I was listening to the New Year’s Day concert by the Vienna philharmonic and wondered who one of the composers was so used a popular song recognition app. (I expected it would make some fuzzy match on the piece and give me the name + composer). To my amazement it did give the name and composer but as played by the Vienna philharmonic in 2005 in the same location. The orchestra does not have the same members as 19 years ago, nor was it the same conductor, so it seemed the piece was matched on the acoustics of the Musikverein where they were playing, which I found astonishing.
It’s funny, because what makes them more like a real person is their inability to be consistent.
Like genuinely, LLMs are indeed cool, but in many ways all we’ve done is create a computer that is as bad at computer tasks as a human.
It lies with the same confidence that it uses to tell the truth, mostly because it can’t actually make distinctions between the two.
What we have done is even worse because by and large the marketing around AI has worked and people categorically trust what a single AI will tell them after asking the question only once and if that doesn’t terrify you, you aren’t paying attention.
If there is a future for ai it is in “ensembles” where you ask several independently trained and developed LLMs the same question, preferably at least checking some of the LLMs for consistency/indepth correct knowledge by asking them the same question twice in two different sessions. Then you evaluate from the ensemble of answers what you think the real answer is.
Sound tedius compared to just using a search engine that works?
…and to that question I respond emphatically - Yes, very much so!