I prefer Gwern's critique of "Tool AI" as a path toward AGI, similar to her 5th point: https://www.gwern.net/Tool-AI https://www.gwern.net/Tool-AI The recent i

I prefer Gwern's critique of "Tool AI" as a path toward AGI, similar to her 5th point: https://www.gwern.net/Tool-AI https://www.gwern.net/Tool-AI The recent interview with John Carmack on Lex Friedman also made me hopeful that at least people in the field (of AGI research) are working of the issues with the current approach of strapping models together directly on the required input, instead of agents in a virtual world (or worse yet, with some connection to the outside world) that's running as fast as possible. Gato's [1] more general approach of using different inputs and actions in the same format is probably a good path towards this kind of agent-in-a-world AI, which is much closer to how a natural organism interacts and lives: serially, always-on, and in a universe. [1] https://www.deepmind.com/publications/a-generalist-agent https://www.deepmind.com/publications/a-generalist-agent