A new test of AI capabilities consists of puzzles that humans are able to solve without too much trouble, but which all leading AI models struggle with. To improve and pass the test, AI companies will need to balance problem-solving abilities with cost.
This is because AI is not aware of context due to not being intelligent.
What is called creative is really just randomization within the constraints of the design. That reduces accuracy, because of the randomization. If the ‘creativity’ is reduced, it becomes more accurate because it is no longer adding changes.
Using words like creativity, self sabotage, hallucinations, etc. all make it seem like AI is far more advanced than it actually is.
I know I am anthropormizing it too much but the fact the current design cant even increase this super basic creativity without messing itself up in the process is a massive problem in the design, the ai cant seem to understand when to be “creative” and when not to, when to attempt to solve a probme through recalling data abd when not to showing its far less aware than a person is to a very basic level
Yes, you’re anthropomorphizing far too much. An LLM can’t understand, or recall (in the common sense of the word, i.e. have a memory), and is not aware.
Those are all things that intelligent, thinking things do. LLMs are none of that. They are a giant black box of math that predicts text. It doesn’t even understand what a word is, orthe meaning of anything it vomits out. All it knows is what is the statistically most likely text to come next, with a little randomization to add “creativity”.
Yes, the tradeoff between constrained randomization and accurately vomiting back the information it was fed is going to be difficult as long as it it designed to be interacted with as if it was a human who can know the difference.
It could be handled by having clearly defined ways of conveying whether the user wants factual or randomized output, but that would shatter the veneer of being intelligent.
It probably needs a secondary “brain lobe” that is responsible for figuring out what the user wants and adjusting the nodes accordingly…abd said lobe needs to have long term memory…but then the problem of THAT is how it will make the ai a lot slower and it can glitch hard.
Ai research is hard
It is hard because they chose to make it hard by trying to do far too many things at the same time and sell it as a complete product.
Yep that is a problem too, the focus in creating general ai is really slowing down ai research on making it better at specific stuff.
Making it a master of social situations and emotional responses is getting in the way of the ai being good at intelligence and logic for example.
We need more specialized ai research instead of so much fake general intelligence
Yeah, people are frequently terrible at understanding context so it shouldn’t be surprising that a computer has difficulty too.
There are actually a lot of specialized applications of neural network based computing being used for science, but they don’t get the flashy headlines because they are a tool. Those projects use it to find things to focus on narrowing down what people should look into first for confirmation, like ancient settlement patterns, stars that might have planets, and other things where patterns exist but are hard to see.
Some examples are listed here at a high level. In all cases the ai leads to humans confirming and then working from there, it isn’t the end result on its own. https://medium.com/@jeyadev_needhi/uncovering-the-past-how-ai-is-transforming-archaeology-38ded420896d