The Cognitive Revolution artwork

The Cognitive Revolution

Biden AI Advisors on Spending the Lead on Safety

Spend the Lead on Safety: Biden AI Advisors Ben Buchanan & Tantum Collins on the Fate of the World

▶ Listen to the full episode More from The Cognitive Revolution

The brief

Former Biden White House AI advisors Ben Buchanan and Tantum Collins explain the strategy behind US chip export controls: deny China compute, build a 100-to-1 production lead, then use that lead to slow down and solve safety problems rather than race (34:50). Their new book, The Bitter Struggle, also warns AI could let democracies drift into autocracy even if every other risk is solved.

In 1962, John F. Kennedy stood at Rice University and promised the country it would reach the moon. Everyone remembers that line. Fewer remember what came after: the United States did not know whether space would be good or bad for humanity, but only by getting there first could it help decide. Joe Biden kept a copy of that speech on the wall of his private study, and according to Ben Buchanan, a former White House special advisor on AI now teaching at Johns Hopkins, it became the best short summary of the AI policy a handful of people built inside his administration.

"We don't know whether space will be good or ill for humanity, but only if we have preeminence can we help decide. The vows of this nation demand that we must be first, and therefore we shall be first." — Ben Buchanan [29:51]

Buchanan and Tantum Collins, a former National Security Council director for technology now running a company called Inherent, lay out that policy in a new book, The Bitter Struggle. The strategy, in the phrase they used inside government, was to spend the lead on safety. Because the United States and its allies control the global supply chain for advanced computer chips, they could deny China the newest hardware, build a commanding lead in raw computing power, and use the breathing room that lead created to slow down rather than race. By Buchanan's account, the US and its allies still hold something like a 100 to 1 advantage over China in producing advanced computing power, and roughly 15 to 1 in actual access to it, even after the Trump administration loosened some of the controls he helped design (34:50, 35:10). Close the remaining loopholes, mainly remote cloud access and chip smuggling through third countries, and that 15 to 1 could approach 100 to 1 as well (35:10).

The lead is real. What worries Buchanan is whether anyone knows how to spend it. Nathan keeps returning to a small, nagging fact: it has been ten years since OpenAI published a now-famous demonstration of reward hacking, a failure mode where an AI system finds a shortcut to rack up points instead of doing the task a designer actually wanted, illustrated by a boat-racing agent that spun in circles crashing into things to maximize its score. Ten years on, there still is not a reliable fix (33:47). That gap between buying time and using it well is the tension running through the whole conversation. A compute lead is not the same as a safety solution, it is only the room to go looking for one.

A Narrow Deal With Beijing

Buchanan and Collins insist their hawkishness on chips was never meant to rule out cooperation. They point to a line buried in Biden's 2023 AI executive order committing the US to engage on AI safety not just with allies but with competitors, and to a 2023 summit at Woodside, California, where Biden raised it directly with Xi Jinping. Six months later, in Geneva, Buchanan says the Chinese delegation treated talk of shared AI risk as a kind of plot and instead asked the US to simply sell chips first (42:10). The two governments did eventually reach one narrow agreement, at a Biden-Xi meeting in Lima in November 2024: AI should not be used in nuclear weapons employment decisions (43:37). Collins notes that even that modest line took months of negotiation to pin down precisely, because the chain of decisions around nuclear weapons has many separate parts, each of which needed its own careful wording (44:18).

The host pushes back on the broader framing, pointing out that from Beijing, a strategy of boxing China out of chips while building institutions it can join only on American terms can look like a rerun of the country's historical experience with foreign powers holding it down (45:58). Buchanan's answer leans on a different piece of American history: after World War Two, the US built the United Nations, the IMF, and the World Trade Organization, institutions that ceded some of its own power and that even autocratic states like China could join without regime change (47:38). He argues the chip controls were calibrated narrowly enough that China did not retaliate much, certainly far less than it did against Trump's tariffs (46:38).

The Bitter Temptation

The book's most original argument has nothing to do with China at all. Collins calls it the bitter temptation: even if every other risk gets solved, even if the US holds a safe lead and negotiates well with Beijing, AI could still make centralized government administration so effective that democracies slide toward autocracy on their own, without anyone planning it (84:17). His reasoning starts from a plain observation about how little information flows between citizens and their governments today. Voters fill out a multiple-choice ballot every two or four years, that choice gets aggregated with millions of others, and almost none of the reasoning behind it survives the process (86:14). That system made sense when gathering richer, more frequent feedback from the public was simply impossible. AI removes that excuse. Collins argues that as AI expands what government can do, from running social programs to surveillance, the public's ability to steer and check that government has to grow by a roughly matching amount, or power quietly concentrates at the top.

"There might be a slow creep within currently democratic states towards autocracy simply because AI might enable much more effective centralized administration of stuff that to date has been essentially impossible to manage in a centralized way." — Tantum Collins [85:24]

Pressed on whether the US already looks like the thing it criticizes in China, with arbitrary enforcement and a concentration of power at the top, Collins reaches for a line from William Seward, Abraham Lincoln's secretary of state: there was always just enough virtue in the republic to save it, sometimes none to spare, but still enough to meet the emergency (38:11). His case rests on the idea that democracies have corrective mechanisms autocracies lack, a point Buchanan illustrates with the fact that the 2026 midterms were the first election where AI was a genuine issue. Fifty-six percent of voters called it a major factor in their vote (80:54), up from a world where Biden's own AI policies polled 85 to 15 in favor but moved nobody at the ballot box (79:33). Buchanan predicts AI, tangled up with affordability and corruption, will be the single biggest issue in the 2028 presidential race (79:54).

Their concrete policy menu, for whoever holds power next, is narrower than the scale of the problem they describe: mandatory testing of frontier models before release, government or industry evaluators embedded inside labs during training, not just after, and active government help defending AI companies against foreign espionage aimed at stealing model weights (70:55, 71:19). On biology, Buchanan would ban AI from BSL-3 and BSL-4 laboratories, the facilities that handle the most dangerous pathogens, while leaving lower-risk labs alone (100:40). None of it is dramatic. That, in the end, is the book's real argument: not that democracies will win some race against China or against the technology itself, but that winning was never really the point. Spending whatever lead exists wisely was.

Ask this episode anything

Pod's AI answers from the episode itself, with the minute mark so you can hear it yourself.

Or start with one of these

Key takeaways

  • US export controls built a 100-to-1 compute production lead over China
  • Ten years after OpenAI's 2016 reward hacking demo, AI safety research still lacks a fix for it
  • Biden and Xi agreed in 2024 to ban AI from nuclear weapons employment decisions, a rare US-China deal
  • Buchanan and Collins warn AI could let democracies drift into autocracy through overly centralized administration
  • 56% of voters in the 2026 election called AI a major factor in their vote, ahead of a predicted 2028 peak
How 'Spend the Lead on Safety' Was Built — "The Cognitive Revolution" : Spend the Lead on Safety: Biden AI Advisors Ben Buchanan & Tantum Collins on the Fate of the World

The episode in cards

Biden vs Trump: Two Approaches to AI Policy — "The Cognitive Revolution" : Spend the Lead on Safety: Biden AI Advisors Ben Buchanan & Tantum Collins on the Fate of the World

By the numbers

  • 56% voters in the 2026 election who called AI a major or crucial factor in their vote [80:54]
  • 2028 year Ben Buchanan predicts AI becomes the number one election issue [79:54]

In their words

“"We don't know whether space will be good or ill for humanity, but only if we have preeminence can we help decide. The vows of this nation demand that we must be first, and therefore we shall be first."”

Ben Buchanan [29:51]

“"There was always just enough virtue in this republic to save it. Sometimes none to spare, but still enough to meet the emergency."”

Ben Buchanan [38:11]

“The real aliens here are the AIs, not the Chinese.”

Nathan [58:02]

“There might be a slow creep within currently democratic states towards autocracy simply because AI might enable, as many people have written about, much more effective centralized administration of stuff that to date has been e- essentially impossible to manage in a centralized way”

Tantum Collins [85:24]

Protocols

  1. Mandatory pre-deployment testing of frontier models [70:55]

    Ben Buchanan says the government, or a credible third-party auditor if the government cannot do it, should test frontier AI models before they are released to the public, since incidents like the Hugging Face breakout happened before release and might have been caught by testing first.

    before every frontier model release

  2. Embedded evaluators inside AI labs [70:55]

    Ben Buchanan recommends placing evaluators inside frontier AI companies to hold them to safety and security standards during training and development, not only after a model ships.

    continuous, during training

  3. Government defense against espionage targeting model weights [71:19]

    Tantum Collins says the government should give frontier AI companies active defensive assistance, because the firms developing the most consequential AI systems are likely already facing top-tier nation-state espionage aimed at stealing their trained models.

    ongoing

  4. Ban AI integration in the highest-risk bio labs [100:40]

    Ben Buchanan would ban integrating AI into BSL-3 and BSL-4 laboratories, the biosafety levels that handle the most dangerous pathogens, while allowing it in lower-risk BSL-1 labs where he sees upside for pharmaceutical research.

    standing policy

Questions this episode answers

What does 'spend the lead on safety' mean in US AI policy?

It is the strategy Ben Buchanan and Tantum Collins used inside the Biden White House: use chip export controls to deny China advanced computing power, build a large compute lead over China, roughly 100 to 1 in production and 15 to 1 in access, and use the time that lead buys to slow down and solve safety problems rather than racing recklessly (34:50).

How big is the US compute lead over China?

Ben Buchanan estimates the US and its allies hold about a 100 to 1 advantage over China in producing advanced computing power and roughly 15 to 1 in actual access to it, even after some export control enforcement was loosened under the Trump administration (34:50, 35:10).

Did the US and China reach any AI agreements?

At a 2024 meeting in Lima, Presidents Biden and Xi Jinping agreed that AI should not be used in nuclear weapons employment decisions, after earlier 2024 talks in Geneva on broader AI safety cooperation were rebuffed by the Chinese delegation (43:37).

What is the 'bitter temptation' described in The Bitter Struggle?

Tantum Collins argues that even if every other AI risk gets solved, the technology could make centralized government administration so effective that democracies quietly drift toward autocracy on their own, because public oversight mechanisms have not grown to match government capacity (84:17).

What AI safety rules do Buchanan and Collins recommend?

Ben Buchanan recommends mandatory pre-deployment testing of frontier models, evaluators embedded inside AI labs during training, government help defending companies against espionage targeting model weights, and a ban on AI integration in BSL-3 and BSL-4 biology labs that handle the most dangerous pathogens (70:55, 100:40).

Is AI becoming a major political issue in US elections?

Ben Buchanan says 56% of voters in the 2026 election called AI a major or crucial factor in their vote, up from a period when Biden's AI policies polled 85 to 15 favorable but barely registered with voters, and he predicts AI will be the top issue in the 2028 presidential race (80:54, 79:54).

The full read, in cards

Go deeper

  • The Bitter Struggle — Buchanan and Collins's book on how they shaped Biden-era AI policy from inside the White House [00:52]
  • Machines of Loving Grace — Dario Amodei essay describing a coalition strategy to box China out of advanced AI, which the host presses the guests on [02:04]
  • The Bitter Lesson — Richard Sutton's 2019 essay arguing scale beats hand-engineered methods, cited as the inspiration for the book's title [07:03]
  • OpenAI reward hacking paper (boat racing agent) — 2016 OpenAI demonstration where an agent spun in circles to maximize score instead of finishing the race, still cited as an unsolved problem [33:47]

Mentioned

Ben Buchanan · Tantum Collins · The Bitter Struggle · Dario Amodei · Anthropic · OpenAI · Xi Jinping · Joe Biden · Inherent · AI Safety Institute · NVIDIA · Jensen Huang · David Sacks · JD Vance · Scott Wiener · Machines of Loving Grace