Ahead of OpenAI CEO Sam Altman’s four days in exile, several staff researchers sent the board of directors a letter warning of a powerful artificial intelligence discovery that they said could threaten humanity, two people familiar with the matter told Reuters.
Current AI isn’t really intelligent at all. It’s essentially just a search engine combined with that robot voice from TikTok videos. Of course it’s more complicated than that but it helps to illustrate the point, which is that the AI you’ve interacted with thus far don’t know if they’re right about what they tell you. They’re just hoping the answer they found was correct and stating it in an authoritative way that can confuse people who don’t know the real answer to the question it was trying to answer.
Actual AI will be able to reason out correct answers from incomplete information and solve complex mathematical equations very quickly. Being able to solve basic math problems without just searching it’s database for the correct answer is an important step towards real intelligence. It means we’re no longer dealing with a hard drive attached to an answering machine, we’re dealing with something that can process information in basically the same way we do, which opens up all sorts of awkward moral and philosophical questions.
From 10 years time…
The good news: a superintelligent AI has cracked faster than light travel that allows humans to travel across the galaxy in minutes.
The bad news: that AI uses its new-found ability to yeet all of us off to some barren rock far away and leaves us to die there with no resources because humanity is such a crazy, deductive pain in the arse.
And all of the humans that survived the landing will be like, "Holy shit it’s clean fucking air!
There’s fucking water without PFAs here!
The AI removed all the microplastics in my brain!
Holy fucking shit, somebody find all the billionaires and kill those fuckers!"
Not sure about this upcoming development, but they had the math part solved already via a Wolfram Alpha plugin which integrated into ChatGPT. As you may already know, Wolfram can already solve complex math problems with just a natural language input, so this isn’t anything revolutionary.
What would be revolutionary though is if it applied that same sort of logic beyond math, like towards language (and visual) outputs and be able to fact check, or at the very least, not contradict itself or hallucinate like it does sometimes.
It’s not a terribly complicated idea, really. You can train it to output formatted calculations when presented with a problem, then something in the middle watches for those and inserts the solution for it behind the scenes. You might even trigger another generation to let it appear more smooth when presented to the user.
I totally get the skepticism. It’s not surprising given how abused the term “AI” is, but they are a much bigger deal than “a search engine with that robot voice”. In fact that’s exactly the thing they are definitely NOT. They are terrible at recall and terrible at prioritizing reference information.
That said, Gpt models are the first, possibly most important piece of an AGI. They are a proof of concept that the ability to draw basic conceptual and linguistic understanding is possible from an enormous amount of data and shockingly little instruction. There’s no real reason to think they should be as good as they are at correctly interpreting written content, but here we are.
People make a big deal out of gpt because they think it will enable rapid improvement, and personally I don’t think that’s a forgone conclusion. It’s probably appropriate to compare it to the development of the first rudimentary computer: by itself it isn’t particularly groundbreaking, but drawn to its maximum it has revolutionary potential. Every additional step from here is likely as big or bigger than the one from gpt2 and to 3 and 4.
This is a great comment. I first learned ML at Google in Boulder in 2017 using TensorFlow. We were introduced by the google images team to re-create Uber’s fare estimation algorithm using 25+ years of New York City taxi data. GPS locations, fares, times of day, routes, etc. As expected, given gradient descent and how people chose to use the parameters, everyone had very different algorithms by the end. This is what has been known with ML for years (even with GPT, just massive models), but something that can process and learn on the fly is something else entirely and is pretty exciting for the future. Philosophical questions abound.