In Minsky’s “Steps Towards Artificial Intelligence”, Planning is the second-last stage. The final stage is Induction, by which he means, making its own models of the world.
As far as the current era of AI goes, you could say we saw the first signs of Planning in the primitive LLM-based agent ChaosGPT. It wasn’t very good at planning, but it did talk to itself about which courses of action to take.
Apart from the method of adding planning “scaffolding” to a transformer LLM, there is the rumor that Google’s Gemini combines the Monte Carlo Tree Search method of policy optimization, used in AlphaGo, with a transformer-like architecture.
I think next year’s AIs will probably be good at planning, and I’ll stick with my timeline, 0-5 years to superintelligence.
In Minsky’s “Steps Towards Artificial Intelligence”, Planning is the second-last stage. The final stage is Induction, by which he means, making its own models of the world.
As far as the current era of AI goes, you could say we saw the first signs of Planning in the primitive LLM-based agent ChaosGPT. It wasn’t very good at planning, but it did talk to itself about which courses of action to take.
Apart from the method of adding planning “scaffolding” to a transformer LLM, there is the rumor that Google’s Gemini combines the Monte Carlo Tree Search method of policy optimization, used in AlphaGo, with a transformer-like architecture.
I think next year’s AIs will probably be good at planning, and I’ll stick with my timeline, 0-5 years to superintelligence.
I think Minsky got those two stages the wrong way around.
Complex plans over long time horizons would need to be done over some nontrivial world model.