No - they've long gone past that limitation, they are now meant to be able to search the web for up to date information etc. but I suspect that whatever LLM he was using was using its internal data - probably using the "fast/quick" option.
No, they've applied AI techniques to a well-understood and clearly defined problem space. A space I have worked in a number of times and can see where it would work. Likewise occasional poster Dessi* has written on her FB page about using a chatbot to write testcases for her code and it did it very well. Again, a well-understood and clearly defined problem space. Both cases include the output being carefully verified, in thePrestige's case by the test infrastructure. In Dessi's case I'd have used standard test coverage tools like SonarQube etc.
Worst remake of the Dead Parrot sketch ever!
But strangely evocative...
Training an LLM model is a time consuming and resource-intensive process. That's what the huge new datacenters with their huge new power and water requirements are doing.
I mean, inasmuch as they "know" (or if you want to be pedantic, appear to know) anything, or "remember" anything, as long as the end result is that the information it outputs is accurate, why should we care about the internal process that led to the correct output? As a shorthand, we will end up using common words like know and remember to describe it.
The distinctions are interesting to me. And I think they're important, too. I think the better we understand what these things are doing, the better we can manage our expectations and our relationships with them.
The way I see it, LLMs can't not know something. They're statistical models. There's always going to be an "answer" that's statistically more likely than other possible "answers". That's where the hallucinations come from. You ask it for a legal citation, it's not going to say, I don't know of any cases that cover this. It's going to come back with the string of words most statistically likely to be a valid citation.
The way I see it, LLM's can't know something. They're statistical models. They just construct an "answer" that's statistically more likely than other possible "answers".
Good thing they're using it to identify and remedy vulnerabilities rather than to exploit them.
Its been in a limited preview, why is that statement funny? Do you believe Anthropic the company has been using it to exploit these security issues?
Yes. Don't you?
No I don't. It would be beyond stupid, it would make the stupidity of accidentally releasing your source code to the whole world seem like Einstein publishing his theory of General Relativity in comparison.
