AI's Rapid Rise: Cost, Creativity, and Geopolitics – Key Takeaways

The Naval podcast recently hosted a discussion with three founders: Gary Tan from Y Combinator, Daniel from Able Police, and Farbood from A-List. The conversation primarily revolved around the rapid advancements and implications of Artificial Intelligence (AI), touching on its impact on various sectors, societal changes, and geopolitical dynamics.

The Unavoidable Topic: AI

The founders acknowledged that AI is the most talked-about and fastest-changing topic, making any discussion on it potentially obsolete by the time it's published. However, its high impact and steep learning curve make it an essential subject for continuous discussion and note-sharing among innovators.

Gary Tan shared his personal journey, moving from not coding at all to becoming proficient enough to teach others and even creating a top 100 open-source package called GStack. He highlighted the rapid evolution of AI models, noting his quick transition from using Claude to OpenClaw.

The Economics of AI: Cost and Compute

A key insight from Gary was the idea that spending around $100,000 a year on AI tokens could enable a lifestyle akin to a "normal citizen in 2028." This is based on the expectation that token costs will significantly decrease, while compute power will increase dramatically—potentially 90,000 times within 24 to 36 months. This exponential growth in compute is seen as a fundamental driver of new AI capabilities, as each order of magnitude brings forth new functionalities.

AI's Capabilities and Human Creativity

The discussion delved into the evolving capabilities of AI. Initially, AI was seen as "mid at everything but pro at nothing." Now, it's approaching a point where it can be "pro at everything," raising questions about the "last creative mile"—the ability to move beyond recombining existing data to create something truly new. The progress AI has made in complex areas like the Erdos math problem was cited as a surprising, if not unsettling, development.

A central question emerged: will a machine with a human in the loop always be better than a machine alone? The example of "centaur chess" was brought up, where initially human-computer teams outperformed computers, but eventually, standalone computers surpassed them.

The Bottleneck: Intelligence vs. Cost

Daniel from Able Police emphasized that for practical AI applications, intelligence is no longer the bottleneck; cost is. His company, which started by converting bodycam footage into police reports, now uses a sophisticated, elastic agentic fleet that has driven down the cost of running AI models significantly. He believes that the more interesting question is what humans will do when these tasks are automated.

However, a more pessimistic view from those in AI labs suggests that AI will eventually leave "nothing left for humans to do," solving fundamental problems in science and engineering. This perspective fuels anxiety and fear, leading to discussions about potential societal unrest and even "nationalization" of AI.

The "AI Writing" Debate

A lively debate ensued regarding AI-generated content. While some argued that AI writing is often verbose and clinical, lacking the human touch, others countered that with advanced "eval harnesses" and multi-stage voice tuning, AI could produce indistinguishable and high-quality content. The core of the disagreement centered on the value of human-authored content and the potential for AI to diminish genuine human connection through communication.

Gary shared his "LSD mode" (Lateral Sarcastic Drift) system, which uses multiple vector spaces and frontier models to generate novel ideas, demonstrating AI's potential for brainstorming and creative assistance. He argued that instead of complaining about AI, people should embrace and leverage it, living in the future to understand its potential.

Open Source vs. Closed Source AI

The conversation shifted to the geopolitical implications of AI, particularly the competition between open-source and closed-source models. The observation was made that China is increasingly leading in open-source AI development, while the US focuses on proprietary models. Several theories were proposed for China's rapid advancement:

1. Independent Pre-training: China conducts its own pre-training using its own compute and data sets, unconstrained by copyright laws, allowing them to crawl vast amounts of web data.
2. Model Distillation: Accusations suggest Chinese entities are querying and distilling proprietary US models, using the generated data to train their own.
3. Algorithmic Breakthroughs and Talent: China boasts a large number of mathematicians and AI researchers, producing more STEM PhDs and Olympiad winners. Many US AI labs also employ Chinese talent, leading to a potential "AI is our Chinese against their Chinese" scenario, where knowledge transfer occurs through informal networks and job hopping.
4. Security Vulnerabilities: AI companies may lack the robust security of national secure facilities, making them susceptible to hacks and leaks of model weights, which can then be used for distillation and training.
5. Hardware Advantage: China's long-standing dominance in hardware manufacturing, coupled with the commoditization of software by AI, positions them to win if software becomes a commodity. The Chinese government actively funds labs to develop open-source AI, leveraging their hardware advantage.

The concern was raised that if AI development becomes too centralized or nationalized, it could stifle innovation and create a dangerous power imbalance. The idea of requiring models trained on open web data to become open source after a certain period was suggested as a potential solution.

The Future of Work and Society

The discussion also explored the broader societal impacts of AI:

• Job Displacement: The rapid pace of AI development is displacing white-collar jobs, including managers, academics, and journalists. However, those who embrace and use AI are finding increased productivity and new opportunities.
• Learning and Education: AI can personalize learning, allowing individuals to learn at their own pace and level, potentially making traditional education models less relevant.
• Startup Ecosystem: While AI can shrink firm sizes and increase leverage, potentially fostering more startups, the commoditization of software by AI raises questions about what new ventures will build and sell.
• Political Implications: The anxiety surrounding AI's impact could lead to political instability, with calls for nationalization or even resistance against technological progress. The idea of an "AI politician" was humorously floated.
• Human Desire and Connection: Despite AI's growing capabilities, human desire remains irreplaceable. As long as humans have desires, and AI and robots can help fulfill them, there will always be a need for humans in the loop, acting as "robot handlers" or "AI trainers."

The podcast concluded with a mix of anxiety and jubilation about the future of AI, acknowledging its transformative power while grappling with its profound implications for humanity.