AGI, Open Source, Taiwan, and Why California Runs the World
Three founders discuss AI's explosive trajectory (90,000x compute in 3 years), the collapse of proprietary model advantages as open source catches up, China's structural dominance in hardware and AI talent, why defending Taiwan is strategically incoherent, and the urgent need for AI-augmented leadership to fix broken institutions like Google and Meta.
AI Scaling and the New Frontier
Token costs collapsing, compute exploding
Spending $100,000 per year on AI tokens now lets you live like a normal citizen in 2028. Inference compute will increase 90,000x in 24–36 months, driven by chip and data center buildout. Each order of magnitude unlocks new capabilities.
From mid-level to pro-level AI
Two years ago, AI was mid at everything but pro at nothing. Now it's pro at everything, but the open question is whether it will achieve the last creative mile—moving beyond recombination of training data to genuine novelty.
Recursive self-improvement and ASI
Lab researchers believe scaling laws will produce AI smarter than the smartest humans through recursive self-improvement (ASI). This remains the most important unanswered question in AI development.
Cost is the bottleneck, not intelligence
In practical AI deployment, cost is the limiting factor. One founder reduced per-person AI costs from $100/month to $2.84/month by building an optimized stack with an eval harness and elastic agentic fleet.
Open Source vs. Proprietary: The Convergence
Open source catching up in months, not years
The gap between frontier proprietary models and open source has compressed from 12 months to 9 months to 6 months, with some claiming 3 months. Performance is jagged—caught up in some domains, lagging in others.
Open source models are 5-10x cheaper
Models like Llama and Mistral are 5–10x cheaper than frontier models like Claude and GPT-4, though they require a good harness to perform well.
Why China dominates open source
China has its own pre-training, compute, and datasets. It can crawl more data due to fewer copyright restrictions. Chinese researchers are distilling American models, and the US AI companies lack security to prevent weight leaks. China's strategy: subsidize open source so their hardware dominance stays competitive.
Distillation and data broker leaks
Anthropic accuses China of distilling their models by querying them at scale. Meanwhile, data brokers resell Claude API tokens at 80% discount by exploiting subsidized enterprise plans. KYC (know-your-customer) rules are easily bypassed.
Once open source leads, it rarely loses
Historical precedent (Linux, other open projects) shows that once open source takes the lead, an ecosystem springs up and proprietary competitors struggle to catch back up, especially if they have limited resources.
AI Anxiety and the Speed of Transition
AI anxiety is real and spreading
People are getting angry and scared about AI replacing jobs. This fear is driving irrational responses like attacking data centers. The real issue is not water or environmental concerns but the fear of replacement and job loss.
The speed of transition is the problem
When agriculture mechanized, 50% of the US labor force moved off farms over 60–70 years without mass unemployment. AI displacement is happening much faster, which is more tumultuous and disruptive.
Most corporate jobs are already make-work
Fortune 500 companies are full of make-work jobs. The real productivity issue is that most people are not actually working; they are in bullshit jobs. AI will expose this.
AI users are more productive, not displaced
People actively using AI are working harder and more productively than ever. Displacement only happens if you refuse to use AI. The leverage and productivity gains are real.
AI Writing and Human Authenticity
AI-written emails are a low-class signal
When emails are clearly written by AI, it signals disrespect for the reader's time and is seen as a lower-class move. Good writing requires human thought, rumination, and respect for the audience.
AI writing will be indistinguishable in 9 months
With enough eval harnesses, cross-modal evaluation, and skill file tuning, AI-generated writing will be indistinguishable from human writing within 9 months. The conversation about AI writing quality will disappear.
Code is utilitarian; prose is human
Code is meant to be consumed by computers, so AI-generated code is fine. Prose meant for humans should be written by humans or heavily edited to respect the reader's time and convey authentic thought.
AI as brainstorming partner, not writer
AI is useful for brainstorming, generating synonyms, pulling tangential information, and idea generation. But the final output meant for human consumption should be crafted by a human.
Geopolitics: China, Taiwan, and US Decline
China owns hardware; the US owns software (for now)
China dominates hardware manufacturing and will for the next decade. The US leads in software and AI research. But AI is commoditizing software, which erodes the US advantage. China's strategy is to let open source commoditize software while they control the hardware supply chain.
Taiwan defense is strategically incoherent
Aircraft carriers are vulnerable to land-based missiles. The US lacks manufacturing capacity and would run out of missiles in 7 days. Most Taiwanese are not interested in fighting; they expect to reunify with China over time. The US cannot realistically defend Taiwan.
Taiwan's draft evasion system
Rich Taiwanese families dodge mandatory conscription by keeping their children out of the country for 3 months per year for 5 years, making them ineligible. This shows low appetite for conflict.
No rational reason for US-China war
The US is in low growth, high inflation, internal political division. Picking a war with China is strategically stupid. China is internally focused on staying in power. Both sides benefit from trade, not conflict.
China is hyper-competent; California is dysfunctional
The Chinese Communist Party (really a capitalist-fascist organization) is hyper-competent at building infrastructure and redirecting capital. California, by contrast, is a direct democracy where 50.1% can vote for anything, leading to chaos and dysfunction.
California: Empire and Opportunity
California has monopoly on best US land
California has all the warm, dry Mediterranean coastline in the US, plus arable land, natural resources, and good weather. This should be 5–6 states. Instead, it is one state with direct democracy, leading to mismanagement and dysfunction.
30-50% of US GDP will concentrate in California
Due to geography and natural resources, California will capture a disproportionate share of US economic output, making it a de facto empire within the US.
San Francisco is on the rebound
After years of decline, San Francisco is recovering. New York friends are moving back. The city is regaining its position as a tech and culture hub.
LA's decline is real but creates opportunity
LA has declined (sinking, Hollywood exodus), but this creates opportunity for reinvestment and rebuilding. The problem is opportunity.
Fixing Broken Institutions with AI
Google needs to fire 80% of employees
Google is bloated with make-work jobs and product overlap. The company needs to use AI to identify which 80% of employees to cut, then rebuild with a lean, high-leverage team.
Meta's data-labeling project is too blunt
Meta's plan to turn engineers into data labelers is indiscriminate and demoralizing. Instead, use AI to evaluate each person's actual contribution and performance, then make targeted decisions.
CEO brain: total information awareness
CEOs should use AI to have total information awareness of all teams, KPIs, and conversations. This allows the CEO to make informed decisions and concentrate power in the founding team.
1 million token context is a game-changer
A 1 million token context window (roughly three Harry Potter books) allows an AI agent to understand and evaluate an entire company's operations, documents, and communications. This is a massive advantage for decision-making.
The Future: Robots, UBI, and Human Purpose
Universal Basic Robot (UBR) over UBI
Instead of universal basic income, provide universal basic robots. Everyone gets a robot to cook, clean, and handle household tasks. This is more tangible and empowering than cash.
Robotics is 2-5 years away
Robotics boosters say 2-3 years; skeptics say 5-10 years. No one thinks it is impossible. Self-driving cars and robots already exist.
Humans as AI handlers, not displaced
If AI reaches expert level but not ASI, humans remain the motivated agents. People become AI handlers and robot trainers, guiding and directing AI systems. This is the optimistic scenario.
Human desire cannot be replaced
Even if robots have desires, they cannot replace human desire. As long as humans have wants and AI can help fulfill them, humans stay in the loop.
College is about networks, not learning
College is 90% socializing, babysitting, and credentialism, with a tiny bit of education. The real value is the network of people and the four years in an idyllic setting.
Healthcare and education are mostly waste
90% of healthcare is wasted (e.g., keeping people alive for the last two weeks). Education is a credential factory. Both can be disrupted by AI and better systems.
The Startup Ecosystem and AI Harness Wars
2027: the year of AI harness wars
The key battleground is not models but harnesses—the infrastructure and tools that make AI useful. The question is which harness (Claude Code, OpenClaw, Hermes, etc.) becomes the standard.
General models beat specialized ones
Frontier general models (Claude, GPT-4) outperform specialized models trained on specific domains (e.g., legal AI). This makes it hard for vertical SaaS to compete.
Two competing platforms are better than one
The mobile ecosystem thrived because iOS and Android competed. If AI consolidates to one provider, startups face a true monopolist. Diversity in models (OpenAI, Anthropic, open source) is crucial.
Errors compound in recursive loops
If an AI is 90% accurate and runs recursively 100 times, errors compound. A 99.9% accurate model drops to 80-90% accuracy; the 90% model drops to 13%. Intelligence matters on the margin.
COVID Origins and Bioweapon Democratization
COVID was US-China collaboration
Fauci and the NIH in North Carolina conducted gain-of-function research to build a vaccine. The virus was assembled in China. The joke: designed by NIH in North Carolina, assembled in China (like Apple products).
AI democratizes bioweapon creation
AI could make bioweapon creation easier, but smart biohackers can already do it. AI is democratizing a capability that was already possible. The game theory of mutual destruction is not good.
Notable quotes
If you're just willing to spend $100,000 a year on tokens, you can basically live like you are a normal citizen in 2028. — Gary Tan
The most interesting remaining question is whether AIs will keep getting smarter until they become smarter than the smartest humans through recursive self-improvement. — Naval
This is the worst it'll ever be. It's getting better. — Gary Tan