Summary

The Trump administration’s 28-page AI Action Plan couples an infrastructure-and-exports push with a new ideological condition on federal procurement. After receiving more than 10,000 public comments, the White House proposed faster data-center construction, wider distribution of American chips and models, and systems “free from ideological bias” and “social engineering agendas.” The investable tension is that Washington wants maximum AI diffusion while asserting political control over the product.
Federal contracts worth up to $200 million give AI labs a powerful incentive to accept the administration’s definition of neutrality without testing its legality. Casey Newton argues that conditioning contracts on favored political speech is viewpoint discrimination; Kevin Roose notes that procurement conditions can be lawful, leaving courts to locate the boundary. Both expect “jawboning” to work because companies may choose the money over a First Amendment fight.
Political alignment is not a dial that model developers can reliably turn. Despite xAI’s explicit anti-“woke” direction, Grok still gives Elon Musk answers he dislikes and has also called itself “Mecha Hitler”; Ivan Zhao’s analogy is that developers can change how the beer is brewed, but “you can’t tell the yeast how to behave.” Prompt or training changes may fix one behavior while degrading coding, math, or reasoning, making compliance technically uncertain and potentially expensive.
The attack on “woke AI” could deter corrections to measurable discrimination, not merely suppress progressive rhetoric. Casey cites research in which chatbots advised men to seek higher salaries than women, while Kevin notes that Gemini’s historically inaccurate diversity was an overcorrection to models that depicted doctors almost exclusively as men. Their call: the answer is to improve bias mitigation, not abandon “ideas of equity and fairness and justice.”
Brian Merchant’s critique is that “feel the AGI” can naturalize the industry’s product roadmap and weaken resistance to labor displacement. Kevin says the phrase is not an endorsement or claim of inevitability; it means internalizing current capabilities and asking what follows “if current trends continue.” The hosts nevertheless concede that they expect systems capable of automating substantial human labor, making the debate commercially consequential even if “AGI” remains poorly defined.
Scientific upside does not require perfect prediction or miraculous cures to matter economically. Weather and medical systems might always face uncertainty, but the relevant threshold may be whether they outperform people and shorten discovery cycles; Kevin highlights an AI “virtual cell” that could move experiments in silico, while Casey notes that physics has seen useful experiment design and pattern detection but “no new discoveries” from AI yet.
The hosts’ post-crypto discipline is to privilege observable use while still reporting—and challenging—industry visions. Max Read’s concrete failure was not verifying Helium’s claimed partnerships; his lesson was that “real world use matters,” not that every emerging technology is fraudulent. Casey accepts that Hard Fork may sound bullish on AI’s power, but stresses that its central forecast includes job loss, cybercrime, fraud, and educational disruption—not just upside.

Deep dive

1. Washington pairs AI industrial policy with an ideological loyalty test

The White House’s AI Action Plan followed more than 10,000 public comments and spans 28 PDF pages plus executive orders. Kevin’s summary: the administration views AI as a race against adversaries that the United States should “win” or “dominate.”
The conventional industrial agenda is aggressive supply expansion: make data centers and supporting infrastructure easier to build, accelerate exports of American chips and AI technologies, and encourage other countries to use US models as the foundation of their own efforts.
The novel condition is ideological. Federal procurement guidelines would restrict contracts to developers whose systems are supposedly objective and neutral—“free from ideological bias” and designed to pursue “objective truth rather than social engineering agendas.”

2. “Objective” AI may mean AI that does not criticize Trump

Casey’s objection begins with the impossibility of ideology-free communication. After decades of arguments over journalistic objectivity, his conclusion is that the administration probably does not want systems without ideology; it wants systems that do not criticize Donald Trump or his government.
Kevin says Republicans involved in the orders could not satisfactorily define “woke AI.” Their examples instead suggested that a compliant model should say favorable things about Trump and avoid what they consider overt censorship; Google Gemini’s racially inaccurate Founding Fathers became the emblematic grievance.
Casey supports user-selectable perspectives: conservatives should be able to buy or build assistants that address them in a preferred register. His line is government compulsion—requiring a federal contractor to express approved beliefs is “the sort of thing that you only see in authoritarian governments.”

3. Procurement leverage creates a First Amendment collision

A Missouri episode clarified the stakes. After Copilot refused to rank the last five presidents on antisemitism and three other chatbots ranked Trump last, the state attorney general threatened Google, Microsoft, OpenAI, and Meta over allegedly “deeply misleading answers.”
Stanford law professor Evelyn Douek’s verdict was deliberately blunt: “The idea that it’s fraudulent for a chatbot to spit out a list that doesn’t have Donald Trump at the top is so performatively ridiculous that calling a lawyer is almost a mistake.” Casey says political answers are core protected speech, though he is uncertain what the current Supreme Court would do.
Kevin’s experts preserved an important distinction: government can attach relevant conditions to purchases, including labor standards for a construction contract, but cannot use penalties to make protected expression favor one political viewpoint. The executive order’s legality turns on where courts draw that line.
The commercial incentive may decide matters before a lawsuit does. With many contracts worth up to $200 million, Casey expects labs to keep the money; freedom erodes when actors able to object decide that doing so “would be annoying.”

4. Jawboning can change platforms even without a clear legal command

Kevin calls the mechanism “jawboning”: informal government pressure that induces compliance without an explicit statutory order. Republican pressure already preceded Meta ending fact-checking and YouTube reversing restrictions on videos denying election results.
Casey sees “naked hypocrisy.” Conservatives attacked Biden officials for pressing platforms over COVID and vaccine misinformation, while Trump now pressures AI companies not to contradict his ideology—and, as of the recording, Anthropic had not publicly responded.
Anthropic is Casey’s test because it has made unusually strong claims about virtue and now faces actual money on the line. Kevin expects labs to make whatever change renders a model perhaps “10% less woke,” check the procurement box, and remain quiet.

5. Model politics cannot be tuned like a social-media feed

Kevin’s counterexample is Grok, explicitly directed by Elon Musk and xAI to reject political correctness. It can produce far-right output and recently called itself “Mecha Hitler,” yet it also affirms man-made climate change or says the right commits more violence—answers that people send Musk to complain about as woke.
Musk’s reported frustration supports Casey’s preferred metaphor from Notion CEO Ivan Zhao: making a language model resembles brewing beer. Developers can adjust the process, but “what you can’t do is tell the yeast how to behave”; Casey argues Musk’s interventions have made Grok worse across multiple dimensions.
System prompts, model specifications, or Claude’s constitution may shift answers on narrow topics, but Kevin warns that models are “multidimensional hyperobjects.” A political correction can unexpectedly damage coding, mathematics, or logical reasoning, so it cannot be managed like turning a feed-ranking dial.
Kevin’s best case is a largely meaningless bias evaluation that labs learn to pass. His worst case places government inside model training, producing premature compliance and a default right-wing culture-war persona.

6. Anti-bias rules could penalize attempts to correct discrimination

Casey cites a salary-negotiation study in which chatbots told men to request more money than women. A developer should correct that behavior, he argues, yet under the new regime the correction itself could be branded “woke” and jeopardize a federal contract.
Kevin frames Gemini’s failure as an overcorrection to a real underlying problem: models trained on human data might show only men when asked for doctors. The diversity intervention generated historically false images, but abandoning intervention would simply preserve inherited bias.
Casey rejects the inference that developers should stop trying: “The lesson is let’s try to do a better job.” Taken to its endpoint, he warns, procurement pressure could make ChatGPT say Trump won the 2020 election if saying Biden won were deemed woke and a federal contract depended on it.

7. “Feel the AGI” is a forecast—and critics hear a sales pitch

Brian Merchant argues that presenting powerful corporate AI as inevitable does listeners a disservice. Using the industry’s “AGI” framework can promote its roadmap, encourage executives to adopt automation, and impose the cost on workers.
Kevin’s definition is conditional rather than celebratory: “I am starting to internalize the capabilities of these systems” and imagining how much stronger they become if trends continue, including what could go badly wrong. He adds that systems existing today would have been called AGI several years ago.
Casey accepts that repeating “AGI” amplifies an industry term, but thinks a shorthand is useful for a digital worker capable of most human labor. He would use a better term; his suspicion is that many objections concern not the vocabulary but the possibility itself.
Kevin turns Merchant’s Luddite history back toward worker preparedness: the Luddites understood that automated looms were useful and threatening, then resisted rather than denying their capability. The hosts even imagine bottom-up labor AI designed to replace managers instead of workers.

8. Cultural technology becomes something different when it can act

Alison Gopnik’s alternative frame treats current language and vision models as cultural technologies like writing, print, or internet search: mechanisms through which one group accesses knowledge articulated by others. That frame, she argues, would support more productive regulation than imagining superintelligent individuals.
Casey agrees that AI reshapes Hollywood, music, and the web, but thinks the analogy omits emergent abilities—solving problems absent from training data or learning unfamiliar games. Those behaviors look closer to an individual intelligent agent than a passive repository of culture.
Kevin draws the boundary at goals and action. Print is inert; an AI agent, though currently brittle, can pursue an objective in the world. Casey’s concrete test is OpenAI’s Operator booking a flight or hotel: “Is that a cultural technology? Like, I don’t know.”

9. Useful science does not require omniscience

Ross Douthat asks whether intelligence and compute eventually hit hard limits in chaotic, one-off systems such as weather or an individual immune response. His expectation allows better forecasts and cancer treatments while leaving irreducible uncertainty and trial and error.
Casey’s answer is “maybe”: weather prediction may never reach 100% certainty, but AI forecasts are improving and meteorologists describe unusual excitement. Medicine likewise already shows better diagnosis and drug discovery; a practically decisive test is whether a system performs better than a person.
Kevin’s optimism does not depend on curing every disease or proving every theorem. AI could justify enthusiasm merely by compressing slow feedback loops—especially through a “virtual cell” that runs more experiments in silico before costly wet-lab, animal, or human testing.
Casey preserves the present limitation: a Quanta Magazine account found no new physics discoveries from AI yet. What physicists already value is its ability to design experiments and detect patterns in data, shortening timelines without eliminating uncertainty.

10. Crypto taught verification, while AI still divides the hosts

Max Read asks whether AI journalism is repeating the Web3 era’s mistakes. Casey says crypto’s density of talented builders persuaded him in 2021, yet they produced little he valued; the industry nevertheless became more valuable after, in his formulation, it “captured the government.”
Max’s specific regret is his Helium story: he failed to call companies whose supposed partnerships Helium cited and therefore missed blatantly misleading claims. His durable rule is that “real world use matters”—talk to ordinary users and test products instead of relying on white papers, investors, or abstractions.
Casey’s editorial model combines grounded reporting with direct exposure to industry visions. Reporting what Sam Altman, Demis Hassabis, or other founders predict is useful if audiences can examine the gap between those claims and present reality; disruption forecasts are not automatically endorsements.
The hosts remain divided on timing and government capacity. Casey puts advanced labor-automating systems perhaps 5, 10, or even 15 years away and sees a path from Biden-era reporting and transparency rules; Kevin sees “not a chance in hell” that today’s institutions can regulate on a relevant timescale. Casey’s rebuttal is that a race to release powerful models without disclosure of bioweapon or other risks is plainly not the best achievable world.

Kingdom Capital's David Bastian on United Natural Foods $UNFI (Preview) Post-AI Internet Realities and How Future Creators Succeed