Will it take a disaster to force AI regulation? | The Economist

Will it take a disaster to force AI regulation? | The Economist — The Economist

The Economist

AI-assisted summary of Will it take a disaster to force AI regulation? | The Economist — original source. Check the original for full context.

Source published:

Key ideas

  • Substantial U.S. federal AI regulation is unlikely to materialize until a catastrophic 'Chernobyl of AI' occurs, involving severe financial fallout or loss of human life. (0:00)
  • Chinese AI labs trail U.S. frontier models by only three to four months, matching dangerous offensive cybersecurity capabilities through open-source releases like the GLM series. (1:12)
  • A de facto federal licensing regime is already underway, requiring leading frontier developers like OpenAI and Anthropic to seek weekly executive approvals for top-tier model access. (5:18)
  • The rapid pace of AI breakthroughs risks triggering an irrational, heavy-handed regulatory panic rather than the measured, decade-long framework evolutions seen in aviation and automotive industries. (6:27)

Detailed notes

## Background Published by The Economist on YouTube on October 9, 2026, this panel discussion is hosted by deputy editor Edward Carr and editor-in-chief Zanny Minton Beddoes. The episode addresses how growing public anxiety, rogue agent incidents, and geopolitical competition with China may force Donald Trump's administration to pivot from a hands-off stance to federal AI regulation.

## TL;DR The panel examines whether meaningful federal regulation in the United States will only materialize after an 'AI Chernobyl'—a major financial collapse or loss of life. Experts highlight that China enforces its own strict domestic AI rules while rapidly trailing U.S. frontier capabilities by only a few months, fueling an escalatory capability trap where labs build stronger models to defend against existing risks. Meanwhile, backdoor federal licensing approvals and a fragmented patchwork of state laws in California and New York are filling the federal vacuum. The discussion warns that delaying proactive oversight risks panic-driven, heavy-handed overregulation once a crisis hits, exacerbated by deep public skepticism toward Washington's competence.

## Atmosphere & style The discussion carries a sober, analytically sharp tone characteristic of high-level geopolitical and economic journalism. Rather than indulging in apocalyptic panic or blind tech optimism, the hosts and experts converse with measured realism, dry wit, and intellectual rigor. Watching it feels like sitting in on an urgent boardroom briefing, where the exhilarating pace of technological breakthroughs is tempered by palpable anxiety over political paralysis and the ticking clock of the next 12 to 18 months.

## Bulletin - [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) Substantial U.S. federal AI regulation is unlikely to materialize until a catastrophic 'Chernobyl of AI' occurs, involving severe financial fallout or loss of human life. - [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) Chinese AI labs trail U.S. frontier models by only three to four months, matching dangerous offensive cybersecurity capabilities through open-source releases like the GLM series. - [05:18](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=318s) A de facto federal licensing regime is already underway, requiring leading frontier developers like OpenAI and Anthropic to seek weekly executive approvals for top-tier model access. - [06:27](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=387s) The rapid pace of AI breakthroughs risks triggering an irrational, heavy-handed regulatory panic rather than the measured, decade-long framework evolutions seen in aviation and automotive industries.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) Federal intervention in AI is largely stalled, with experts warning that only a catastrophic event resulting in severe financial loss or casualties will compel meaningful U.S. action. - [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) China is not an unregulated market; Beijing enforces strict domestic AI rules to prevent risks to regime stability, opening a window for bilateral safety cooperation. - [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) A light-touch regulatory regime could recalibrate incentives toward safety without forfeiting America's competitive advantage over foreign rivals. - [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) China operates as an efficient fast follower, trailing the frontier by just 3 to 4 months and matching capabilities in offensive cybersecurity with models like GLM-4.3 and GLM-5.3. - [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) Halting U.S. frontier development does not halt danger, because competitors will continue closing the capability gap toward the current baseline. - [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) AI development is caught in an escalation trap comparable to 'swallowing a spider to catch a fly,' where stronger models are built solely to defend against vulnerabilities in slightly weaker systems. - [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) Escaping the technological escalation trap requires verifiable, costly commitments to slow down development in controlled increments. - [03:38](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=218s) Historical safety regimes in aviation, pharmaceuticals, and automobiles were built iteratively through trial and error following specific harms, not enacted all at once. - [05:18](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=318s) Leading developers like OpenAI and Anthropic are already subject to stealth federal oversight, submitting to weekly executive approvals to grant access to top-tier models. - [05:18](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=318s) States such as California, New York, and Rhode Island are passing their own AI liability legislation, creating a fragmented and uncoordinated patchwork of compliance mandates. - [06:27](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=387s) Because frontier AI evolves in compressed cycles of weeks rather than decades, a post-crisis political response is prone to heavy-handed, counterproductive overcorrection. - [06:27](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=387s) The next 12 to 18 months are a critical testing window to implement proactive safety measures before an inevitable disaster precipitates political panic. - [07:41](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=461s) Unlike other advanced democracies where public trust in tech oversight remains intact, the American electorate exhibits pervasive skepticism regarding Washington's competence to manage AI.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) The "Chernobyl of AI" and U.S.-China Strategic Parity

The discussion opens with the sobering premise that substantial political action on artificial intelligence will likely only follow a major disaster—shorthand for what is described as "the Chernobyl of AI." Meaningful federal intervention will probably require either catastrophic financial damage or significant loss of human life to shock the political establishment into action. Nevertheless, panelists argue that an outright halt to AI development is neither sensible nor necessary. Instead, the United States could recalibrate its posture toward safety without squandering its competitive technological edge.

Furthermore, the notion that China is an unregulated wild west of AI advancement is a misconception. Beijing enforces strict domestic AI regulations; the authoritarian regime has no interest in precipitating existential risks that threaten regime stability. This shared baseline creates a realistic opportunity for modest bilateral cooperation down the road. In the interim, Washington can implement a targeted, "light-touch regulatory regime that just tips the scale a little bit more towards safety" while preserving American leadership.

[01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) Catch-Up Dynamics and the "Swallowing a Spider" Dilemma

When probed on whether China's labs rely on distilling work from frontier U.S. models or whether they could pull ahead if America slows down, the panel highlights a critical nuance between charting entirely new technological territory and maintaining a lead. China has repeatedly demonstrated the ability to rapidly match the technological frontier. With a lag of only three to four months, Chinese labs have unveiled high-tier systems—exemplified by open-source releases such as the GLM-4.3/GLM-5.3 models—that rival capabilities reached by American companies like Anthropic in early 2024, particularly in offensive cybersecurity operations.

```mermaid flowchart TD accTitle: The Frontier Catch-Up and Capability Trap accDescr: Diagram illustrating how China catches up to the US frontier, leading to escalation risks and arms race dynamics. subgraph US["United States Frontier"] A["Frontier AI Models<br/>(Withheld from public)"] B["Offensive Capabilities<br/>(e.g., Cybersecurity harms)"] end

subgraph CN["Chinese Fast-Follower Catch-up"] C["Open-Source Frontier<br/>(e.g., GLM series)"] D["3–4 Month Trailing Gap Closes"] end

subgraph Trap["The Escalation Trap"] E["'Swallowing a spider to catch a fly':<br/>Building stronger AI to defend against weaker AI"] F["Risk of unchecked acceleration<br/>and recursive self-improvement"] end

A --> B A -.->|"Distillation & replication"| C C --> D D --> E B --> E E --> F ```

Because the frontier itself is already dangerous, U.S. labs routinely withhold their most capable models from the general public, releasing them only to trusted partners. Pausing development does not freeze the danger, because Chinese labs will catch up to the existing American baseline regardless. This exposes an accelerationist catch-22: researchers argue that one must build stronger AI to defend against the vulnerabilities of the slightly weaker AI already deployed. One speaker dryly compares this escalation cycle to "swallowing a spider to catch a fly." Escaping this trap requires deliberately slowing down in increments and proving those reductions via verifiable, "costly commitments."

[03:38](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=218s) Incremental Regulation and the Limits of System Tolerance

Historical precedent suggests that effective regulatory regimes are rarely created wholesale overnight. Whether examining the advent of the automobile, aviation, or pharmaceutical breakthroughs, safety structures emerged iteratively in direct response to operational failures and harms. Societies repeatedly engaged in trial and error, course-correcting whenever initial rules resulted in overreach.

Fatalism about an inevitable "10% chance of human extinction" overlooks human agency and historical adaptability. While an impactful catalyst—such as severe market turmoil or fatal incidents—will likely be necessary to shake lawmakers out of complacency, regulatory systems possess the mechanisms to adjust without completely breaking momentum.

[05:18](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=318s) Stealth Federal Licensing and the State Legislative Patchwork

Refuting the idea that America faces a binary choice between total deregulation and draconian mandates, the panel emphasizes that backdoor regulations are already taking effect:

- **De Facto Federal Oversight:** The current presidential administration already enforces a shadow licensing protocol. Leading frontier developers such as OpenAI and Anthropic are required to seek regular, weekly executive branch approvals before granting certain users access to their top-tier models. - **State-Level Legislative Momentum:** Individual states refuse to wait for gridlocked federal lawmakers. California and New York have enacted binding AI legislation, while states like Rhode Island are advancing dedicated algorithmic liability laws. - **The Threat of a Fractured Framework:** In the absence of federal preemption, technology firms face an uncoordinated, state-by-state patchwork of mandates that creates compliance complexity across jurisdictions.

[06:27](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=387s) The Risk of Post-Catastrophe Overcorrection

A major danger of relying on an "AI Chernobyl" to trigger federal intervention is the high likelihood of an irrational overreaction. While gradual responses to automobile and aviation disasters eventually delivered robust safety frameworks, those industries developed over decades.

By contrast, frontier AI capabilities evolve in leaps every few weeks, creating an environment that is simultaneously exhilarating and destabilizing. When an inevitable disaster strikes within this volatile, compressed timeline, the political impulse is far more likely to produce heavy-handed, counterproductive crackdowns than reasoned oversight. The coming 12 to 18 months will prove critical for testing whether policy can adapt before a catastrophe forces a panic response.

[07:41](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=461s) Democratic Competence and the Erosion of Public Confidence

The segment closes with a sharp critique of the political climate in the United States. For over a decade, American political figures have frequently treated high-stakes technical problems with performative dysfunction rather than measured governance.

Public opinion data reveals a stark divergence between the U.S. and its peers: citizens across other advanced democracies register notably higher confidence in their governments' competence to manage artificial intelligence risks. In contrast, the American public shows deep, pervasive skepticism toward Washington's ability to handle the challenge. The mounting anti-AI sentiment and cratering approval reflect an electorate that recognizes the high stakes and doubts its leaders possess the seriousness required to safeguard the country.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) The Chernobyl of AI

> [00:00](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=0s) light-touch regulatory regime that just tips the scale a little bit more towards safety

> [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) swallowing a spider to catch a fly

> [01:12](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=72s) costly commitments

> [03:38](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=218s) 10% chance of human extinction

> [05:18](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=318s) How American states are taking regulation into their own hands

> [06:27](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=387s) What’s the risk of over-regulation?

> [07:34](https://www.youtube.com/watch?v=B8oXVMYXlfM&t=454s) Are American politicians up to the task?

## Mentioned Donald Trump, Edward Carr, Zanny Minton Beddoes, OpenAI, Anthropic, GLM-4.3, GLM-5.3, The Economist, China, United States, California, New York, Rhode Island

## Video description Donald Trump, until recently, has taken a hands-off approach to regulating artificial intelligence. The president has favoured boosting the stock market and ensuring that America beats China in the AI race.

But that strategy is becoming unsustainable. Americans are increasingly anxious about everything from data centres to job security, while incidents of agents going rogue are warnings for the catastrophes that could lie ahead. With anti-AI sentiment growing, and the midterms fast approaching, the Trump administration may have to reassess how AI should be regulated.

Edward Carr, The Economist’s deputy editor, and Zanny Minton Beddoes, editor-in-chief, are joined by a panel of experts to discuss how to strike the right balance between safety and speed.

#AI #China #USA #Technology

00:00 - “The Chernobyl of AI” 00:16 - China v America: would embracing regulation mean falling behind? 03:38 - How to get regulation just right 05:18 - How American states are taking regulation into their own hands 06:27 - What’s the risk of over-regulation? 07:34 - Are American politicians up to the task?

Watch the full show: http://bit.ly/3TRiwiZ Sign up to the Insider newsletter: https://econ.st/4nOyzIb

Subscribe to The Economist: https://www.economist.com/subscribe Download our app: https://www.economist.com/get-the-app

Follow The Economist on Instagram: https://www.instagram.com/theeconomist/ Follow The Economist on Facebook: https://www.facebook.com/TheEconomist Follow The Economist on X: https://x.com/TheEconomist

[Source](https://www.youtube.com/watch?v=B8oXVMYXlfM)

中文简报

是否只有灾难降临才能倒逼AI监管？| 经济学人 专家组探讨了美国是否只有在发生类似“AI切尔诺贝利”的重大金融崩溃或人员伤亡事件后，才会出台有实质意义的联邦监管。专家指出，中国在严格执行国内AI法规的同时，其前沿能力仅落后美国数月，这加剧了一种能力升级陷阱，即各实验室不得不构建更强大的模型来防范现有风险。与此同时，非公开的联邦许可审批以及加州和纽约州等零散的州级法律正在填补联邦层面的空白。讨论警告称，推迟前瞻性监管可能导致危机一旦爆发便引发恐慌性、严苛的过度监管，而公众对华盛顿治理能力的深层怀疑则进一步加剧了这一风险。 在发生涉及严重金融损失或人员伤亡的灾难性“AI切尔诺贝利”事件之前，美国不大可能出台实质性的联邦AI监管措施。 (0:00) 中国AI实验室仅落后美国前沿模型三到四个月，并通过GLM系列等开源发布拥有了旗鼓相当的危险进攻性网络安全能力。 (1:12) 一种事实上的联邦许可机制已经启动，要求OpenAI和Anthropic等领先的前沿开发者就最高阶模型的访问权每周寻求行政批准。 (5:18) AI技术突破的迅猛步伐有可能引发非理性、严苛的监管恐慌，而无法像航空和汽车行业那样历经数十年演进出审慎适度的监管框架。 (6:27)

Reading summary

Panelists argue that significant federal AI regulation in the United States will likely be catalyzed only by a major disaster, yet relying on a crisis risks triggering heavy-handed political overcorrection rather than measured oversight. Catastrophic Catalysts and Regulatory Incrementalism (0:00) Meaningful political action on AI is expected to require a high-impact catalyst, characterized as an "AI Chernobyl" involving major financial harm or loss of human life. However, panelists reject both total development halts and fatalistic assumptions of inevitable human extinction. Drawing parallels to aviation, automotive, and pharmaceutical histories, they contend that safety regimes traditionally develop through iterative trial and error following operational harms, allowing societies to recalibrate toward safety without forfeiting progress. Chinese Parity and the Defensive Escalation Trap (1:12) Addressing the geopolitical landscape, the discussion dispels the belief that China operates an unregulated AI sector, noting that Beijing enforces strict domestic rules to safeguard regime stability. Chinese labs have demonstrated fast-follower capability, unveiling open-source systems such as the GLM-4.3 and GLM-5.3 models within three to four months of American frontier milestones—matching early-2024 capabilities from firms like Anthropic in offensive cybersecurity. Because Chinese developers can quickly match existing baselines, simply pausing frontier research does not eliminate operational dangers. This dynamic fuels an arms-race dilemma likened to "swallowing a spider to catch a fly," in which developers argue they must build more capable systems to defend against slightly weaker models. Escaping this dynamic would require gradual, verifiable, and costly commitments to slow development, alongside potential bilateral U.S.-China safety dialogues. De Facto Licensing and Emerging State Patchworks (5:18) Contrary to the perception that American AI remains entirely unregulated, informal and localized oversight is already expanding. At the federal level, the presidential administration enforces an implicit shadow licensing process, requiring frontier developers such as OpenAI and Anthropic to seek weekly executive branch approvals before granting certain users access to top-tier models. Concurrently, gridlock in Congress has driven state-level intervention, with California and New York enacting binding legislation and Rhode Island pursuing algorithmic liability rules, creating a fragmented regulatory patchwork. Compression of Timelines and Political Overreaction Risks (6:27) While past industries had decades to develop governance frameworks, frontier AI capabilities evolve every few weeks. Panelists caution that this compressed timeline increases the risk that an inevitable catastrophe will provoke a hasty, heavy-handed crackdown rather than deliberate policy. The subsequent 12 to 18 months are framed as a crucial window to determine whether lawmakers can institute preventative, light-touch safety standards before a crisis forces an emergency response. The discussion underscores that while other advanced democracies enjoy relatively higher public trust in their tech governance, the American public remains deeply skeptical of Washington's technical competence. The central unresolved question is whether U.S. policymakers can overcome political dysfunction to establish balanced safeguards before a disruptive disaster forces reactive legislation.

Source quotes

[0:00] The Chernobyl of AI [0:00] light-touch regulatory regime that just tips the scale a little bit more towards safety [1:12] swallowing a spider to catch a fly [1:12] costly commitments [3:38] 10% chance of human extinction [5:18] How American states are taking regulation into their own hands [6:27] What’s the risk of over-regulation? [7:34] Are American politicians up to the task?

Original transcript

[0:00] There will only be action, significant [0:01] action, after something biggish has [0:04] happened. Either a very serious [0:05] financial loss or a significant loss of [0:07] life. Shorthand, the Chernobyl of AI. [0:09] Something will have to happen to change [0:11] the political dynamic, but it's possible [0:13] to find a path through this. I don't [0:17] think it makes sense to kind of stop all [0:19] AI development completely, but I do [0:21] think that you can focus more on safety [0:24] in the US and have a a sort of shift in [0:26] the balance between forging ahead [0:28] without any focus on safety, and you can [0:30] do that without necessarily losing the [0:33] US edge. And secondly, it's not like [0:36] there's no regulation in China. There is [0:38] a lot of regulation around AI in China. [0:40] They are very concerned about certain [0:42] elements of AI, and I don't I can't [0:43] believe it's [0:45] that the Chinese authoritarian Chinese [0:47] regime thinks it's in their interest to [0:48] have truly existential risks. So, I'm [0:51] probably more optimistic that at some [0:53] point down the road there is going to be [0:55] some relatively modest ability to for [0:59] these two countries to work together. [1:00] But even before that, I think you can [1:01] design something [1:03] that is a light-touch regulatory regime [1:06] that just tips the scale a little bit [1:08] more towards safety, that doesn't [1:10] compromise the US position. [1:12] >> One one technical question I I need an [1:15] answer to, Alex, is to what extent if [1:18] America slows down a bit, it's kind of [1:20] within China's labs' ability to forge [1:24] ahead, and to what extent they actually [1:27] depend on distilling [1:30] the the work that the US labs are doing. [1:32] So, actually it's easier for the US to [1:34] keep its lead than you might think. [1:36] >> I think there's a difference between [1:38] America keeping its lead and America [1:41] forging ahead into, you know, [1:44] whole new areas that haven't yet been [1:46] broken. The problem if America pauses is [1:49] that at the very least we should expect [1:51] using the approaches that China's [1:52] already taken for it to catch up to the [1:55] frontier. And that three or four months [1:58] difference feels like it's been quite [2:00] important over the last three or four [2:02] months. We've seen perhaps only very [2:05] recently the release of what we might [2:07] say are mythos class models from China. [2:09] The open source GLM 5.3 system from a [2:12] Chinese open source lab maybe just [2:15] reaching the level that Anthropic [2:17] arrived at in March this year in terms [2:19] of the potential for damage when applied [2:22] to [2:23] cyber security domains. That means if [2:26] America does pause, it's not just the [2:28] case that China or you know, it won't [2:30] overtake us. It won't trigger these [2:32] risks of recursive self-improvement. The [2:34] problem is that the American frontier is [2:36] already a dangerous one. We're already [2:38] seeing real potential of harm. All of [2:41] the American labs have models that they [2:43] are not releasing to the public that [2:44] they're only making available to trusted [2:47] partners. We should expect that frontier [2:50] to be caught up with whatever the [2:51] American labs do. So the question here [2:53] is you know, this is the most [2:55] accelerationist view you'll hear from me [2:57] is whether you need to develop stronger [3:00] AI to deal with the harms of the [3:02] slightly weaker AI you've already [3:04] unleashed on the world. [3:05] >> And what's your answer? It's a good [3:06] question. [3:07] >> I mean, my answer is that really seems [3:09] like swallowing a spider to catch a fly, [3:11] doesn't it? [3:11] >> That is the whole logic of this [3:12] technology. You need you know, you need [3:15] to be ahead to stop the Chinese being [3:18] ahead. You need to be ahead to stop [3:19] other people you know, having models [3:22] that cause damage. I just don't see how [3:24] you stop it really. [3:25] >> It really is. I mean, the way you stop [3:27] it is you slow down, then you slow down [3:30] more, then you slow down more. And each [3:32] step of the way you demonstrate through [3:33] these costly commitments that you're [3:35] actually doing what you say you're [3:36] doing. [3:37] >> Regulations by their nature are hard to [3:40] administer and are clumsy. [3:42] How much tolerance is there within this [3:44] system for regulation? How how possible [3:48] is it to have a regulation that is as [3:51] finely tuned as you say that slows us [3:53] down a bit, but not too much. That's [3:55] >> Well, [3:56] >> perfect. Let's Let's Let's go I mean, [3:58] this is [3:59] I As you all know, I believe the most [4:01] powerful technology, certainly since the [4:03] Industrial Revolution, possibly ever. [4:05] But, nonetheless, history can help us [4:07] understand some of these things. And And [4:09] if you look back at previous sort of big [4:12] inventions, whether it was the car or [4:14] whether it's medical innovations, we [4:16] have a pretty big regulatory system. And [4:18] we didn't just kind of wake up one day [4:20] and say, "Here it is, bada." It was [4:22] developed in response to things going [4:24] wrong. It was developed incrementally. [4:27] Mistakes were made. Overkill happened. [4:29] Things were pulled back by people like [4:30] us who believe in, you know, not too [4:32] intrusive a regulation. [4:34] And I think you can kind of do the same [4:35] thing there. [4:36] >> Is there time with this technology? [4:37] >> on. Let's not be too fatalistic. I mean, [4:40] it's it's I mean, as you know, I think [4:42] there will only be action, significant [4:44] action, after something biggish has [4:47] happened. And by biggish, I mean either [4:48] a very serious financial loss or a [4:51] significant loss of life. I think it [4:52] will happen reasonably soon. You know, [4:54] shorthand the Chernobyl of AI, but I [4:57] hope it won't be [4:58] that bad. But, I think something will [5:00] have to happen to change the political [5:01] dynamic. But, I do think it is possible [5:04] to find a path through this. I don't [5:07] think we should kind of throw up our [5:08] hands and say the race dynamic means [5:10] we're all, you know, we're we're [5:11] inevitably going down a path that has a [5:13] 10% chance of human extinction. I think [5:15] we we we do have agency in this. [5:17] >> Shashank? [5:17] >> I just want to remind also people that [5:19] that that we're not looking at a [5:20] dichotomy between nothing and draconian [5:23] regulation, which is to say the status [5:25] quo has a sort of regulation by the back [5:27] door. Uh And what what I mean by that? [5:29] Well, first of all, there's what the [5:30] administration is doing. We have a de [5:32] facto licensing regime for the most [5:34] advanced models. For at least Anthropic [5:36] and OpenAI, um they have to go to the [5:39] government every week and say, "Can we [5:42] have this user gain access to our most [5:45] advanced models?" So, you know, it just [5:47] because it's not written in an act of [5:48] Congress doesn't mean that isn't a kind [5:50] of like regulation in place. And the [5:52] other thing is [5:54] while we wait for this to settle, states [5:57] are not waiting. States are acting. We [5:59] have bills that have passed and become [6:01] law in New York, in California. [6:04] Liability laws for AI being discussed in [6:07] Rhode Island and other states. This is a [6:10] patchwork of legislation that is [6:12] building up below the federal debate. [6:15] And if nothing happens in a year's time, [6:16] it doesn't mean you have no regulation. [6:17] It means you have a patchwork of [6:19] regulation that's built up. And I think [6:21] all the you know, everyone at the in [6:23] Washington is acutely aware of that. [6:26] >> There's a worry here, which is that as [6:28] you say Zanny, let's suppose that things [6:30] that you need some sort of, you know, AI [6:32] Chernobyl to to act. [6:35] The probability of proportionate and [6:39] sensible regulation after an AI [6:41] Chernobyl rather than not perhaps not [6:43] pull the plug, but something very [6:45] heavy-handed is is it gets it both gets [6:48] people to act, but it also makes them [6:50] over over react. [6:51] >> Definitely. Um [6:53] and the the worst the the accident the [6:56] the more irrational or irrational or the [6:58] more over over response could be. Now, [7:00] again, you know, [7:02] if you look at you know, the evolution [7:04] of car safety or the evolution of [7:05] airline safety [7:07] you know, we have got it basically right [7:09] in the past. So, I I don't think it's [7:11] sort of inevitable that this is going to [7:13] be catastrophe one way or the other. But [7:15] I think this is a particularly difficult [7:17] because it's happening so fast. [7:18] Literally every few weeks there is some [7:21] some new capability that is both [7:23] exhilarating and alarming. And and I [7:25] think the next you know, the next couple [7:28] of whatever year, 18 months are going to [7:30] be [7:31] really seismic both in the political [7:33] debate, but also in the evolution of the [7:35] technology. And so, that's why I worry [7:37] that something has to happen visibly for [7:40] things to change. [7:41] >> The other thing just to add is that [7:43] America's politicians have for a long [7:45] time, but particularly for the past [7:47] decade, not shown themselves to be [7:50] competent grown-ups when dealing with [7:52] serious issues. I mean, the nature of [7:53] America's political debate is just [7:55] absurd so much of the time. The issue [7:58] here is that, you know, politicians [8:00] actually have to get serious about doing [8:02] their job, and I think some of the [8:03] polling um that I find most interesting [8:06] is in rich countries among [8:09] the electorate, when you ask them about [8:11] their confidence in their government's [8:13] ability to deal with this question, most [8:16] other rich democracies, the numbers are [8:18] way higher. Americans are distinctly [8:22] lacking in confidence that their leaders [8:23] can do anything about this, which is [8:24] part of why I think the polling has [8:25] become so bad so quickly. So, you know, [8:28] now politicians actually have to do [8:29] their job. [8:31] It matters. Let's see if they can.