28 months of AI lessons in 32 minutes

Despite significant investment and some speculative behavior, AI is not a bubble due to its demonstrable utility, massive revenue growth from key companies, and advancements in reinforcement learning, but the industry faces challenges with compute bottlenecks and a lack of innovation in many startups.

• AI is fundamentally transformative, showing real utility and revenue growth, not a speculative bubble.

• Reinforcement learning and specialized smaller models are key to future AI advancements and applications.

• Compute is the primary bottleneck for AI, and a lack of true innovation marks many current startups.

AI is not a bubble, characterized by genuine utility and unprecedented revenue growth from leading companies, contrasting sharply with past speculative markets. While significant private investment in unproven startups indicates some bubble-like behavior in the venture capital world, the fundamental technology and its applications are robust. The industry is rapidly evolving with open-source models challenging closed-source counterparts and a shift towards smaller, specialized, and faster AI models.

AI Bubble vs. Reality

• 00:00:17 The current AI surge is not a bubble, despite intense investment in both established companies like Google and OpenAI, and early-stage startups with unproven products. Unlike the 2021 crypto bull run, AI offers immediate, tangible use cases and leading AI companies like Anthropic are demonstrating massive, unprecedented revenue growth, which indicates genuine value rather than pure speculation. While a short-term market pullback is possible, an 80% crash like the dot-com bubble is unlikely given AI's transformative nature.

Reinforcement Learning Gains

• 00:04:34 Significant advancements in reinforcement learning (RL) are driving AI progress, especially when combined with test-time compute. This frontier allows AI models to achieve and master specific benchmarks, particularly in deterministic tasks like coding and math where synthetic data can be effectively generated and validated. Companies are now creating specialized RL environments for AI agents to perform complex actions, providing new data sources that are crucial for continued model improvement beyond the saturated internet text data.

Emergence of Open & Smaller Models

• 00:13:11 Open-source AI models are rapidly catching up to, and in some domains, surpassing closed-source models, although this trend receives less publicity due to commercial incentives and inference optimization challenges. Simultaneously, there is a growing trend towards smaller, more specialized, and faster AI models, such as Anthropic's 'Haiku' series. These models, while not necessarily more powerful, are more useful due to their lower cost and increased speed, enabling wider application and improving user productivity by maintaining a 'flow state'.

Innovation & Compute Bottlenecks

• 00:22:54 The AI industry is plagued by a lack of true innovation, with many startups creating identical 'vibe coding tools' or 'workflow builders,' leading to a 'winner-takes-all' market. Many of these startups are unprofitable, reselling tokens at a loss. The biggest bottleneck remains compute power, with massive investments pouring into building multi-gigawatt data centers to meet insatiable demand. This focus on infrastructure and the predictable revenue of data centers attracts smart money from institutional investors, while a looming wave of job displacement from AI-powered automation is predicted to cause social unrest.