The Formative Years and Early Challenges of OpenAI

Key Insight

OpenAI commenced operations chaotically on January 4, 2016, initially from Greg Brockman's San Francisco apartment due to an absence of office space. Key figures like Ilya Sutskever, described as an 'artist who expresses himself through machine learning' and a 'musician whose preferred instrument was the whiteboard', struggled without proper facilities. Brockman, the fourth employee at Stripe, shifted his career focus 'from ego to cause', undertaking tasks like washing cups to contribute to AGI development. The team, aiming to pursue DeepMind, initially lacked a clear direction for their work.

Inspired by DeepMind's Atari-playing agent, which combined deep neural networks with reinforcement learning, OpenAI's early efforts focused on similar approaches. Andrej Karpathy proposed expanding AI agents from 1970s Atari games to current computer screens, leading to the ambitious 'Universe' project. Zaremba suggested building 'Gym', an infrastructure for reinforcement learning research, driven by a strong desire to 'do something significant quickly' and 'prove ourselves' to avoid evaporation of the initiative. Brockman learned crucial lessons about collaboration between AI researchers and software engineers during the Gym project, finding the slow pace painful and eventually taking over to optimize their work.

Despite early projects, OpenAI faced significant challenges. DeepMind's AlphaGo achieved a major breakthrough in October 2015, defeating a former European Go champion, and then the world champion Lee Sedol in March 2016, watched by 200 million people, demonstrating that AGI might not be far. This victory intensified pressure on OpenAI. However, their efforts, particularly Karpathy's plan to train an AI agent to click a mouse or use a keyboard for rewards, failed. They consumed an 'ungodly amount of compute' but saw no improvement, realizing their models were not intelligent enough to tackle complex 'exploration problem[s]' directly. Brockman concluded they were 'targeting this castle in the sky, when what we needed was a little shack', prompting a shift to smaller teams and diversified goals.