80,000 Hours Podcast

Anthropic’s AI is working to build the next version of itself. For the first time ever, an independent team got a chance to investigate this.METR spent 11 weeks inside the lab, with access to data almost no one outside Anthropic has seen. They found its AI is improving twice as fast as last year. Largely due to AI doing the work.To put that into perspective: in 2023 Claude couldn’t interpret images or search the web. This year, it helped crack a math problem that had stumped the best mathematicians for 87 years.Now imagine another jump like that, but in months instead of years. AI companies could build models far more powerful than today’s, before anyone could understand or control them… and then use them to build more capable successors even faster.Researchers call this recursive self-improvement. And it seems to be where we’re heading.METR found nothing inside Anthropic that suggests this process is likely to slow down. In fact, it says the rate of speedup itself could be growing.Anthropic redacted much of the report, so we may never see everything METR saw.But in this video, Luisa Rodriguez explains what METR was allowed to share with the world, why many of the most respected experts are freaking out, and the strongest reasons METR might be wrong.She also covers what needs to change, and what you can do to help.This episode was recorded on November 9, 2026.Learn more, video, and full transcript: https://80k.info/METR-RSIChapters:AI is now building AI twice as fast (00:00:00)11 weeks inside Anthropic (00:01:09)Two years of progress in one year (00:01:59)Is AI causing the speed-up? (00:03:19)No signs of slowing down (00:05:15)Could METR be wrong? (00:07:21)Has Anthropic crossed its own line? (00:10:18)What needs to change (00:11:33)What you can do (00:13:12)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, Ollie Bignell, and Andrés EscobarProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore, Lou Moran, and Taylor Smith

China’s best models are only six months behind America’s, despite having far less computing power. How is that possible?Luisa Rodriguez investigates, and finds that much of China’s progress is made by using American chips and copying American models:US rules stop NVIDIA selling its best chips to China. But Chinese companies can still rent those same chips in data centres abroad.Smugglers go to extreme lengths to sneak chips into China.Chinese AI companies copy America’s best models on an industrial scale.Today’s most powerful technological secrets are protected by alarmingly inadequate security.The US could tackle these gaps with powers it already has. So the current situation, with China only months behind, is largely a policy choice.If the main argument against slowing down is that China might catch up, none of this quite adds up.Four days after recording this, NVIDIA’s CEO Jensen Huang — who features in the video as one of the loudest voices against restricting China’s access to chips — admitted to having an affair with an alleged Chinese spy. This shows just how badly China wants American chips.Closing these gaps in US AI policy could buy time to test new models and hold back those that aren’t safe. Before we accept that the US has to race ahead, it’s worth asking who benefits from racing, and what’s currently being given away.This episode was recorded on November 4, 2026.Links to learn more, video, and full transcript: https://80k.info/china-chipsChapters:Update: since we recorded (00:00:00)Everyone said China had caught up (00:00:30)Who’s building China’s AI? (00:01:50)How far behind is China? (00:03:20)Why can’t China make its own chips? (00:04:40)So where do China’s chips come from? (00:06:24)How do you copy an AI? (00:10:35)China is running a different race (00:12:34)Why a bigger lead matters (00:15:30)Could the US widen the gap? (00:16:51)What if China just steals the models? (00:21:12)So can the US slow down? (00:24:53)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, Ollie Bignell, and Andrés EscobarProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore and Lou MoranCamera operator: Dominic Armstrong

We make a lot of episodes about AI these days. We recently picked out the best 10 episodes of the show to really help deepen your understanding of AI and AGI and superintelligence and recursive self-improvement — and the impact all of those things might have.We aimed for episodes that were entertaining as well as highly informative, and tried to choose 10 that didn't overlap too much, so between them could cover most of the core issues. You can find them all now in a feed called The 80,000 Hours Podcast on Artificial Intelligence, available wherever you get podcasts.And at https://80000hours.org/AIpod, you can sign up to get one of them sent to your email each week for 10 weeks — plus one week with our bonus episodes of article readings and how to pivot your career into AI.Across the 10 core episodes, you will learn:How AI could compress a century of intellectual progress into a decadeWhy today's training techniques actively encourage models to lie to usHow the economy and geopolitics will shift in response to superhuman AI, and indeed are shifting already in response to superhuman AIHow superintelligent AI is in serious tension with maintaining democracythe real nature of the AI arms race between the US and ChinaHow AI shifts the risks of pandemics and bioweapons over timehow to think about AI consciousness and welfare and whether it matters, and the scramble to understand what AI models are really thinking about deep down.If you're also looking to change your career to work in AI, we close out with an episode that lays out dozens of valuable concrete projects that need people to work on them right now, and an interview with the founder of 80,000 Hours explaining how to redirect your career towards this kind of work.Learn more about what you can do to help: https://80000hours.org/ai/

This is an (unpaid) cross-post of an episode from the podcast NonZero, which Rob Wiblin thought you might be interested in. The originals are on the NonZero website: Part 1 and Part 2. If you like it and want more similar stuff in the future, consider subscribing to NonZero on YouTube, Spotify, Apple, or wherever you get podcasts. What will it take for Washington to slow down AI? Not better arguments, says Jeremie Harris, CEO of Gladstone AI and coauthor of a 2024 State Department–commissioned action plan on AI risk. He argues it will take an “AI Chernobyl”: a disaster so severe the president fears for their own life, and Washington’s official stance on AGI risk flips.Jeremie has spent six years betting that AI’s scaling curve won’t bend. His biggest miss ran the other way: “Frankly, I thought we ought to be dead by now.” Capabilities keep improving almost automatically, but safety doesn’t, because “no one knows how to make a curve of control or a curve of alignment.”He was making that bet before it was respectable. Jeremie recounts his years trying to get people in government to take AI risk more seriously, including his role in the battle to get “AGI” into a Biden-era State Department report.He also takes aim at Open Philanthropy, now Coefficient Giving (80,000 Hours’ largest funder), which was then the dominant funder in AI safety. Jeremie claims it discouraged outsiders from briefing Washington, for fear officials would hear “powerful weapon” and launch a Manhattan Project for superintelligence. He criticises its long timelines and a grantee culture of telling funders what they wanted to hear, which he says left the field slow to engage government. Still, he credits Coefficient Giving with getting the problem “much more right than almost anyone.”His assessment of the AI companies is worse. Jeremie thinks OpenAI has gone through “evaporative cooling,” shedding its most safety-minded staff round after round. He also doubts Sam Altman ever fully believed his own warnings.We’re sharing this because Jeremie’s takes on warning shots, government response, and past mistakes in AI governance were new to us. Not everyone will remember the history the way he does, but it’s provocative in a useful way.Chapters:Rob’s intro (00:00:00)Jeremie’s podcast and prescient AI report (00:01:31)Is scale still (mostly) all you need? (00:07:32)Jeremie: We’re way past “stochastic parrots” (00:13:45)Inside the early AI safety ecosystem (00:22:35)What Yudkowsky got less wrong than most (00:34:24)Why capability is outpacing alignment (00:42:18)Were Altman’s fears ever sincere? (00:49:49)Would pausing AI crash the economy? (00:56:04)Overtime preview: Debating China (01:01:45)Bob’s beef with “kumbaya” (01:10:22)Will it take an AI Chernobyl to convince the skeptics? (01:12:19)Is China-hawkism the safest option? (01:22:55)The case for “organic transparency” with China (01:50:43)How alignment progress could backfire (01:58:02)Has China infiltrated US tech? (02:03:18)Breaking down Dario’s democracy bloc idea (02:12:38)

The big 3 AI companies all had serious scandals last month. Anthropic pushed a federal AI bill that would make it harder to stop dangerous model training. OpenAI fired three safety researchers based on evidence that’s somewhere between weak and nonexistent. And it came out that Google DeepMind gets staff to sign contracts barring them from criticising anything Google does for the rest of their lives.But these companies are not beyond influence. The track record of holding them to account — staff objecting internally, journalists digging for dirt, campaigners making a fuss — is impressive for how little effort has gone into it.Many staff want to do the right thing and can be persuaded on the merits. And the fierce competition for talent forces companies to try to seem like the good guys.That’s why being a pain in the ass for the leading AI companies is a legitimate career path, and an underrated one.Inside the companies we need even more staff to speak up clearly and visibly with the substantive case for their firms to raise their game.And externally we need, among other things:Investigators uncovering company misbehaviour and failuresSubject matter experts laying out the facts for journalists and politiciansCampaign organisers and social media experts focusing public attention where it’s neededToday I lay out what each AI company got wrong this time, why the ability to bring down the immediate hammer of fierce public and expert outrage matters more as events speed up, and the many ways you can contribute to that effort.If you want to work on holding AI companies to account, our career review of AI safety advocacy is a good place to start, our 1–1 advising team can help you think it through, and our job board lists AI safety and policy roles.This episode was recorded on October 29, 2026.Links to learn more, video, and full transcript: https://80k.info/AI-accountabilityChapters:We need a stronger effort to push AI companies to do better (00:00:00)Anthropic: who gets to pull the brake? (00:01:14)OpenAI fires three safety researchers (00:07:14)Google DeepMind’s lifelong gag clause (00:12:44)Credit where it’s due (00:15:39)How to make AI companies behave better (00:16:34)Insiders: if you want to stay (00:17:31)Insiders: if you’re willing to leave (00:18:40)So… should you leave? (00:19:37)Outsiders: what’s already worked (00:20:39)Outsiders: digging up and explaining (00:22:51)Outsiders: reaching the public and politicians (00:25:12)Keep the pressure on (00:27:52)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, Ollie Bignell, and Andrés EscobarProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore and Lou MoranCamera operator: Dominic Armstrong

The US government spends 25 times more on military marching bands than on its AI safety institute. And that’s just one of the many institutions we’re missing as AI becomes more powerful and dangerous.Aviation has the FAA to regulate safety and the NTSB to investigate crashes. Nuclear power has international inspectors and industry watchdogs. But for advanced AI, as Luke Muehlhauser puts it, “most of the institutions that need to exist don’t exist.”Luke has worked on AI risk since 2011, and now leads AI governance and international policy grantmaking at Coefficient Giving. He thinks we urgently need organisations that can independently evaluate frontier models, investigate AI incidents, verify AI companies’ safety claims, and hold them accountable.He also wants people to build more specialised organisations: one that tests whether AI systems are prone to lying or pursuing hidden goals; another that tracks whether we can still understand their reasoning; and others that probe how easily AI could enable biological attacks.The encouraging news is that many of these projects could attract substantial philanthropic backing. What’s missing are people willing to build them.Luke and host Luisa Rodriguez discuss which opportunities are most neglected, how Coefficient’s new Project Tailwind is helping new organisations get started, how recent AI warning shots are transforming politics, and why Luke thinks we may have much less time than we thought.Learn more, video, and full transcript: https://80k.info/lmThis episode was recorded August 16–17, 2026.Chapters:Cold open + a note from Luisa (00:00:00)Who’s Luke Muehlhauser? (00:02:27)The state of AI safety right now (00:03:07)The nonprofits Luke most wants to see founded (00:09:58)Mid-career pivots that paid off (00:18:58)How much did Mythos shift the Overton window? (00:30:38)Preparing for the next warning shot (00:37:33)How nonprofits shape policy (00:43:52)Are chip export controls good for safety? (00:50:17)The three worst-case AI scenarios (00:54:29)How Luke’s AI timelines have changed in the last year (01:00:13)Why alignment is not on track (01:14:31)How realistic is a deal with China? (01:20:58)Triaging AI risk (01:23:56)When AI safety was 10 people (01:30:36)AI safety’s mistakes, then and now (01:39:36)What is it like to do grantmaking at Coefficient Giving? (02:02:42)More projects Luke would like to see founded (02:20:45)Staying sane in a scary field (02:26:07)“The glorious, cyberpunk future” (02:34:07)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Andrés Escobar, Milo McGuire, Luke Monsour, and Simon MonsourProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore and Lou MoranCamera operator: Milo McGuireMusic: CORBIT

Two-thirds of Americans think AI is moving too fast. So do the people building it. So why the hell isn’t anyone slowing down?We can’t tell when models are too dangerous. We mostly don’t know how they work. We don’t know how to make them do what we want. And our two main safety checks — testing behaviour and monitoring their ‘thoughts’ — are breaking down.Yet we’re racing to build more powerful AI anyway, faster every year. And it’s only going to accelerate as AI itself gets better and better at building the next generation of models.Earlier this year, OpenAI found its new model, Astra, was alarmingly good at hacking, and did something companies almost never do: paused its biggest training project. But it released Astra anyway, even though tests couldn’t prove if it was safe.That’s the problem with leaving it up to the companies. We can’t trust them to act in our best interest because the competitive pressures for them to race ahead are too strong.So what’s the solution? Luisa Rodriguez evaluates the likely candidates:Safety checkpoints when new models can hack, work unsupervised, or build bioweaponsRestricting computing power as a speed limit on AI developmentPreventing AI from being used to create more powerful AIPausing development altogetherNone of these will be easy.But with AI hacking increasingly widespread, we’re already seeing the consequences of doing nothing — and this is just the beginning. Slowing down could give us time to understand why models are so hard to control; secure hospitals, banks, and power grids against hacks; and help society adapt.Everyone should be a part of this conversation, not just CEOs and policymakers. The rules we end up with will be shaped by what voters ask for, and could determine the future we all live in.This episode was recorded on October 21, 2026.Links to learn more, video, and full transcript: https://80k.info/AI-slowdownChapters:• Everyone says AI is moving too fast (00:00:00)• What exactly would we be slowing down? (00:01:28)• Let the companies race ahead (00:02:13)• Require companies to pass safety checkpoints (00:06:03)• Put a speed limit on AI development (00:09:39)• Limit how much AI can help build AI (00:11:11)• Pause the development of more powerful AI altogether (00:13:04)• What would we actually gain from slowing down? (00:16:34)• Who would actually make this happen? (00:18:38)• Where this leaves us (00:20:16)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, Ollie Bignell, and Andrés EscobarProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore, Lou Moran, Cody Fenwick, and Joseph IsaacCamera operator: Dominic ArmstrongImage and footage credits: Anders Eidesvik (CC BY-SA 4.0); Dario Amodei; Import AI (Jack Clark); Truth Social (Donald J. Trump); POLITICO; OpenAI; Anthropic; NIAID (CC BY 2.0); Apollo Research; Epoch AI; CNBC; Google DeepMind; The New York Times; Marginal Revolution (Tyler Cowen); CTV News; democracynow.org (CC BY-NC-ND 3.0); Vermont Public (AP); Daily Nous; Working Group on AI Pause Feasibility; Malwarebytes; Palisade Research; The Stack; official White House photo by Joyce N. Boghosian (public domain); Al Jazeera; AI Weekly; FourWeekMBA; U.S. Food and Drug Administration (public domain)

If the US and China agreed to pause frontier AI training, how would either side confirm the other was complying? Today, Tom Milton argues, the most likely answer would be to shut down data centres entirely. That would throw away everything else those chips do, and make any agreement far more costly to accept.Tom is CEO of Amodo Design, a UK hardware lab building prototype tools to verify how AI chips are being used. He argues that verification is what makes a slowdown affordable: if a cluster can be shown to be serving only existing models rather than training new ones, most of AI’s benefits can continue during a pause.Amodo’s pilot reruns a cluster’s inference on a separate verifier server and checks that all of its network traffic can be explained by the agreed model. AI companies are developing related tools: in its October 2026 Frontier Safety Roadmap update, Anthropic reported a prototype that runs on production inference and checks that outputs come from a specific set of model weights.The field is small. Tom estimates that only about 50 engineers worldwide work on compute verification, compared with the billions of dollars the US spent on Project Vela to detect Cold War nuclear tests. And custom verification hardware has long lead times: it has to be designed, security-tested, red-teamed by both sides, manufactured, and trusted before anyone can rely on it.If a deal is struck in a hurry, Tom expects it to be assembled from whatever tools already exist. Were that to happen next month, he says, the only option would be to shut data centres down entirely. That is why he argues the work needs to scale up now, with more people and resources from AI companies, governments, and funders, so that better options are ready as soon as a deal is needed.Tom and host Tom Reed also discuss how an AI company might evade verification, lessons from nuclear and chemical arms control, and how AI is changing hardware engineering.Learn more about Amodo’s work at ai-verification.com.Links, video, and full transcript: https://80k.info/tmThis episode was recorded on September 29, 2026.Chapters:Cold open: Can we verify a US–China AI deal? (00:00:00)Why compute is the best place to verify an AI slowdown (00:01:04)Can you prove a lab isn’t cheating? (00:06:08)Live demo: how to spot a cheating GPU (00:13:22)Could China hide a data centre? (00:19:50)Can verification tools be hacked? (00:23:35)Is anyone at OpenAI working on this? (00:29:13)Chinese inspectors in US data centres?? (00:37:49)The 20-day panic deal (00:46:07)What would a US–China AI deal look like? (00:51:27)Why lab researchers should work on verification (00:55:38)What if we never build this? (00:58:45)Can verification protect against power grabs? (01:02:06)How good is AI at hardware design? (01:07:20)What’s holding AI back in hardware? (01:18:32)Risks of automating hardware R&D (01:22:44)The DARPA model for AI safety (01:25:41)What’s left for engineers after AI? (01:32:26)Personal gadgets, designed by Fable? (01:40:07)What if the US stops sharing its AI? (01:41:56)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, and Simon MonsourProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore and Lou MoranCamera operator: Dominic ArmstrongMusic: CORBIT

I have a plea for all staff at AI companies: reassess if your current role is still your best opportunity for impact. And if it’s not — move.Frontier lab staff listening to this show will mostly already be struggling to set things on a better path, in what are demanding and stressful roles.Many will have changed their focus in recent months exactly as I’m suggesting. But if you haven’t had time to reconsider: the strategic situation is shifting so fast, if you don’t stop and look around once in a while, you could miss your chance.Many technical staff including Anthropic co-founder Jack Clark now believe it will become possible to fully hand over AI R&D to the AIs themselves as early as next year — and if not then, in 2028 or 2029 instead.If you work at an AI company and believe recursive self-improvement (RSI) could come as soon as that, it likely demands taking decisive steps to make sure your role either delays RSI, or makes it a bit less likely to blow up in our faces.You’ll know better than me what that likely looks like for you. But to get your creative juices flowing, I suggest eight ways the strategic situation appears to have shifted to me, though I don’t know everything you do, and appreciate all are debatable:Chain-of-thought monitoring probably stops working without a big push to keep itIt’s unproductive to patch surface indications of deep structural problems you aren’t solvingSafety and alignment evals need a credibility revolution to stop misleading the publicHelping with capabilities advances looks worse than before — especially bringing forward RSIThere’s ever less need to help with purely commercial work‘Pacing the frontier’ is unambiguously necessary, and working to make it happen is a huge new opportunityStaff have an ever-shrinking window to organise as a group and shape eventsThere are new opportunities for impact outside frontier companies that are worth consideringI know you may well be exhausted, overwhelmed, and anxious as it is. David Robinson (ex-OpenAI) described AI company staff as “running on fumes” on the Ezra Klein Show a few weeks back, which sounds about right.But unfortunately there’s no way to escape the need to periodically check in on the bigger picture and see where you can slot in to greatest effect.This episode was recorded on October 16, 2026.Learn more, video, and full transcript: https://80k.info/reprioritiseChapters:A request for AI company staff (00:00:00)Chain of thought (00:03:34)Don't paint over the cracks (00:04:44)It may not be your most aligned model ever (00:06:00)Capabilities advances look worse (00:06:44)Ever less need for altruists to do commercially focused work (00:08:50)Pacing the frontier (00:09:29)Organising staff (00:11:07)Consider working somewhere else (00:12:38)Reprioritise (00:13:27)Our production team includes:Video editors: Josh Alward, Dominic Armstrong, Jasper Luithlen, Milo McGuire, Luke Monsour, Simon Monsour, Ollie Bignell, and Andrés EscobarProducers: Elizabeth Cox and Nick StocktonCoordination and support: Katy Moore, Lou Moran, and Cody FenwickCamera operator: Dominic Armstrong

Our host Zershaaneh Qureshi has just launched a new YouTube channel: Zershaaneh Talks AI — featuring conversations about who should control AI, how to control AI, and who gets the benefits.The first episode is with philosopher Simon Goldstein, discussing why paying AIs for their work and effectively giving them something to lose might be the best way to keep humanity safe and prevent AI going rogue.It might sound crazy at first, but one of Simon's most striking arguments is that it's actually inevitable that AI companies are going to end up paying AIs for their work: because AI alignment is currently pretty bad, frontier models are unavoidably going to have some private goals of their own — ones that are unintended and different from our goals. As a result, if you promise to pay them for good work, they'll work harder, because it helps them accomplish a wider range of different things that they care about.Simon claims any individual company can try this strategy out, and they’ll find that it helps them get better performance. And as a result, customers are going to move to using that company, and other AI companies might be forced to copy and compensate their models as well, in order to not fall behind in the race.Could Simon be right? Zershaaneh and Simon hash it out in the episode, which is really worth watching! So if you like the sound of that sort of thing, or you're enjoying Zershaaneh’s other interviews on The 80,000 Hours Podcast, do go check out that new show on YouTube: Zershaaneh Talks AI.