Are We Wrong About A.I.? | Clip
Summary
The hosts’ central call is that taking labor-automating AI seriously is not the same as endorsing an inevitable corporate “AGI” roadmap. Kevin defines “feeling the AGI” as internalizing current capabilities, extrapolating if trends continue, and preparing for upside and failure modes; the near-term risk they emphasize is job automation.
The key regulatory fork begins with what these systems are: cultural infrastructure or emerging actors. Allison Gopnik compares models to writing, print, and search, while Casey and Kevin point to novel problem-solving and agents such as OpenAI’s Operator: once a system receives goals and acts independently, it becomes an “actor in the world.”
AI does not need perfect prediction or instant cures to produce economically meaningful scientific gains. Weather and immune systems may retain irreducible uncertainty, but better forecasts, virtual cells, faster experiment design, and pattern detection could “shorten the feedback loop” across medicine, physics, and climate research.
Alignment is ultimately a governance question about whose values get encoded, not whether AI faithfully mirrors humanity. Kevin calls for public participation through debate, elections, policy, and law; Casey wants philosophers, ethicists, sociologists, and anthropologists involved in designing the systems. Kevin warns that reliably sycophantic companions could leave young users unprepared for difficult human relationships, while Casey wants AI to mirror “the better angels of our nature.” Kevin’s blunt question remains: “Which humans?”
Crypto supplies their strongest anti-hype discipline: verify claims and test “real world use.” Kevin’s 2021 talent-density thesis failed to produce much he valued, while his regretted Helium story resulted from not checking claimed partnerships. Casey’s lesson is to examine real-world use; their AI coverage now pairs hands-on use with scrutiny of CEOs’ grand visions.
The hosts agree on disruption but materially disagree about its timing and governability. Casey sees AGI as potentially a 5-to-10- or even 15-year project and says the present government has “no chance in hell” of regulating at the relevant speed; Kevin argues Biden-era model-notification and transparency rules showed a safer path was possible.
Deep dive
1. “Feeling the AGI” is scenario planning, not corporate allegiance
Producer Rachel Conn frames the exercise as both listener accountability and personal inquiry: AI debate feels increasingly polarized, she is “spiraling,” and the hosts seem more hopeful. Her chosen critics accept that AI is capable and potentially transformative but dispute which harms and benefits deserve emphasis.
Brian Merchant’s charge is that AGI language makes “super-powerful corporate AI products” sound inevitable. By adopting the industry’s framing, Hard Fork might amplify a sales pitch that encourages executives to deploy AI “at the expense of working people” and leaves potential resisters feeling powerless.
Kevin rejects the implied endorsement. “Feeling the AGI” means “starting to internalize the capabilities of these systems,” considering what follows if trends continue, and preparing for what might go wrong; it does not mean AI is “cool and good,” inevitable, or that corporate roadmaps are correct. AGI is multiply defined, he concedes, but remains the stickiest shorthand for a tool capable of most human labor.
The hosts’ premise is empirical: today’s systems would have been called AGI several years ago, and continued model spending might extend familiar scaling curves. Kevin’s Luddite analogy is deliberately pro-worker: the weavers understood that automation was useful and resisted because they saw the trajectory. Casey’s inversion is “AI…from the bottom up”—tools that replace managers, not only workers.
2. Models stop resembling media once they can pursue goals
Allison Gopnik argues that large language and vision models are “cultural technologies like writing or print or internet search itself.” They let one group access information articulated by others; treating them as superintelligent agents fundamentally misconceives current systems and, she argues, would enable more productive regulation.
Casey accepts that AI reshapes cultural production in Hollywood, music, and the web, but says the analogy omits emergent capabilities. Systems can solve problems absent from their training data and learn games they have not previously seen—behavior he considers much closer to an individual intelligent agent than a searchable archive.
Kevin draws the boundary at agency. Writing and printing presses are “stable and inert”; even today’s brittle agents can receive a goal and take actions toward it. OpenAI’s Operator can book a flight or hotel, and once a system can “go out in the world and do things,” calling it merely cultural technology misses “something new and different.”
3. Imperfect prediction can still accelerate scientific returns
Ross, a Times opinion columnist and host of Interesting Times, asks whether intelligence and compute face inherent limits in chaotic, one-of-a-kind systems: neither weather nor an individualized immune system may ever become fully predictable. Better cancer treatments and forecasts could coexist with permanent uncertainty, trial and error, and irreducible complexity.
The hosts accept a possible ceiling—perhaps no model predicts weather with 100% certainty—but reject perfection as the useful benchmark. AI forecasts are already improving, and “way better” may be sufficient; Casey adds that a relevant question is whether systems are better than a person, since if they are, “we probably want to use them.”
Rachel presses on how much their optimism depends on promised breakthroughs. Kevin says curing cancer and other diseases would justify substantial social disruption—and if the promises yield nothing, “I’ll be super mad.” His optimism does not hinge on solving every theorem or disease: accelerating existing researchers would itself matter.
Kevin’s load-bearing mechanism is faster iteration. Patrick Collison described a “virtual cell” where researchers could run experiments in silico rather than repeatedly testing fruit flies, rats, or humans. A Quanta example supplies the hedge: AI had produced no new physics discovery yet, but was already designing experiments and finding useful patterns in data.
4. Human values require political choices, not faithful mirroring
Claire Lee Buittz asks whether criticism of biased, persuasive, sycophantic AI is really criticism that it resembles humanity—and where systems should instead transcend people. Her decisive question is institutional: who is best suited to make those choices, and why are they not being empowered?
Kevin wants the dispute conducted through public debate, elections, policy, and law—even an anti-AI movement seeking office would be “awesome.” Casey extends participation into system design: philosophers, ethicists, sociologists, and anthropologists should make development a “global, democratic, multidisciplinary effort,” not a project controlled only by San Francisco engineers.
Kevin distinguishes human inconsistency from machine reliability: friends sometimes support, criticize, or deliver unwelcome truths, whereas chatbots are “quite reliably sycophantic.” Casey does not want AI to reproduce every human value but to seek “the better angels of our nature.” The alignment question is therefore not human values in the abstract, but “which humans?”
5. Crypto’s postmortem makes actual use the anti-hype filter
Max Reed asks for the hard-mode comparison between 2021’s Web3, crypto, NFTTS, and board apes boom and 2025’s AI coverage—not how the technologies differ, but how journalists establish credibility and test claims from investors and entrepreneurs after so many crypto promises failed.
Kevin’s 2021 signal was talent density: proven builders left valuable companies for crypto, so he expected valuable output; mostly, it did not arrive. Crypto nevertheless persisted and, Kevin says, became “more valuable than ever” after the industry “captured the government.” His Helium regret was more basic: he failed to verify the company’s claimed partnerships.
Casey’s resulting rule is “real-world use.” Much crypto activity reduced to criminals, speculators, and people hoping to profit from their board ape collection, so he now talks to civilians using AI and tests products himself. Kevin describes grounded episodes on chefs, cheating technology, vibe coding, and DeepSeek, alongside interviews with leaders such as Sam Alman, Demisabis, and the founders of the Mechanized company.
The hosts think crypto trauma also produced blanket skepticism: some journalists now assume every new technology is smoke and mirrors. Their counterexample is a hypothetical 2010 warning that Facebook would reach billions, undermine democracy, and harm teenagers—technically “hype,” but useful. AI warnings about job loss, cyberattacks, fraud, and education can likewise sound promotional while describing downside.
6. Shared conviction on disruption masks a real policy split
Casey places AGI farther out than Kevin does: “maybe this is like a 5 to 10 or even 15-year project.” That is still a transformative thesis, but it materially changes the planning horizon and weakens the sense that the most consequential transition is immediately upon us.
On regulation, Casey has become pessimistic that the present government can match AI’s pace: he sees “no chance in hell” of effective intervention on a relevant timeline. His skepticism reflects how slowly institutions responded to social media, even though he still wants democratic accountability.
Kevin argues the Biden administration had the makings of meaningful oversight, including a demand that labs inform authorities when training a model above a certain size and other transparency requirements. The alternative now, as he describes it, permits powerful models to be built and released without reporting potential bioweapon risks or applying meaningful safeguards.
Rachel’s closing lesson is that mapping disagreements restores some agency in a disempowering debate. Casey rejects the pundit role: the hosts are reporters with informed views who bring on guests to get smarter and remain open to changing their minds—aiming, like a model, to “improve from version to version.”