Episode Details

John Schulman is a research scientist at OpenAI. He co-leads the Reinforcement Learning group and works on agent learning in virtual game worlds (e.g., Dota) as well as in robotics. John, Corey, and Steve talk about AI, AGI (Artifical General Intelligence), the Singularity (self-reinforcing advances in AI which lead to runaway behavior that is incomprehensible to humans), and the creation and goals of OpenAI. They discuss recent advances in language models (GPT-2) and whether these results raise doubts about the usefulness of linguistic research over the past 60 years. Does GPT-2 imply that neural networks trained using large amounts of human-generated text can encode “common sense” knowledge about the world? They also discuss what humans are better at than current AI systems, and near term examples of what is already feasible: for example, using AI drones to kill people.

Resources

• John Schulman
• OpenAI
• Better Language Models and Their Implications (GPT-2)
• Transcript

Announcing this for some friends at Mechanize - a startup that builds environments for training and evaluating frontier LLMs. Its customers include the top AI labs, and it has contributed to the breakthrough in coding capabilities of frontier models. 

Mechanize is hiring! 

https://mechanize.work/b/hsu 

Compensation is extremely competitive. For technical roles, $300-500k. They are also seeking smart generalists. 

For example: 

Research Engineer, Alignment: Build evals that test for misaligned model behaviors  $500K salary 

Puzzle Maker: Design interesting and original puzzles that LLMs can’t yet solve  $300K salary 

Mechanize understands that my readership is highly selected. There is a VERY GOOD CHANCE you will be interviewed if you apply via the link above.