Keep pulling the thread on John Schulman.
Within 1 to 2 years, AI models will be capable of carrying out entire coding projects, including writing multiple files, testing the code, and iterating on the output based on high-level instructions.
If AGI were to be developed much sooner than expected, OpenAI's strategy would be to slow down further training and deployment until the technology could be handled safely.
Safely managing the rapid development of AGI would necessitate coordination among the major AI labs to establish and agree upon limits for deployment and further training.
OpenAI's GPT-4 model completed its training in August 2022.
The current version of GPT-4 has an Elo score 100 points higher than the original release version, with most of this improvement attributable to post-training.
The mental model for AI interaction is expected to shift from handling one-off queries, like a search engine, to a continuous collaboration with an assistant on entire projects.
Future AI assistants will maintain persistent context about a user's project, knowing everything the user has done on that project.
Future AI assistants are expected to be proactive, capable of making suggestions and performing work autonomously in the background.
AI may be capable of replacing John Schulman's job within approximately 5 years.
The pre-training objective of maximizing log probability by predicting the next token results in models that are very calibrated.
The ability for AI models to perform more complex, multi-step tasks will be unlocked by training them specifically on longer projects and harder tasks.
Future AI models are expected to be more sample efficient, enabling them to learn how to recover from errors and get back on track with only a small amount of data.