At WebGPT: Improving the Factual Accuracy
of Language Models through Web Browsing, OpenAI describes their model:
We’ve fine-tuned GPT-3 to more accurately answer open-ended questions using a text-based web browser. Our prototype copies how humans research answers to questions online—it submits search queries, follows links, and scrolls up and down web pages.
To me that sounds like they decided to let GPT-3 interact with the open internet which is the common standard for what it means to be left out of the box. If my assessement correct? If so, what do we learn from this?
OpenAI is an AI capabilities firm & nonprofit. They spend almost all their time pushing the forefront of generalist AI reasoning. The reason they have not let an AGI proper out of the box already is because they do not have an AGI yet. Not sure what you expected of them, exactly, especially when it comes to their smaller models.
:goose10:
GPT isn’t an AGI but it’s certainly an AI. In this case, it seems they let it use a web browser and provide information for queries that will be sent to outside servers.
How is GPT-3 not general? It seems like the I should be more in doubt than the G.
My understanding is that “AGI” means AI that is able to do anything that a human can. GPT-3 can do well at some tasks (writing sentences) but presumably isn’t very good at driving a car or many other common human tasks.
GPT-3 can play Turing Test, and I think Turing was right to exclude things like “driving a car” from the definition of AGI. To quote:
Makes sense. I’ve always been using “G” more literally as “general”, but it’s certainly possible that this is unusual.
Given the success of transformer models at every other subfield of AI, I’m sure a GPT-like model would make a decent self driving car.
To be clear, I mean an AI which can pilot a car correctly on US roads 99% of seconds, which has already been achieved by many companies including Waymo, Tesla and Comma.ai. None of these companies have “self driving cars” in the sense of a legal, marketable L5 car, because the standard for legal, marketable and L5 is way beyond 99%.
Meant AGI when I said AI, sorry.
Yes:
What they mean by following is, specifically, only allowing HTTP GET/HEAD requests, rather than PUT/POST/DELETE etc. According to the standards, a GET is supposed to be free of (observable) side-effects and never have consequences worse than wasting bandwidth. In practice, on many websites, a GET will do all sorts of stuff, as I’ve learned the hard way, and is one reason that browsers no longer prefetch links aggressively (too many cases of users logging themselves out, deleting files, banning people etc).
Note that while WebGPT had live Internet access, DeepMind’s competing work is even more alarming from a tool-AIs-want-to-be-agent-AIs perspective.
I think we have a core lesson here. Anybody who talks about boxing should just acknowledge that the kind of organizations that would actually develop AGI don’t engage in such activities.
That would force them to admit that tool AIs want to be agent AIs and boxing was always irrelevant. (True AI boxing, comrade, has just never been tried!)