I’ve been daily driving a few different AI models for several months now, and my life will never be the same. Models like GPT-3 and GPT-j-6b have completely re-defined my definitions of creativity and knowledge.
They do not possess mind, but they can perceive, and, on the plane of our perception, contain their own notion of ‘truth’. This is how we control AI.
GPT models have no conceptual understanding of anything as we do, so describing them as having the capacity for ‘understanding’ and ‘truth’ is merely a useful personification of the ‘behavior’ of an absolutely alien and unworldly thing.
We must think about AIs in terms we understand.
In some ways, these AIs are exponentially more creative than any human could possibly be. But the things created by computational statistics are worthless unless humans deem them otherwise.
This is why we should fundamentally use AI to inspire and create, not make decisions.
This is an impossible question to answer. Usefulness and value vary entirely on context.
But months of playing with GPTs daily has taught me how to ‘make them think’ in the ways I want them to:
When the AI outputs some language, it is useful to think of this as the “AI’s truth.” It has no inherent connection with *our human *notion of truth, but if we imprint our notions of truth on these NLP generators, they usually do a good job of operating within those boundaries.
This is how we make NLP models useful: We must ‘program’ their fundamental ‘understanding’ of ‘truth.’
Despite being creative super-geniuses, these AIs have no inherent understanding of any aspect of reality. They need to be given a ‘foundation of truths’ upon which they can ‘build’ what we could perceive as ‘their own internal logic’ which ‘determines’ their outputs.
Axiomatic monads of truth no longer need to be boiled down into compiled code for computers to factor those axioms into their outputs. You simply tell the AI what the truth is, so it’s your job to figure out what useful philosophies to program it with.
No, you will never possess an ‘absolute’ representation of truth, and neither will your AI. Sociolinguistic truth is fluid and dynamic. But a simple linguistic seed of truth is often good enough for AIs to generate something useful.
You can now directly program computers with philosophy. Didn’t think I’d say that for a while.
In fact, if you don’t program your AIs with well-rounded philosophy, they will be uncontrollably wild and useless.
OpenAI’s mission is to make AI benefit all of humanity. That goal is made muddy by different definitions of usefulness, value, and ‘benefit.’ Posting AI-generated fake news certainly benefits the people profiting from it.