> the main argument against automating these tasks was that machines can’t think creatively. Now this argument gets weaker by the day.
Citation needed.
Citation needed.
A: needs to have a close to complete understanding of the problem to solve and the tools they have in languages to complete this in the most efficient way and..
B: need to be able to iterate through all of the different options it can generate and run `perf` or what have you on the outcomes of each and present the top few winners
AGI might not be a thing we can make, but if we can make enough semi-intelligent AIs communicating via a standardized format that are each 95% right on their respective domain we might be at a level that is good enough. Add in for critical domains some sort of polling that is done between a few admin AIs to make sure the winning results are not a hallucination.
AI may very well plateau, but it might not.
Is it? I see no evidence that machines are any closer to "thinking creatively" than they ever have been. We certainly have been developing our capacity for computation to a great extent, but it's not at all evident that statistical methods and creativity are the same thing.
That being said, I think we could make such a system. It just has to have training data that is competent...
Some systems built on top of LLMs have this built in - Perplexity searches for example usually ask a follow up before running the search. I find it a bit annoying because it feels like about half the time the follow up it asks me isn't necessary to answer my original question.
That's rather missing the point. If your question makes no sense it will not ask a followup, it will spit out garbage. This is pretty bad. If you are competent enough to ask it to ask for followup then you are probably already competent enough to either not need the tool, or competent enough to ask a good question.
I don’t use it for dev, but other things and I get chatgpt asking me follow up questions multiple times a day.
The thing is that statistical models only need to be fed large amounts of data for them to exhibit what humans would refer to as "creativity", or "thinking" for that matter. The process a human uses to express their creativity is also based on training and the input of other humans, with the difference that it's spread out over years instead of hours, and is a much more organic process.
AI can easily fake creativity by making its output indistinguishable from its training data, and as long as it's accurate and useful, it would hold immense value to humanity. This has been improving drastically in the last few years, so arguing that they're not _really_ thinking creatively doesn't hold much water.
It is. And I bet most people would agree with GP. Most people (including engineers building these systems) have experienced surprise with some of the outputs of these models. Is there anything better to gauge creativity by than perceived surprise?
I think there has to be, since such surprise can be generated through purely random mechanisms, and I don't think anyone would call a purely random mechanism "creative".
Sometimes it adds an extra detail or two even!
A few years ago that was almost unthinkable. The best we had was arithmetic over abstract language concepts (KING - MAN = QUEEN)
We don’t have a solid definition of “creativity” so the goalpost can move around a lot, but the idea that a machine can not create new prose, for example, is just not true anymore.
That’s not the same as creativity, sure, but it definitely weakens the “computers can’t be creative” argument.
Sometimes. If the embeddings were trained well.