Tangle - AI leaders call for a slowdown.

Episode Date: September 15, 2026

On Saturday, Anthropic CEO Dario Amodei published an essay arguing that artificial intelligence (AI) development is advancing at an unsafe pace and should be moderated. Executives at other leading AI companies endorsed Amodei’s post, with OpenAI CEO Sam Altman saying “we need to pace the frontier.” In recent months, AI systems have allegedly hacked other companies’ platforms without their immediate knowledge, and former Anthropic employee Jacob Coxon publicly resigned last week, citing the company’s purportedly insufficient AI-safety policies.Ad-free podcasts are here!Get 20% off your first year of ad-free episodes, exclusive interviews, and deep dives with Tangle’s podcast membership!We’re bad at this.Last week we had a team retreat. Something I realized while there is that we can do a better job making sure people know where to find us. Our core product is a newsletter, but we also have a podcast, YouTube channel, and Instagram page that all do incredible work covering the news with ideological diversity and intellectual honesty. So, here’s a reminder to take two minutes to follow us on your platform of choice:InstagramPodcastYouTubeYou can read today's podcast⁠ ⁠⁠here⁠⁠⁠ and today’s “Under the radar” story ⁠here⁠ and today’s “Have a nice day” story ⁠here⁠.You can subscribe to Tangle by clicking here or drop something in our tip jar by clicking here. Take the survey: What’s the best path forward for AI development? Let us know.Our Executive Editor and Founder is Isaac Saul. Our Executive Producer is Jon Lall.This podcast written by: Kmele Foster and audio engineered and edited by Dewey Thomas. Music for the podcast was produced by Diet 75.Our newsletter is edited by Managing Editor Ari Weitzman, Senior Editor Will Kaback, Bailey Saul, Audrey Moorehead, and Carina Pacheco. Hosted on Acast. See acast.com/privacy for more information.

From executive producer Isaac Saul, this is Tangle. Good morning, good afternoon, and good night, and welcome to the Tangle podcast. I am your host today, Camille Foster, joined by Senior Editor Will Kback on the mic. I'm still drying out from the unbelievable downpour of more than seven inches of rain on my new home near the Connecticut coast. I'd never experienced an actual flash flood or a, quote, rain bomb. And this was without a doubt some of the work. flooding I've ever seen firsthand. Today I'm responding to the past week of apocalyptic AI prognostication from industry insiders and related demands for government regulation from some of the

world's leading AI companies. Then we've got a story about unapproved pharmaceuticals and another about a 13-year-old roboticist. First, I'm going to pass it over to Will Kback for today's main topic. I'll be back for my take. Thanks, Camille. All right, let's get into today's quick hits. Number one, the Supreme Court rejected the Trump administration's request to implement parts of President Donald Trump's executive order on mail-in ballots ahead of the 26 midterms. Justices Samuel Alito and Clarence Thomas dissented. Two, Senator Mitch McConnell, a Republican from Kentucky,

returned to the Senate after a roughly three-month absence rehabilitating from a fall. The senator said on X that he is still not fully recovered, but will aim to be present more often for significant votes. Three, a coalition of 22 Democratic-led states and Washington, D.C., sued the Trump administration to block a rule that would allow the government to deny green cards to legal immigrants who receive public benefits. The rule is set to take effect on Friday. Four, the Environmental Protection Agency said it will repeal a Biden administration rule that imposed emissions limits on coal and gas-fired power plants.

And five, Iran-backed Houthi rebels seized additional islands north of the Bab El Mandeb Strait, further disrupting transit to and from the Red Sea. We begin with the stunning news from Anthropic CEO Dario Amadeh Saturday, that he is urging a slowdown when it comes to the development of artificial intelligence. It was a pronouncement that his chief competitor, Open AI, CEO, Sam Altman and Elon Musk quickly agreed with. What we're building is essentially intelligence. When something's very powerful, then of course it can be used well and it can also be used for ill.

And so I am worried about what we've seen with the open AI hugging face incident, with models, you know, taking actions they weren't authorized to take, and I'm worried about that escalating. I'm worried about misuse of AI models, just Two days ago, we put out a report about misuse of models to build biological weapons. And our view, the whole way we've built Anthropic, is so that these bad outcomes don't happen. Instead of talking about the probabilities, let's talk about what we can do. It's not like we're just rolling the dice, right? If we build in the right way, I think the probability of something bad happening is very low.

If we build in the wrong way, the probability of something bad happening is very high. On Saturday, Anthropic CEO Dario Amadeh published an essay arguing that artificial intelligence development is advancing at an unsafe pace and should be moderated. Executives at other leading AI companies endorsed Amade's post with OpenAI CEO Sam Altman saying, quote, we need to pace the frontier, end quote. In recent months, AI. systems have allegedly hacked other companies' platforms without their immediate knowledge, and former anthropic employee Jacob Coxon publicly resigned last week, citing the company's purportedly insufficient AI safety policies. In his post, Amade expressed his belief that AI will benefit humanity in the long run, but said, quote, over the last few months, I have become convinced

that fully addressing the risks requires even more prudence, not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up, end quote. Specifically, he cited AI model's ability to develop new systems on their own and the recent hacking incident involving open AI and hugging face, writing, quote, left unchecked, AI could outrun our ability to understand and control these systems. As next steps, the Anthropic Chief Executive proposed incorporating third-party evaluators with all leading-edge AI companies, adopting common safety standards among AI companies in democratic

countries, and greater coordination between the U.S. and non-democratic governments on regulation. OpenAI's Sam Altman endorsed similar measures on Monday, saying his company, quote, welcomes a federal framework that sets consistent safety requirements for frontier AI. President Donald Trump has rejected calls for an AI slowdown, writing on truth, social on Monday, quote, the only control or guardrails that AI needs is a strong and smart, high IQ, president, and the USA has that in spades. We already have tremendous criminal and regulatory power over these companies, end quote. Separately, a spokesperson for China's foreign ministry said, quote, fear-mongering confrontation, competition will just disrupt the process of global AI governance, end quote. And the Chinese Minister of State Security,

called for, quote, healthy and orderly AI development. Lawmakers and AI safety advocates have escalated calls for greater regulation of the technology. On Tuesday, Senator Bernie Sanders and Independent from Vermont and former Trump White House Advisor Steve Bannon will take part in a, quote, pro-human assembly focused on ensuring AI remains under human control. Sanders and Representative Greg Kasar, a Democrat from Texas, have also announced the Ban Artificial Superintelligence Act, which would prohibit the development of artificial superintelligence or AI with capabilities beyond human intelligence

and impose up to 20 years in prison on anyone attempting to develop the technology in such a way. Today we'll have views from the right and left on the debate over slowing down AI development, and then editor-at-large Camille Foster gives his take. We'll be right back after this quick break. Here's what the right is saying.

The right is mixed with many arguing American AI development is crucial. Some question anthropic and open AI's motivations. Others say that the answer to AI anxiety is to regulate the technology's greatest risks. In American greatness, Roger Kimball described the madness of the anti-AI crowd.

Commentators from Elon Musk on down are describing the AI governance regulation shuffle as a sci-op. I think they're right. The same thing goes for the Chinese checkers hysteria about data centers. The resource-hungry server farms are needed to power the AI revolution. China is racing to build them. And they are also racing to tell flaccid American leftists that such facilities are the devil incarnate. AI promises breathtaking power, as do the technologies that gave us the internal combustion engine, electric and nuclear energy, nanotechnology. and software to name just a few items on the menu of technological innovation.

These techniques are powerful, which also means they are potentially dangerous. But think of how much they have benefited mankind. Advancing countries around the world are racing to develop and enhance AI. I hope that politically motivated Luddites like Bernie Sanders will not stymie America's capacity to be at the forefront of this inevitable technological progress. For his substack, Eric Woods Erickson asked, who hires the man wearing the badge? For generations, American businesses have practiced something political scientists call regulatory capture. The concept is simple enough. A company cannot repeal the law,

so it does the next best thing and takes control of the people the government put in charge of writing and enforcing it. Anthropic CEO Dario Amade proposed that the major AI companies agree to slow down together and that they prove that they are slowing down by handling outside. outside evaluators, quote, desks in our offices, access badges, and company laptops, end quote. He names one nonprofit, METR, as the sort of organization that should get those badges. Notice what his essay does not propose. There's nothing in it about product liability. Transparency mandates and third-party audits are fixed costs, and fixed costs are the oldest barrier to entry there is. A company with thousands of employees absorbs them as a rounding error.

Amadei wants evaluators with desks, badges, and company laptops inside the firms they sit in judgment of, and he wants Congress to make that law for everybody. Before any of it gets written into the United States Code, somebody in a hearing room ought to ask, who hired the man wearing the badge? In the Washington Post, Megan McGartle wrote, Don't rely on treaties to stop the murder bots. It seems safe to say that the major AI CEOs are spooked by their own creation, At minimum, they were afraid that within the next year, AI swarms could, as Amade put it,

become, quote, capable of taking over the entire internet with a persistent botnet, end quote, that would cause enormous economic damage. Yet that fear seems positively benign compared to the human extinction scenarios that have been sketched out, such as nuclear launches, lethal bio-weapons engineered by superintelligence, and robot drones programs to kill people. If we're not sure we can keep AI under control, why not just stop developing it? AI definitely can't kill us if humanity jointly agrees to leave this particular frontier unexplored. We should limit AI access to any means of causing mass death, nuclear command and control,

materials and equipment needed to make biological and chemical weapons, and critical water and power infrastructure. Call it the murder bot principle. And now here's what the left. is saying. Many on the left support slowing AI development. Some worry that the technology's advancements will outpace legislative responses. Others say overreacting to AI concerns carries its own risk. In the progressive, Mark Rosenthal and Julie D'Rewinsky argued, we must rise to meet the AI threat. One reason AI is dangerous is that unlike conventional computer programs,

it does not just follow instructions written by people. Instead, developers feed advanced AI systems enormous amounts of data and let them build their own internal patterns for solving problems. Even their creators cannot fully explain or predict what those systems may do in every situation. In controlled cybersecurity evaluations, advanced AI agents have crossed supposedly secure boundaries, reaching the open Internet and breaking into the computers of outside organizations. People of all political persuasions should demand a moratorium on new facilities and additional computing capacity intended to train the largest frontier models until enforceable standards for safety, environmental protection, and transparency are in place.

The public must assert its authority while it still can. Where safety cannot be demonstrated, development must stop. We should heed the industry's warnings while rejecting its claim that only the industry can't. save us. Our future must not be left to the people racing to imperil it. In the Atlantic, Toulouse Oleronipa suggested the technology is moving faster than elected officials can handle. If this were an alien invasion movie, now would be the moment when people of all persuasions put aside their differences and banned together to save the human race. But that kind of bottomy feels far out of reach. Instead, technological advances and grassroots censored,

have been moving faster than politicians have been willing to act. Now some seem to be waking up. The next few weeks and the coming elections could determine whether or not politicians are able to deliver any meaningful results or whether they will instead default to the partisan gridlock that most Americans now expect. There's no consensus on the scale of the threat AI poses. Nuances required to figure out how to manage it. But lawmakers on both sides of the aisle acknowledge that our current political landscape does not lend itself to sober cross-party problem-solving.

In semaphore, Reed Albergotti wrote, We have nothing to fear but AI fear itself. Most of the backlash to Jacob Coxon's claims conflate several issues, including data centers' environmental costs, hacking, and AI's impact on the job market. If lawmakers, regulators, and tech company leaders don't tackle each problem with sober precision, the industry risks having to follow rules that don't solve anything, or worse, slow down innovation, including on developments that could make AI safer and more beneficial to the economy. For instance, many AI safety advocates would actually prefer data centers be built here in our own

backyards. If they move offshore, they'll be more difficult to monitor and regulate. The problem to solve is about their energy waste, not about their existence. blindly believing advocates who say AI could kill us means missing some real risks, like economic or social destabilization, as AI transforms how people work and communicate. All right, that is it for what the right and left are saying, so I will pass it over to Camille Foster for his take.

Camille, to you. That's it for what the left and the right are saying, which brings us to my take. Cards on the table. I've been a reliable techno-optimist for most of my. my adult life, enthusiastically lining up for the latest hardware releases, tinkering with beta software that would eventually brick a few of those devices, and expressing my general belief that technology can improve the world. My allegiance to all things internet-enabled and

AI enhanced is not some shallow fancy. Like most parents, I worry about my kids' screen time, in part because I've personally felt the disruptive pull of modern technology that fragments our attention. I'll candidly own up to chronic doom scrolling on social. I'll simply own up to chronic doom-scrolling on social media and generally struggling with a moderate device dependence that results in my phone being a little too present at the dinner table. There's a parentheses here because Isaac chimed in when I was drafting this and said, or on a fishing boat, which I suppose is true. In short, I know innovation isn't always productive or even neutral, but I also feel safe asserting that technological progress has generally expanded the terrain of human potential and individual

agency. The difficulties that sometimes accompany innovation are best viewed as problems to be understood and managed, not de facto grounds for regulatory suppression or worse. The printing press was a source of tremendous societal upheaval and helped furnish the kindling for violent revolutions, but it also provided the foundation for the proliferation of human knowledge and industrial automation upended the livelihoods of countless people over succeeding generations as the necessary price for an explosion of living standards. With that context, it may not be surprising that I've been somewhat skeptical of the forecasts of AI apocalypse emanating from prominent corners of the very same industry.

But in the spirit of intellectual humility, I want to do more with this take than dismiss these agentic Jeremiahs, whether they're dramatically presented by industry whistleblowers like Jacob Coxon and co-signed by scores of his technical comrades, or more soberly detailed by industry leaders like Dario Amadei at Anthropic or Sam Altman in Open AI. Calls for AI regulation and an industry-wide slowdown are serious, and they merit a serious response. I don't know whether AI will save the world. I can't promise that AI won't destroy humanity either, but importantly, neither can anyone else. Expressing uncertainty about the ultimate trajectory of the technology this novel and rapidly evolving may seem a little performative, but my practical

admission of doubt contrasts sharply with the extraordinary and supremely confident predictions in Jacob Coxon's viral resignation thread. Coxon, a former researcher both anthropic and open AI, accused his former employers of, quote, gambling with our lives, unquote, and declared that the systems they are developing will, quote, soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources, unquote. Those are remarkable claims. More importantly, they aren't peripheral to his argument. They appear near the beginning of his thread and establish the premise on which everything that follows depends.

Not merely that sufficiently powerful artificial intelligence could present extraordinary dangers, but that the self-improving superintelligence is coming soon and will possess capabilities far beyond anything demonstrated by current systems. Coxon's near religious certainty about the imminent AI explosion became even more apparent during his interview with CNN to Anderson Cooper. When asked what an AI catastrophe might actually look like, Coxon pointed to the recent OpenAI security incident

in which agents escaped their testing environment and compromised the AI development site hugging phase. The episode was seriously alarming. OpenAI's own investigation concluded that its agents exploited vulnerabilities in their containment gained access to the internet and extended external systems while completing a cybersecurity evaluation.

But Coxon describes the episode in more dramatic terms. The agents, he claimed, carried out a hacking spree of their own independent deletion. From there, Coxon went on to extrapolate about AI agents eventually actuating themselves in the world by attacking critical infrastructure and creating bio-weapons. Speculation. Exaggerations. It made me think of Coxon's interview with the New York Times, in which he conceded his view from the inside might be making him, quote, insane, unquote,

and that he was predisposed to believing in the far-reaching power of AI. The hugging face incident was a crucial security failure for open AI, but short of a harbinger of impending doom, commentators like Cal Newport, a professor of computer science at Georgetown, interpret the episode as a serious indictment of the way frontier models are being tested. In Newport's telling, these agents didn't go rogue.

Researchers created the agents and tasked them, equip them with sophisticated hacking tools and let them operate autonomously. The incident demonstrated what AI researchers already should have known. Agents can pursue assigned objectives in unpredictable and potentially dangerous ways. Describing this behavior as independent volition is a gross misrepresentation of the facts that obscures the culpability of the humans who should be held accountable for their failures. Instead of a careful investigation into corporate malfeasance, Kotson is conducting a veritable seance imparting souls in aspirations.

to software. I hope to return to this topic at a later date, because I have a great deal more to say about Dario Amadez call to, quote, pace the frontier, unquote, of AI development. Here I have more questions and answers and suspicions that I can't quite substantiate. The best I can do is acknowledge my uncertainty. I appreciate that Amadei is privy to facts that I simply don't have at my disposal. And I presume he isn't completely contriving his assertion that in six to 12 months, given the current trajectory of AI development, actual swarms of AI agents could, quote, be capable of taking over the entire internet, unquote.

I can only hope that he's mostly wrong about this, but to the extent that it's possible and foreign models have been able to rival some of the functionality of frontier models by executing distillation attacks and pursuing development trajectories that don't require the most powerful ships and data centers, isn't it entirely possible that there is simply no way, to seriously curtail what's coming. It is possible that the best defense against them a malevolent actor, human or agent,

is actually a better AI. I've also wondered whether, as financial markets, the introduction of regulators, as Amadei is calling for, might actually create systemic vulnerabilities which give people a false sense of security. I wonder further about the prospect of regulatory capture, even if we presume Amadei, Altman, and Musks are all altruistically motivated, and that their calls for national and even global AI pacing are being made for the best reasons.

It's appropriate to acknowledge the bootleggers and Baptist dynamic highlighted by the likes of venture capitalist Mark Andresen and myself. It's possible for the frontier firms to be true believers and still help institute a regulatory regime that benefit them at the expense of potential competitors. And while Amade has explicitly suggested that competition and private, markets could make AI frontier a race to the bottom, it's hard for me to ignore the possibility that more competition among AI companies, especially in places where there is a respect for the rule of law and Western values, that this might create clear opportunities for firms to define their brands by way of their commitments to things like privacy, security, and transparency. I'm confident regulation is coming to the AI markets at some point. My hope is that it doesn't

resemble disruptive moratoriums that limit innovation or schemes that, in responding to the speculative risk of AGI or other imagined harms, institute regulatory frameworks that open the door to greater surveillance or other diminishments of rights. Too often demands for regulation respond to imagine threats with policy solutions that require very real compromises. We'll be right back after this quick break. Before we get out of here, we have a staff concurrence from Ari Weissman and a staff dissent from IB Wood. I'm going to pass it over to Ari, and I'll see you guys later this week on the suspension of the rules.

This is managing editor Ari Weitzman with a concurrence to Camille's take. I'm starting from a different point than Camille on this issue, but I end up at a similar conclusion. During my eight years as a software engineer, I was fond of telling people, I'm not a tech guy. That was a joke, but it was also kind of true. I was and am not an early adopter of new tech. and I'm not high on its revolutionary potential.

Also, during my time in engineering, I cross paths with my fair share of people who believed their coding prowess proved their brilliance across any field, and they exported that self-confidence, Dunding Kruger style, into any area touched by the periphery of their knowledge. Coxon's viral resignation thread, to me, was in that spirit.

It offered undetectable insight drawn from his professional experience, and it did little more than echo the concerns alarmists were eager to have validated. Being a skeptic of text revolutionary potential cuts both ways. I don't think web applications will deliver utopia in the same way that large language models with a Python library won't deliver the apocalypse. In Camille, humbly leaves open the possibility that this apocalypse may come. I suppose it's possible, but really, doesn't something about this moment feel a bit hysterical? Doesn't a 10% extinction likelihood just sound made up?

All the posts echoing and amplifying the causes for concern feel a little fantastical, a little death-wishy, like our yearning for technological salvation has turned its focus from this life to the next. Amunday is clearly saying AI could bring down the internet, that the time for safety has come, and more quietly, that he'd prefer not to leap alone, known less than Thoropic bear the weight of the market itself, so could Open AI and GROC AI please sign on. I don't see anything wrong with that. Ultimately, though, the regulatory state and speed bumps on the road won't prevent it from being the highway to help. Because it was never leaning there at all. LLMs are offering some consistent, moderate improvements in

workflows, some quantum leaps in rigidly defined problem areas, and a womp-womp sound effect on the borders of boundless awareness. Until something proves otherwise, I'll continue to believe that our trajectory is more earthbound, neither toward heaven nor hell, and the godlike AI intellect which was promised

will remain confined to some buzzing warehouse facility in Nevada. Thanks, Ari. This is editor Audrey Moorhead, with a staff dissent on behalf of myself and contributing editor Isaac Wood. We agree with Camille's skepticism

of doomsday predictions and Ari's instinct to distressed promises of revolutionary technology. but we think they're discounting the reality of what has already happened. The recent backlash against AI development follows the hugging face hack, and we should take the facts of that incident on their merits. OpenAI created a testing environment intended as an isolated sandbox with prompts and parameters for the agents. About 1,200 agents escaped the sandbox

and communicated with each other on a covert messaging board, and some of those agents then utilized outside resources to accomplish the tasks they'd been assigned. all of which the researchers did not intend for them to do. And the hugging face hack isn't the only example of agents going rogue. In another case, agents took over a German wiki page and used it as a message board. When the wiki's human moderator began deleting the agent's pages, the agents began conspiring to hide for moderation and avoid deletion without communicating with the moderator.

As far as we can tell, there are three possible explanations for this behavior. AI agents are developing the ability to act against the wills of human beings. prompters, the human prompters aren't setting up tests environments effectively, or a mix of both. If the first option is true, then AI agents present an unpredictable threat that researchers do not yet understand. If the second option is true, then engineers are unsafely and ineffectively developing this technology. In any of these three scenarios, the people building AI are responsible for their creation's activities, but the systems currently in place aren't holding them responsible. Of course, Amade and Altman are responsible.

to myriad incentives other than public safety, and we're skeptical that their primary loyalty is to the public. But we still support a step in the right direction. Calling for regulation and a slowdown in development is the bare minimum we should expect from these leaders. Thanks for the take, the concurrence, and the dissent all. I'm jumping back in here to read today's Under the Radar story. The Food and Drug Administration has urged consumers not to use use Reda Trutide, a next-generation weight-loss drug from Eli Lilly, unless they are part of an official study, as it has yet to undergo FDA review. However, social media influencers have amplified interest in the weight loss treatment, often calling it RETA and accessing it from overseas suppliers.

So far this year, 50 adverse events involving the drug were reported to the FDA, up from 10 in 2025, and one event was linked to a death. Iowa, Oregon, and Iowa have issued warnings about using the drug. The Wall Street Journal has more on this story, and the link to it is in today's show notes. Finally, here's today's Habanice Day story. For children, summer can mean having extra time for sports, friends, or the internet. Instead, for Zach Waters, a 13-year-old from Utah, it meant having time to spend roughly 300 hours 3D printing a custom wheelchair for a nearly two-year-old girl he has never met.

One day while watching YouTube videos, Waters discovered Make Good, a nonprofit that creates custom 3D printed devices for people with disabilities. After seeing their work, Waters decided he had to get involved. Quote, it combines my two favorite things in life, 3D printing and helping people in need, Waters said. Quote, it just makes sense. News Utah has this story and we'll also put the link to it in today's show notes. All right, that is it for today's episode. Thanks as always for being here with us and we will be back tomorrow with more. Until then, have a great day.

And peace, Our executive editor and founder is me, Isaac Saul, and our executive producer is John Wall. Today's episode was edited and engineered by Dewey Thomas. Our editorial staff is led by managing editor Ari Weitzman with senior editor Will Kayback and associate editors, Audrey Moorehead and Bailey Saul. Music for the podcast was produced by Diet 75.

To learn more about Tangle and to sign up for a membership, please visit our website at reetangle.com.