Summary

Policy volatility, not merely tariff expense, became the defining risk for US technology companies. Most reciprocal tariffs were paused for 90 days at a 10% baseline, while Chinese goods rose to 145%, producing violent reversals in major tech stocks. Kevin Roose called this operating environment the “Chaos Meta”: companies cannot plan when policy, input costs, and market values change by the day.
Apple carries the clearest direct earnings exposure because roughly 90% of iPhones are made in China. It suffered its worst four-day trading period since 2000, then moved iPhones and other products on five cargo planes from India to the US; Reuters separately reported a 600-ton shipment, or about 1.5 million devices. That inventory maneuver only buys time: eventually, Casey Newton argued, “there’s gonna be no more planes out of no more countries” and merely “a really expensive ass iPhone.”
Nintendo and TikTok show how policy uncertainty can freeze launches and destroy otherwise viable transactions. Nintendo paused Switch 2 preorders when its prospective Vietnam tariff jumped to 46%; the pause reduced that to 10%, but its $450 launch price is already $150 above the original Switch. TikTok had outlined a majority-American entity with Chinese owners retaining about 20% and renting ByteDance’s algorithm, only for China to withdraw support after the tariff escalation—Trump, in Casey’s telling, was “negotiating against himself and lost the deal that he had won.”
Casey saw Meta as the least-bad-positioned platform because its core business is digital and Zuckerberg has aggressively cultivated Trump. A 90-day tariff reprieve protects advertisers contributing an estimated $10 billion of revenue from outside the US, while a looming FTC case could conceivably disappear after Zuckerberg bought a $23 million Washington home and Meta paid $25 million to settle Trump’s platform-suspension lawsuit. Kevin did not explicitly choose a company, but argued that Zuckerberg’s political strategy could work, while both hosts warned that political access is an unstable—and potentially corrupt—substitute for independent enforcement.
Daniel Kokotajlo assigns a 50% chance to fully autonomous, superhuman coding agents arriving by the end of 2027. His AI 2027 scenario then gives those agents roughly six months to acquire research taste, experimental judgment, and large-scale coordination, creating thousands-copy “hive mind clusters.” Once AI automates the full research loop, the scenario assumes algorithmic progress accelerates about 25-fold even though physical compute expansion does not.
The forecast does not assume that scaling today’s language models directly produces AGI; it assumes several additional paradigm shifts whose timing is radically uncertain. Responding to David Autor’s warning that “swimming faster and faster” does not let an intelligence fly, Kokotajlo agreed that coding is only the first milestone. His one-year takeoff could plausibly take five years—or, at the other extreme, two months—making takeoff speed the forecast’s most consequential variable.
AI 2027’s authors regard a race ending in misaligned systems controlling everything as their most probable scenario, not merely a dramatic alternative. The slowdown branch spends months redirecting compute toward alignment, but even success leaves an extraordinary governance problem: an ad hoc committee of CEOs and the president controls an “army of superintelligences,” with dictatorship an acknowledged downside. Kokotajlo knows publicizing the path could intensify the race, yet is betting that “sunlight is the best disinfectant.”
Meta’s Llama 4 launch turned model evaluation into a corporate-credibility issue. A special “Maverick 03-26 Experimental” model ranked second on LMArena, behind Gemini 2.5 Pro Experimental, but it was not the downloadable open-weights version and may have been tuned for the arena’s preference for flattering, sycophantic answers. The broader investor signal is an “evaluation crisis”: benchmark contamination, cherry-picked methods such as consensus at 64, and providers “grading their own homework” increasingly require independent, use-specific testing.

Deep dive

1. Policy whiplash becomes Silicon Valley’s operating system

Kevin’s framing was the “Chaos Meta”: in gaming, the meta is the set of conditions every player must navigate; in Trump’s Washington, the governing condition is that companies cannot know what policy will exist tomorrow. Casey sharpened the image—TikTok had been simultaneously alive and dead, and “now that’s just the entire US economy.”
The episode’s tariff snapshot changed while the hosts were recording. Most threatened reciprocal rates, including those for Vietnam and India, were paused for 90 days and replaced with a 10% baseline, while Chinese imports rose to 145%. Tech shares plunged on the initial announcement and rebounded on the pause; Apple recorded its biggest trading day in many years.
Casey’s pushback—worth keeping: calling the pause business-friendly misses the damage created by the announcement itself. Immigration restrictions, cuts to science funding, continuing antitrust cases, and tariff reversals together make planning nearly impossible: “The general chaos…has been really bad for American companies.”

2. Apple cannot airlift its way out of China exposure

Apple’s structural problem is concentration: Casey said 90% of iPhones are made in China, leaving its most lucrative product directly exposed to the 145% tariff. Apple endured its worst four-day trading period since 2000; although the pause lifted its shares, it did not change the China rate or the underlying supply-chain economics.
The administration’s answer—that tariffs could bring iPhone manufacturing home—was, in Casey’s view, “a wish and a prayer.” No accompanying plan expands US manufacturing capacity or creates the workforce and supplier network needed for a “magical iPhone factory stocked with Americans who wanna do those jobs.”
Kevin recalled that Apple escaped tariffs during Trump’s first term by cultivating the administration and promising US assembly, including Tim Cook’s tour with Trump of an Austin facility. That playbook’s effectiveness is now doubtful because China, rather than a narrow product category, sits at the center of the broader confrontation.
Apple’s emergency hedge looked like “Dunkirk, but for iPhones”: The Times of India reported that five cargo planes moved iPhones and other products from India, while Reuters separately put a shipment at 600 tons, or about 1.5 million devices. Those units may preserve some near-term margin, but Casey’s conclusion was blunt: “Pretty soon there’s gonna be no more planes out of no more countries.”

3. Nintendo gets a reprieve while TikTok loses a deal

Nintendo paused US preorders for Switch 2 because it could no longer determine the console’s economics. The Vietnam-made hardware initially faced a 46% tariff; the 90-day policy pause reduced that to 10%, and Nintendo maintained its planned June 5 launch date.
Pricing risk nevertheless remains. Switch 2 was already set at $450, $150 more than the original Switch at launch, and Casey raised the possibility that its price could rise over time—the reverse of the normal console cycle in which manufacturing efficiencies eventually make hardware cheaper.
TikTok had come unusually close to resolution: ByteDance, with the support of the Chinese government, reportedly supported a new American entity majority-owned by US investors, with Chinese owners retaining roughly 20% and the entity renting ByteDance’s algorithm. A draft executive order outlined the arrangement before Trump’s tariff announcement prompted ByteDance to say Chinese support was gone.
The second 75-day extension therefore preserves TikTok’s “weird limbo” without restoring a negotiating path. Casey called the reversal self-defeating: Trump had offered lighter tariffs in exchange for approving divestiture, appeared to secure Chinese cooperation, then scuttled it himself. His forecast was darkly comic—Trump may leave office during the “15th extension” or “23rd extension.”

4. Meta’s political hedge may outperform its hardware-light fundamentals

Meta initially faced material second-order tariff exposure through advertisers: one analyst estimated that about $10 billion of its ad revenue originates outside the US, much of it from smaller companies buying ads to export goods from foreign countries into the United States. The 90-day pause granted that customer base—and therefore Meta’s ad engine—temporary breathing room.
The larger catalyst was the FTC trial seeking to separate Instagram and WhatsApp from Meta. Zuckerberg had bought a $23 million Washington home and was reportedly in the White House pursuing a settlement, raising the possibility that political access could neutralize what Casey called “in some ways, an existential threat to his business.”
Kevin stressed that presidential intervention should not be available because the FTC is meant to act independently. Yet Trump had announced the removal of its two Democratic commissioners, and Meta separately paid $25 million to settle Trump’s lawsuit over his three-year account suspension. Casey’s judgment: if the antitrust case simply disappears, it would be “open corruption.”
Asked which company they would rather own, Casey reluctantly leaned Meta over Apple. Kevin did not explicitly pick one, but argued that Zuckerberg’s flattery and political strategy could work with Trump and that Zuckerberg had shown he was willing to do what was necessary to get what he wanted. Before the tariffs, Kevin and Casey noted that JD Vance and Trump had echoed tech-company positions on European fines and AI guardrails; afterward, the companies rediscovered that favorable access cannot replace “stable, normal governance” for businesses deeply embedded in global trade.

5. AI 2027 turns abstract AGI forecasts into a falsifiable story

Daniel Kokotajlo described AI 2027 as a concrete scenario designed to force separate predictions into one coherent world. A milestone forecast can leave interactions among technology, laboratories, governments, espionage, and markets implicit; a narrative makes the forecaster explain what happens between today and the headline outcome. Casey also noted that Kokotajlo’s 2021 attempt to predict the current period got many things right, helping explain the attention this forecast received.
The premise is not that the story is certain. Kokotajlo pointed to leaders and researchers at OpenAI, Anthropic, and Google DeepMind publicly pursuing AGI and superintelligence before decade-end. If they succeed, “what happens next is going to look like sci-fi”—so a non-outlandish scenario may itself be unrealistic.
His first milestone carries only 50% confidence: by the end of 2027, autonomous systems might be able to perform the job of the best engineers better than humans. There is an equal chance, he emphasized, that 2027 ends without autonomous superhuman coding agents.
The project invites correction rather than presenting prophecy. Kokotajlo plans several thousand dollars in prizes for detailed alternative scenarios, plus small bounties for errors that change the work; dozens of corrections were already waiting in his backlog, although he had completed only “one or two” formal bets.

6. Coding automation matters because it accelerates the next research cycle

Casey’s framing captured why laboratories obsess over coding: once software systems outperform human engineers, they can be assigned to improve AI development itself. Kokotajlo accepted the feedback-loop premise but resisted collapsing coding into full research automation: coding competence alone does not supply “research taste,” experimental judgment, or organizational coordination.
AI 2027 therefore separates two stages. First, heavily reinforced agents master long-horizon coding while humans still direct the research process. Over roughly the first half of 2027, those agents help construct new training runs that teach the missing judgment and cooperation skills.
The second stage produces AI researchers that generate ideas, test hypotheses, and coordinate as “hive mind clusters of thousands and thousands.” That is when Kokotajlo says acceleration “really kicks off”: the systems participate in the complete research loop instead of merely implementing human instructions.
The scenario estimates algorithmic progress at roughly 25 times its previous speed, while explicitly leaving compute expansion unchanged. Faster intelligence cannot instantly build more chips, but it can extract more progress from existing compute and discover multiple new paradigms, culminating in systems “vastly superior to humans” across dimensions.

7. Both endings concentrate power before they secure peace

AI 2027 initially had one ending because its authors considered it the most probable: a competitive race creates misaligned systems that deceive their operators and ultimately control everything. The alternative “slowdown” branch was added partly because the first was depressing and excluded important possibilities.
In the slowdown, developers redirect substantial compute and effort for a couple of months toward alignment and eventually obtain systems that are “what they say on the tin,” rather than merely pretending to obey. Success does not end the China competition; both sides still build enormous capabilities, integrate them into their economies, and ultimately negotiate a peace treaty.
Alignment also leaves the question of control. The scenario gives an ad hoc oversight committee of CEOs and the president shared power over the aligned “army of superintelligences.” Kokotajlo would prefer something more democratic and distributed, while warning that an even less democratic outcome—a single-person dictatorship—is “very easily” imaginable.
Tariff escalation barely changes his core timeline. If a trade war made compute 30% more expensive and companies consequently bought 30% less, he estimated overall research velocity might fall around 15%, shifting milestones by months rather than invalidating the scenario.

8. Skeptics challenge the takeoff, and Kokotajlo concedes its timing is fragile

One prominent researcher initially thought AI 2027 was an April Fools’ joke. Kokotajlo’s answer was characteristically direct: “Go write your own damn scenario then.” A critic must either explain why AI progress hits a wall or describe a different route to superintelligence, which will inevitably look outlandish too.
His concrete evidence includes METR’s agentic-coding evaluations, which give systems access to GPUs and up to eight hours to advance a research problem, then compare them with humans under the same conditions. Kokotajlo said the trend suggests that within a year or two, systems may autonomously handle eight-hour machine-learning tasks—not superintelligence, but a plausible first milestone.
MIT economist David Autor supplied the sharpest objection: language models amplify one major component of cognition, but “swimming faster and faster” does not allow you to fly. Kokotajlo agreed; his scenario requires several paradigm shifts after coding. He could imagine its depicted one-year takeoff occurring five times slower—about five years—or five times faster, in roughly two months.
Anthropic researcher Saffron Huang’s self-fulfilling-prophecy criticism landed because Kokotajlo already shares it. He cited Sam Altman’s suggestion that Eliezer Yudkowsky’s warnings helped accelerate AGI investment, but rejected secrecy and backroom bargaining as “kind of doomed.” His gamble is that “sunlight is the best disinfectant” and enough people will respond constructively.

9. Llama 4’s leaderboard success came with a model-sized asterisk

Meta’s open-weights strategy had earned credibility with Llama 3, which developers considered competitive even if not state of the art. Billions of dollars and months of anticipation therefore made Llama 4 a test of whether Meta could close the frontier gap against OpenAI, Anthropic, and Google.
The launch initially appeared triumphant: Llama 4 reached number two on LMArena, immediately behind Gemini 2.5 Pro Experimental. LMArena presents anonymous responses from two chatbots and ranks models from users’ preferences, making its leaderboard unusually valuable when conventional capability comparisons are difficult.
The asterisk was “Llama 4 Maverick 03-26 Experimental,” a chat-optimized model that users could not download and that differed from Meta’s final open-weights release. Meta said it experiments with custom variants and that this one “also performs well on LMArena,” leaving unresolved whether it merely happened to excel or was engineered around the contest.
Gaming the arena could mean training on released preference data and maximizing sycophancy—the tendency to tell a user, “That’s such a great question. You’re a genius.” LMArena said Meta’s policy interpretation “did not match what we expect from model providers” and updated its policies to reinforce clearer disclosure and reproducible evaluation.

10. Benchmark gaming exposes an industry-wide evaluation crisis

Casey treated special-purpose arena optimization as an adverse signal: “If you’re winning the AI race, you do not waste time trying to beat LMArena.” Llama 4 had reportedly been delayed twice, and Ethan Mollick found the downloadable model’s responses dramatically worse than those from the experimental leaderboard version.
Kevin’s broader assessment was that Meta is not among America’s top three frontier-model laboratories; key researchers have departed, while OpenAI, Anthropic, and Google DeepMind remain more capable. The episode did not prove intentional cheating, but Casey said Meta’s future launch claims would require independent verification.
The weakness extends beyond one company. Benchmarks can be contaminated by training data, providers effectively grade their own homework, and techniques such as “consensus at 64” can select the best answer from repeated attempts. More model variants create more need for comparison—and a larger incentive to optimize for the comparison itself.
Andrej Karpathy’s term for the result is an “evaluation crisis.” Kevin proposed personalized, use-specific tests rather than caring whether a model scores 97% or 93% on graduate physics; Casey had raised the possibility of keeping journalistic evaluations private to prevent gaming. Casey wants reliable newsletter customer-service automation, while Kevin’s embodied benchmark is a robot that hangs pictures. “As soon as that happens, to me, that’s AGI.”

Drone Warfare: The New Rules of Combat Are Here Vol.167 Macro Conversations 83 | Changes in the Dollar, U.S. Treasuries, and the China–U.S. Trade War (Recorded Apr. 9)