toward the much harder question of how
• 48:48
  Speaker 1: advancement can be managed. As the technology becomes more deeply
• 48:53
  Speaker 1: integrated into the economy, governments and businesses will have to
• 48:57
  Speaker 1: decide which risks can be handled through voluntary measures which
• 49:01
  Speaker 1: require enforceable rules and which may demand international cooperation. The
• 49:07
  Speaker 1: first major decision concerns the role of speed itself. Jensen
• 49:11
  Speaker 1: Huang's position rests on the belief that technological acceleration should
• 49:16
  Speaker 1: not automatically be treated as a safety problem. In his view,
• 49:20
  Speaker 1: companies should continue developing increasingly capable systems while ensuring that
• 49:25
  Speaker 1: products are tested and safe before they are released. That
• 49:29
  Speaker 1: philosophy could become a dominant model if AI companies successfully
• 49:33
  Speaker 1: demonstrate that technical safeguards can evolve as quickly as the
• 49:37
  Speaker 1: capabilities they are designed to control. Under such a system,
• 49:41
  Speaker 1: development would remain highly competitive, with companies racing to improve models, infrastructure, robotics,
• 49:49
  Speaker 1: software agents, and specialized applications. Safety would become an engineering
• 49:54
  Speaker 1: layer built into the development process rather than a broad
• 49:57
  Speaker 1: reason to reduce the pace of research. Companies could create
• 50:01
  Speaker 1: increasingly sophisticated evaluations, red team programs, access controls, monitoring systems,
• 50:08
  Speaker 1: and automated defenses. The critical assumption would be that the
• 50:12
  Speaker 1: industry can learn quickly enough to correct problems before they
• 50:15
  Speaker 1: become irreversible. Whether that assumption holds will depend on the
• 50:20
  Speaker 1: ability of researchers and engineers to identify emerging risks before
• 50:24
  Speaker 1: those risks become widespread in real-world systems. A second path
• 50:29
  Speaker 1: would place greater emphasis on deliberate pacing at the frontier
• 50:33
  Speaker 1: of AI. Under this model, The most advanced laboratories would
• 50:37
  Speaker 1: continue conducting research but would agree to additional conditions before
• 50:42
  Speaker 1: pushing certain capabilities into broad deployment. The process could include
• 50:47
  Speaker 1: independent evaluations, additional cybersecurity testing, stronger monitoring requirements, and more
• 50:53
  Speaker 1: extensive assessment of systems that demonstrate unusually powerful autonomous behavior.
• 50:59
  Speaker 1: Such an approach would not necessarily mean stopping AI progress. Instead,
• 51:04
  Speaker 1: it would create additional checkpoints between the creation of a
• 51:07
  Speaker 1: capability and its release. The objective would be to ensure
• 51:11
  Speaker 1: that safety procedures have time to develop before each successive
• 51:15
  Speaker 1: generation of systems becomes significantly more powerful. This approach reflects
• 51:20
  Speaker 1: the concern that technological progress can sometimes move more quickly
• 51:24
  Speaker 1: than institutions can adapt. It also responds to the argument
• 51:28
  Speaker 1: that once a powerful system is widely distributed and controlling
• 51:32
  Speaker 1: its downstream use can become substantially more difficult. The European
• 51:37
  Speaker 1: Union already provides an example of how this second philosophy
• 51:41
  Speaker 1: can translate into formal governance. The EU AI Act entered
• 51:45
  Speaker 1: a significant new enforcement phase on August 2, 2026, when
• 51:50
  Speaker 1: enforcement powers began applying to several provisions covering prohibited AI practices,
• 51:56
  Speaker 1: transparency requirements, and obligations for providers of general-purpose AI models.
• 52:03
  Speaker 1: The framework also establishes additional obligations for systems involving systemic risks,
• 52:08
  Speaker 1: while some rules for high-risk AI applications are scheduled to
• 52:12
  Speaker 1: take effect later. The European approach demonstrates that governments can
• 52:17
  Speaker 1: regulate AI through a risk-based system rather than attempting to
• 52:20
  Speaker 1: treat every AI application identically. It also demonstrates that regulation
• 52:25
  Speaker 1: can be implemented progressively, with different obligations arriving at different
• 52:30
  Speaker 1: stages as institutions and companies prepare for them.