Summary
Highlights
Will recounts his economics background at Cambridge and his initial venture into algorithmic trading using machine learning for sports predictions. After 10 successful years, he pivoted to focus on large language models (LLMs), recognizing their disruptive potential in consumer technology. He noticed the difficulty for individuals to access and run powerful LLMs from giants like OpenAI and Google, leading him to envision a platform for specialized AI models, similar to the fragmented financial market.
Will argues that current AI models function more like search engines than true reasoning engines, optimized to human preferences rather than deep understanding. He uses an analogy of a child's IQ to describe the progression of GPT models, predicting they might top out around a 15-year-old's intelligence. He highlights the common trend of AI performance reaching a plateau around human capabilities, with the final "5%" of performance proving incredibly challenging due to edge cases.
Will shares insights into building his successful $450 million startup, Chai AI, with only 12 people. He attributes this to his experience in competitive finance, where being among the top performers is crucial for success. He emphasizes the exponential cost of complexity in software development and the importance of hiring creative, intelligent individuals with high standards who can reduce complexity rather than add to it. He contrasts startups with large tech companies, highlighting the need for startups to take risks and innovate.
Will discusses the balance between moving fast and breaking things, a common startup mantra, and the need for discipline. He argues that startups must take risks to innovate, unlike large, established companies that prioritize stability. However, he also stresses the importance of rigorous experimentation, good data collection, and refactoring code to manage complexity and ensure scalability. Without discipline, risk-taking can lead to chaotic outcomes or transient success.
Will explains Chai AI's mission: to serve as a platform for aggregating specialized language models. Initially a consumer app with its own optimized LLM, Chai pivoted to support open-source models. Through Chiverse, their Python package, researchers and data scientists can submit models, receive user feedback, and refine them. The best models are financially rewarded, with Chai aiming to redistribute a third of its revenue to creators. The goal is to combine a diverse array of unique LLMs into a single meta-model that outperforms monolithic AI. He emphasizes that fine-tuning and prompt engineering add more value than simply scaling model size.
Will discusses the value chain in language models, arguing that while base models are important, the real value lies in data curation, fine-tuning, and prompt engineering. He foresees a market for data retrieval, where content providers like The New York Times can monetize up-to-date data feeds for LLMs. He stresses the need for talented engineers to continue to build and innovate for Chai AI, particularly in research, Python development, and backend roles to leverage human resources in advancing AI.