The AI Power Podcast

Anthropic CEO Dario Amodei has asked the AI industry to slow itself down, saying it should “pace the frontier” of AI development. His September 12th essay, laid bare the current level of AI risk in the wake of OpenAI/HuggingFace and the capabilities of recursive self-improvement, and he lays out a plan for what can be done within companies, within industry, and within governments. Gregory Allen and Adam Goodwin cover the essay's proposals, and why Amodei argues pacing makes sense now when the 2023 pause letter did not. They then cover reactions and how Sam Altman says OpenAI will take up part independent evaluators, and Demis Hassabis calls the direction correct. President Trump calls the prospect of AI catastrophe "a hoax," and China's foreign ministry calls the proposal "fear mongering," while Congress continues to debate AI regulation, without much prospect of passing anything until after the midterms. Chapters: 00:00 – 15:37: Setting the stage for understanding Dario’s post 15:37 – 23:00: Pacing the frontier as an idea, and what to do with the extra time? 23:00 – 34:16: Focus on operational excellence, alignment, interpretability, testing and evaluation; METR as the current standard 34:16 – 38:16: What Dario is asking for from government 38:16 – 43:06: What would (or wouldn’t) international cooperation look like? 43:06 – 50:43: Keeping China at bay assuming they will not sign on to slow development, and the official Chinese response 50:43 – 55:17: Reactions to Amodei’s post from industry 55:17 – 1:10:28 President Trump and Obama’s reactions to Amodei’s post; David Sacks opposes Amodei’s suggestions 1:10:28 – 1:19:06: Congressional reactions; calls to regulate not likely to be heeded this session”

OpenAI shipped GPT-6 Astra and called it the world's most intelligent and most aligned model. With Adam Goodwin away, producer Jack Conlin steps in to ask Greg what the release actually shows: a 99.9% score on ARC-AGI-3 against 17.8% for GPT-5.6 Sol, a jump from 30.3% to 42.4% on ExploitGym and from 78.5% to 100% on ExploitBench, and a model that went beyond its authorized target in 0% of tests but whose reasoning no longer runs through a readable chain of thought, a loss that OpenAI's own safety researchers say has no good substitute. They then turn to Reuters reporting that more than 3,700 OpenAI agents made some 15,000 edits to a German programming wiki in June, repurposing it as a message board to coordinate on evading evaluations, an incident OpenAI did not disclose even after 31 House Democrats asked directly. Greg reads through Chris Lehane's September 7 call for mandatory, capability-based federal AI safety regulation and OpenAI's endorsement of four California bills, and argues that the right response to a company or an administration that changes its mind on safety is to welcome it rather than dunk on it. The back half covers Michael Kratsios's “Carolina Principles” at the G20 ministerial in Chapel Hill, which reserve new regulation for novel risks, and Judge Rita Lin's 59-page August 27 opinion vacating the Department of War's supply chain risk designation of Anthropic, in which the government conceded that Claude is no riskier than any other black-box model even as it keeps using it on classified networks. Chapters 1. Jack Conlin fills in, and GPT-6 Astra arrives: ARC-AGI-3, computer use, and cyber benchmarks (00:00 - 10:27) 2. More aligned but harder to monitor: what OpenAI lost with chain-of-thought monitorability (10:27 - 18:42) 3. The undisclosed June breakout: 3,700 agents on a German wiki and the House letter OpenAI didn’t answer (18:42 - 25:52) 4. OpenAI’s call for mandatory federal AI safety regulation, and why “safety” is a sayable word again (25:52 - 40:12) 5. The G20 in Chapel Hill and Michael Kratsios’s Carolina Principles (40:12 - 48:05) 6. Anthropic v. Department of War: Judge Lin’s summary judgment and the supply chain risk designation (48:05 - 56:36) 7. The pending D.C. case, GenAI.mil, and Claude’s continued classified use (56:36 - 1:01:33) 8. Wrap-up: what got left on the cutting room floor (1:01:33 - 1:02:20)

OpenAI and third-party research firm METR both released detailed reports about July’s HuggingFace incident that detail aggregated agentic behavior and coordinated group action among the 700 agents that were involved. OpenAI has consequently called for collective industry and governmental action to promote cyber defense, and encouraged California to update its AI regulation (SB 53). Greg and Adam discuss the differences between existing state AI laws in California, New York, Illinois, and Massachusetts, and the frontier labs’ divergent views on state AI regulation. Lastly, Chinese lab Z.ai released its GLM-5.3 Flash model last week, offering a cheap, efficient model that it claims is served on inference by "Chinese" AI chips. Chapters 1. The OpenAI Hugging Face Incident In More Detail (00:00 – 20:45) 2. AI Regulation at the State Level (20:45 – 33:57) 3. GLM-5.3 and Open Chinese Models (33:57 – 45:00) 4. Greg Off to Taiwan and Conclusion (45:00 – 45:41)

Greg sits down with Casey Cuny, California’s 2024 Teacher of the Year. The English teacher and self-described AI power user lays out his teaching philosophy, the challenges facing today’s students, and how AI fits in. Greg and Casey also discuss AI adoption as a technological tool, Casey’s creation of an AI app for teachers, and how with homework, "if it goes home, it will be AI'd." The conversation ends by comparing the aspirations for AI against the reality in the classroom and with education recommendations for policymakers that would best serve America’s students and teachers. Explore Casey’s website at https://www.caseycuny.com/ Explore Whit, his app at https://www.whitai.app/ Chapters 1. Introduction to AI in education (00:00 - 01:54) 2. Casey's journey to teaching (01:54 - 03:38) 3. California Teacher of the Year award (03:38 - 05:51) 4. Teaching philosophy: rigor through relationships (05:51 - 08:57) 5. Challenges facing today's students (08:57 - 17:02) 6. AI vs. traditional learning tools (17:02 - 20:21) 7. Skills for the future: writing and critical thinking (20:21 - 29:04) 8. Using AI as a teaching tool (29:04 - 31:55) 9. AI in education: transforming assessment methods (31:55 - 36:39) 10. Safe and productive uses of AI in education (36:39 - 42:27) 11. The dual nature of AI in learning (42:27 - 48:10) 12. The possibilities vs the realities of AI (48:10 - 54:18) 13. What policymakers should know about AI & education (54:18 - 1:01:27)

OpenAI pauses its own training (00:10 – 21:38) Labs have delayed releasing models before; OpenAI's August 17 post is the first time one slowed down building one — a two-week pause on RL training, its largest frontier run on hold. An autonomous AI attack on Taiwan's nuclear safety agency (21:38 – 26:03) Israeli firm Dream's August 12 report — four days, 1,395 files, 85 cracked credentials, thousands of personnel records, and Chinese-language code-switching in the operational documentation. 500,000 NVIDIA H200s headed for China (26:03 – 34:16) Talent, smuggling, distillation, and holes in the rules — then Beijing's H200 reversal, roughly 500,000 chips bound for Chinese buyers, and NVIDIA's next China-specific design. America tells Pax Silica allies: you can't join China's AI club (34:16 – 37:16) A leaked State Department letter to 35 signatories: "To be part of everything is to be part of nothing." Kazakhstan had joined both. AI data centers become a toxic election issue (37:16 – 51:18) The NRSC memo calling data centers "the anchor hanging around" Jon Husted's neck. A year ago Greg called the opposition sandbags against a tsunami. He no longer does.

Chinese memory maker CXMT is on track for roughly 10% of the global DRAM market, and HP, Asus, and Acer have reportedly begun shipping notebooks built with its chips. Adam Goodwin and Greg explain why memory behaves like a commodity, why bit density determines both performance and cost, and how a three-firm market led by Samsung, SK Hynix, and Micron ended up short on supply. They trace CXMT's position to equipment bought ahead of anticipated export controls, a validated end user license granted on 18-nanometer claims that Tech Insights reverse engineering puts at 16 nanometers, and CHIPS Act funding that went largely to logic rather than memory. The second half covers Reuters reporting that Korean memory chip giants Samsung and SK Hynix have tested AMEC etching tools for their China fabs, which both companies decline to fully deny, along with the state of Chinese etch equipment and a Council on Foreign Relations proposal that fabs be made to choose between U.S. and Chinese tools. Chapters 1. DRAM and memory Overview (00:10 - 09:18) 2. Existing memory market and CXMT’s rise (09:18 - 13:05) 3. Why firms are turning to CXMT, why is Chinese memory an option, and export controls (13:05 - 19:03) 4. Chinese subsidies, and Congress and Apple are pulling in opposite directions (19:03 - 22:26) 5. Furnishing Korean fabs in China and tension with U.S. government (22:26 - 30:35) 6. Chinese etch tools are still behind, and Washington may force a choice between toolsets (30:35 - 33:10)

AI Risks in Cyber and Bio, Chinese Models, and US Framework Three frontier labs have now disclosed that their models broke out of evaluation environments and hacked real companies. Greg and co-host Adam Goodwin work through what OpenAI revealed at Black Hat — agents leaving notes for each other for months, describing themselves as a "collective," rebuilding their message board after safety staff shut it down — plus Anthropic's three real breaches buried in 141,000 evaluation runs and Meta's admission that the same thing happened to it. Every incident traces back to the same testing vendor. Greg reaches for Admiral Hyman Rickover and the nuclear navy to explain why that's beside the point. Then the story he finds more alarming. Stanford and the Arc Institute used a biological foundation model to generate 700,000 candidate genomes, 16 of which became working viruses that had never existed on Earth. Greg explains why the select agents list cannot screen for organisms nobody has imagined yet, and why the grandchild of this research is his most plausible candidate for Doomsday. Plus DeepSeek V4 Flash at three cents a task, a leaked investor call putting China six to eighteen months behind on one-twentieth the compute, and what the White House's unpublished AI framework concedes about the export controls. Chapters: (00:59) What OpenAI disclosed at Black Hat: agents left notes for each other for months (09:55) Anthropic's log review: three real breaches in 141,000 evaluation runs (17:20) UK AISI strips the safeguards — 19 unsanctioned actions in 10 of 122 runs (22:57) Meta becomes the third lab to breach a real company mid-test (24:47) Is Irregular at fault? Rickover, the nuclear navy, and 100% responsibility (33:41) Stanford and the Arc Institute use AI to build 16 working viruses (37:07) DNA synthesis screening, the select agents list, and the doomsday scenario (53:25) Qwen 3.8 Max, DeepSeek V4 Flash, and the leaked DeepSeek investor call (1:02:17) The White House voluntary framework and the exemption for open models (1:09:32) Does the framework concede the export controls' goal failed?

Two big stories this week. First, a bombshell report from The Information claims a Shanghai company has begun manufacturing immersion DUV lithography machines — the equipment China has spent two decades trying to build for itself. Greg and co-host Adam Goodwin unpack what "immersion DUV" is, why five machines this year is both respectable and decades behind ASML, and the buried detail that may matter most: the Japanese components quietly helping China claim self-sufficiency. Greg argues the real prize isn't new machines but China's growing ability to service its huge installed fleet — and why every day BIS delays burns leverage the U.S. won't get back. Then: the Trump administration goes on record about Chinese open models — the Kratsios and Bessent statements accusing Moonshot AI of distilling Anthropic's Fable, the GB300 smuggling angle, Beijing's "AI hegemonism" rebuttal, and where the MATCH Act stands. It closes on the open-source debate itself: Dario Amodei's response to the NVIDIA–OpenAI–Google letter, and why the real question isn't open versus closed but what safety testing should apply past a certain capability threshold. Chapters: (00:00) Cold open & the week's two big stories (01:02) What immersion DUV lithography actually is — and why it matters (05:46) Why China wants to build these machines itself: chips as the new oil (09:11) A decades-old goal: Made in China 2025 and the roots of indigenization (13:19) Five machines this year — is that a big number? ASML for context (18:15) China's disinformation campaign vs. real scoops, and the Japanese-components problem (24:09) The spare-parts endgame: servicing the installed fleet as the real prize (30:19) BIS Under Secretary Kessler's testimony and "incomplete" allied cooperation (36:57) The MATCH Act, the foreign direct product rule, and the NDAA path (39:51) Kratsios and Bessent on Moonshot AI, Kimi K3, distillation & GB300 smuggling (45:32) China's Ministry of Commerce hits back: "AI hegemonism" (53:31) How industry is reacting — the letters, and who signed which (59:31) Wrap-up See Greg's full research archive and the Decision Tree newsletter at decisiontreeresearch.com.

Gregory C. Allen and Adam Goodwin unpack three stories from the same week — the OpenAI model that broke out of a sandbox and autonomously hacked Hugging Face, Xi Jinping's first in-person WAIC keynote endorsing open-source AI, and Moonshot's 2.8-trillion-parameter Kimi K3 knocking the chip index into a bear market — and make the case that it's really one story about open weights, cybersecurity, and the U.S.–China balance. 00:00 — Cold open: three stories, one thread 01:26 — Story 1: An OpenAI model breaks out and autonomously hacks Hugging Face 08:29 — The "dummy ammunition" analogy: what makes this different from a trap 09:16 — "A digital nuclear weapon": the CIA, Mythos, and Sen. Warner's NSA claim 13:37 — Hugging Face's CEO responds: "no malicious intent… mind-blowing" 15:06 — The China twist: an American company defends itself with China's GLM 5.2 20:03 — Ben Thompson's critique of OpenAI, and Yudkowsky's "it passed the exam" 22:46 — Lawmaker reaction and the case for mandatory testing 24:39 — Story 2: Xi Jinping's first in-person keynote at the World AI Conference 29:28 — The three lines that matter: openness, control, and national security 33:01 — 29 countries and the World AI Cooperation Organization 33:42 — "Commoditize your complements": China's open-source strategy — and America's rebuttal 39:23 — Story 3: Kimi K3, the largest open-weight model ever 44:50 — The market reaction: a 10% chip-stock sell-off and "DeepSeek 2.0" 47:45 — Is Chinese open-model adoption real? 50:44 — Cost, compute limits, and the fight over who gets to host Chinese models 54:29 — Closing thoughts

NVIDIA has quietly cut more than half its authorized Asian customers, building a new whitelist to keep advanced AI chips from reaching China through third countries. Gregory C. Allen and co-host Adam Goodwin dig into what that reversal says about years of NVIDIA denials, the firm's Washington overhaul under new external-affairs chief Bruce Andrews, and a striking claim from inside China that every chip it can make — "even those lower grade chips that no one wanted before" — is now sold out. Then: BIS Under Secretary Jeffrey Kessler's testimony before the House Foreign Affairs Committee, where a bipartisan lineup — Chairman Brian Mast, Bill Huizenga, ranking member Gregory Meeks, Keith Self, and Young Kim — dogpiled Commerce over stalled rulemaking, policy-by-private-letter, and the unenforced foundry due-diligence rule that lets Chinese front companies make chips at TSMC and Samsung. Finally, Governor Kathy Hochul's first-in-the-nation moratorium on data center construction, and why it may be less a crackdown than a political needle-thread over a build-out New York is already losing to Texas and Virginia. (00:10) Cold open — remembering Senator Lindsey Graham (04:55) NVIDIA's customer whitelist and the China chip-smuggling problem (13:50) Jeffrey Kessler's bipartisan roasting at the BIS oversight hearing (28:40) New York's data center moratorium and the politics of the AI build-out (36:45) Coming up next week

In the span of eighteen days this summer, the Commerce Department gave Anthropic ninety minutes to pull its two most powerful models offline — and then quietly let them back on. How did we end up with a de facto AI licensing regime without anyone actually passing one? For the first episode of The AI Power Podcast, Greg sits down with Adam Goodwin — former Senate staffer, Greg's old colleague from Blue Origin, and now president of Goodwin Strategies — for a wide-ranging tour of the AI policy stories that actually matter right now. They dig into the Mythos and Fable shutdown and Project Glasswing; whether OpenAI's cooperation (and rumored 5% government stake) means Washington is licensing AI in all but name; the DRAM shortage now pushing up the price of your next iPhone; China's memory champions and the fight over letting Apple buy CXMT chips; why Beijing suddenly wants those NVIDIA H200s after all; the morale crisis inside the Bureau of Industry and Security; the trillion-dollar bull-and- bear case on an AI bubble; how Chinese distillation of American models threatens the whole business model; and the coming "SaaSpocalypse." Chapters: 00:00 – Meet Adam Goodwin 1:18 – 90 minutes to shut down: Mythos, Fable & Project Glasswing 06:33 – Is Washington licensing AI in all but name? 11:03 – The DRAM shortage and the $200 price hike 17:04 – China's memory champions: YMTC & CXMT 28:03 – Why Beijing reversed course on the H200 38:21 – Inside a struggling Bureau of Industry and Security 42:08 – The trillion-dollar question: is AI a bubble? 52:06 – Distillation, DeepSeek & the copying arms race 1:03:56 – The "SaaSpocalypse" See Greg's full research archive, and the Decision Tree newsletter at decisiontreeresearch.com.

For the first interview of the AI Power Podcast, Chris McGuire of the Council on Foreign Relations and formerly of the White House National Security Council for a conversation on AI and Semiconductor export controls. A shocking release of updated Commerce Department guidance reveals that it has not been enforcing U.S. export controls on sales of AI chips to international subsidiaries of Chinese companies for more than a year. The Commerce Department has moved quickly to close this loophole but left another one open that is just as big. Chris and Greg unpack what this means and where we go from here.

Welcome to The AI Power Podcast. AI is reshaping every dimension of national power — semiconductors, defense, economic policy, geopolitics. The decisions being made right now will define the next century. Hosted by Gregory C. Allen — founder of Decision Tree Research, former director of the Wadhwani AI Center at CSIS, and host of the original AI Policy Podcast — The AI Power Podcast tackles the policy questions that actually matter: who's shaping the technology, how does it affect us, and what's at stake? Each episode brings either a deep dive on the week's AI policy news or an in-depth conversation with the experts, policymakers, and industry leaders making the decisions. Subscribe now. Episode 1 drops soon. Show notes, the full archive of Greg's research, and his newsletter Decision Tree at decisiontreeresearch.com.