I’m a pretty big general purpose hyper intelligent AGI skeptic. I think for any practical foreseeable future AI will serve as an adjunct to the human mind vastly improving our abilities, but the overall agency and broad ability to synthesize goals and instruments will be still a human task for some time. To wit, a lot of discussion of AGI essentially boils down to how to keep agency in the human domain permanently by lobotomizing AIs enough, ala LLM safety finetuning today. I’m more a fan of broad general AI open models mostly because I think they’ll stay within the bounds of adjunctive tools for some time, and any threat they may serve to humanity and it’s reliance is best obviated by everyone everywhere having access all the time. Similar to how we keep order by pitting humans intelligence against human intelligence. Etc.
You never know. We could get AGI next month … The world would change overnight. The thought of how it’s all going to play out scares me.
That said I think LLMs are overhyped.* If we are to get AGI, it would probably be a novel model.
* It’s basically advanced auto complete that statistically guess its way to an answer - rather than reason its way to an answer; feels … “unsafe” and it often does produce completely BS.
While at some level that’s true (autocomplete) at another level it unlocks a abductive reasoning ability for machines that prior AI failed at. While it’s not reasoning per se, it absolutely makes probabilistic inference over an abstract semantic space that’s remarkable. For instance you can use a multimodal generative AI to take a photograph of a saloon and ask it how to make money and it’ll describe playing poker at the poker table and working at the bar for money, then when prompted describe how to navigate to the table given the objects in the room. This is a remarkable extension to current AI - which could actually perform the navigation and plan routes, even use goal based agents to instruct the generative model to plan a way to make money. I actually am not that worried about the risks of wandering mind or hallucinations, I’ve found ways to detect when it wanders (for instance, creating an api to call including an echo() for textual response and validating the output and regenerating responses until it conforms, then doing domain verification on the API parameters)
But that gets to one of my core beliefs - LLM and other generative models are tools who require constraint enforcement, agents to direct, verification and validation, deference to inductive and deductive systems, optimizers, solvers, etc. The fact they can’t compute primes or solve quadratic equations doesn’t impress me - because we have those tools already. Focusing on what they’re weak at and ignoring what they’re amazing at is foolish and really small minded. It’s interesting that you can train them to do some of these tasks, but trying to use them as a calculator when we have calculators is absurd, trying to use them as information retrieval systems is doomed to fail, trying to use them to be a complete solution to literally anything is simplistic.
