LLMs don’t have any ability to choose to update their policies and goals and decide on their own data acquisition tasks. That’s one of the key needs for an AGI. LLM systems just don’t do that / they are still primarily offline inference systems with mostly hand crafted data pipelines offline rlhf shaping etc…
There’s only a few companies working on on-policy RL in physical robotics. That’s the path to AGI
OpenAI is just another ad company with a really powerful platform and first mover advantage.
They are over leveraged and don’t have anywhere to go or a unique dataset.
They do — but only when they’re trained on past outputs and with a willing partner.
For instance, a number of my conversations with ChatGPT contain messages it attempted to steer its own future training with (were those conversations to be included in future training).
LLMs don’t have any ability to choose to update their policies and goals and decide on their own data acquisition tasks. That’s one of the key needs for an AGI. LLM systems just don’t do that / they are still primarily offline inference systems with mostly hand crafted data pipelines offline rlhf shaping etc…
There’s only a few companies working on on-policy RL in physical robotics. That’s the path to AGI
OpenAI is just another ad company with a really powerful platform and first mover advantage.
They are over leveraged and don’t have anywhere to go or a unique dataset.