Riding the AGI Wave: AI Anxiety, Open Models and the Taiwan Question with Naval

Three founders sit at Naval's table: Y Combinator president Gary Tan, Daniel from Able Police, which grew from police-report tooling into compliant AI chat, and Farbood from A-List, a health super app. The format is declared up front: nobody will pitch products; the subject is what they learned while building.

AI is, in one guest's phrase, the most perishable yet highest-impact topic. The first concrete payoff comes from coding practice: a founder who started on Claude Code switched to OpenClaw within a day and now drives his home computer from his phone. After months of three-hour nights, his verdict is that creative productivity crossed an inflection point in recent months.

The boldest number in the episode comes from token economics: spending a hundred thousand dollars a year on tokens feels like living as an ordinary citizen in 2027, and inference volume could rise ninety-thousand-fold in 24 to 36 months. That opens the Nvidia debate: the stock may be mispriced by being far too cheap, not too expensive.

The access question hardens the tone. Citing a new coding model distributed gradually to government-approved partners, the guests warn that safety rules could concentrate AI in a few hands and turn it into a defense asset. Both risks are weighed: control by a small group versus access for everyone, including the bioweapon scenario.

On COVID's origins, Naval borrows an old tech joke: designed in North Carolina, assembled in China. He criticizes the cowboy culture of function-enhancing experiments done for vaccines, while arguing that AI democratizes such knowledge but the underlying problem already exists in the field.

AI anxiety is the emotional core. The guests call AI-written emails a lower-class signal: long, clinical texts not worth reading. The advice is blunt: compress the message to its essence, or let your AI talk to their AI. Good writing is the output of good thinking, and the muscle atrophies when unused.

The counterargument gets its hearing: holding a high-bandwidth conversation with a fairly smart model all day is a new way to learn. Skill files, retrieval contraptions that collide vector spaces, and evaluation harnesses are named as the craft tools of this era. One challenge is issued: take the system, open an anonymous account, reach three hundred thousand followers, then come back and prove it.

December 2025, the Claude Code moment, is dated as the tipping point, preceded by diffusion images, chatbots and reasoning models. The refrain returns: this is the worst it will ever be. Open models such as MiniMax are described as mind-bending at a fifth to a tenth of the cost, provided they run inside a good harness.

Four theories explain China's catch-up: its own pretraining over the whole web without copyright constraints, distillation of American models (including data brokers reselling corporate-plan tokens cheaply past weak identity checks), intense researcher-hopping between labs, and weights leaking through weak security. The biggest claim is strategic: a public-private partnership in which Beijing funds labs to commoditize software and win on hardware, where Shenzhen's ecosystem rules. Software got commoditized, hardware belongs to China, and the only thing not commoditized is AI research itself.

The two-kings thesis describes the power balance: OpenAI and Anthropic pull away because they earn direct model revenue and harvest reinforcement-learning trajectories from active users; Elon is the third-ticket candidate with a public-offering war chest and data centers in space; Google takes the criticism for product sprawl, a backgrounding bug and heavy product management.

Riding the AGI is the episode's slogan: a million-token context window holds three Harry Potter books while the human mind holds seven plus or minus three items. The prescription for CEOs is total information awareness: a personal AI wired into every company system that evaluates everything and reports back. The guests predict 2027 as the year of the harness wars, adding that a two-horse race, like Apple versus Google in mobile, keeps an ecosystem healthier.

The Taiwan passage is the most contentious. The guests claim wealthy Taiwanese systematically shuttle children abroad to dodge conscription, the second party leans pro-China, and most of the island expects a Hong Kong path. Carriers are declared dead against land-based missiles and drones; a slow reunification over ten to twenty years that saves everyone's face is predicted, with tariffs and a level playing field prescribed instead of war.

The California-empire thesis is built on geography: since all of America's warm, dry Mediterranean coastline sits in one state, a third to a half of US GDP could concentrate in California. Direct democracy, where 50.1 percent can vote anything, is criticized; San Francisco is said to rebound while Los Angeles sinks, with China far ahead in high-speed rail and governing capacity.

The close covers freedom and the future: the claim that lockdowns lifted under pressure from armed state militias; the prophecy that a fallen America becomes Latin America with cartels and crime, not a European museum; universal basic robots instead of universal basic income; the caregiver shortfall, wasted healthcare spending, college reduced to credentialism; the irreplaceability of human desire; the anecdote of a customer hunting his AI sweetheart in a server box; and the closing joke that the whole podcast was AI-generated.