LLMs show that a lot of human intelligence comes from (and is encoded in) our linguistic abilities, but it's still missing really important context that forms a hard ceiling on its performance compared to a sentient agent - specifically an awareness of time, its environment, other agents and long term memories.
Although at this point it feels like these are just engineering problems as opposed to deep philosophical questions. The capabilities of ChatGPT are emergent phenomena created from the extremely simple training task of next word prediction. IMO this is very strong evidence that the rest of our cognitive abilities can be replicated this way as well, all it takes is the right environment and training context. It might start with something like this: https://www.deepmind.com/blog/building-interactive-agents-in... that uses cross-attention with an LLM to predict its next actions.
Some speculative ideas I've had:
- Brains (in animals) have largely evolved to predict the future state of the environment, to evade predators, find food and so on.
- To be effective, this predictive model must take its own (future) actions into account, a requirement for counterfactual thinking.
- This means that the brain needs a predictive model of its own actions (which does not necessarily align with how the brain actually works)
- Consciousness is the feedback loop between our senses (our current estimated state) and this predictive model of our own actions.
- All of this is to better predict the future state of the environment, to aid in our survival. For a hypothetical AI agent, a simple prediction loss may well be enough to cause these structures to form spontaneously. Similarly a theory of mind is the simplest, "most compressed" way to predict the behavior of other agents in the same environment.
Although at this point it feels like these are just engineering problems as opposed to deep philosophical questions. The capabilities of ChatGPT are emergent phenomena created from the extremely simple training task of next word prediction. IMO this is very strong evidence that the rest of our cognitive abilities can be replicated this way as well, all it takes is the right environment and training context. It might start with something like this: https://www.deepmind.com/blog/building-interactive-agents-in... that uses cross-attention with an LLM to predict its next actions.
Some speculative ideas I've had:
- Brains (in animals) have largely evolved to predict the future state of the environment, to evade predators, find food and so on.
- To be effective, this predictive model must take its own (future) actions into account, a requirement for counterfactual thinking.
- This means that the brain needs a predictive model of its own actions (which does not necessarily align with how the brain actually works)
- Consciousness is the feedback loop between our senses (our current estimated state) and this predictive model of our own actions.
- All of this is to better predict the future state of the environment, to aid in our survival. For a hypothetical AI agent, a simple prediction loss may well be enough to cause these structures to form spontaneously. Similarly a theory of mind is the simplest, "most compressed" way to predict the behavior of other agents in the same environment.