I agree, we need trustworthy AI not AI's that pass as dumb humans.I find it irritating that the answers to "What is 2 + 2" are all wrong although some of the answers hedge their bets. I expect a computer to give a Spock answer and not attempt a human interpretation of the question.
To get sensible results, the question has to be correct and the answer should answer the question. When you have a computer giving the best guess to a question it has mangled, what use is it other than a novelty?
It would be more to the point if the LLM fired back questions until it thinks it can give an accurate answer, it would also educate users to be more accurate in their questioning.
The history of AI and Government has a example in the Australian Tax Office.
It has gone from itself being a dumb human to catching dumb humans.
Of course it seems to be targeted at everyday people and not the rich like politicians who can hide their money in "legitimate" schemes.
I think AI should be more like Star Trek Data not Snake Oil salesman, Lawyers or others that speak with forked tongue.
A LLM that is an apologist then in the next paragraph say you are too stupid to understand anyway is not impressive.
Well it is impressive in one way, very human, but I don't trust humans most of the time.
And I sure don't want an AI acting like a human.
AI's that lie, make things up or are DEI biased are not a good way to convince me to trust them.
So I test them, which I can do online and locally now.
I just wish I could test our leaders the same way.
Some good news, I am not the only one who thinks this way.
The MoE, Mixture of Experts AI model is an improvement, Agentic methods are all helping AI's get better faster and perhaps useful.
Still waiting for the Cures we were promised at the start of this AI binge of spending tax payers and Gov money.
Of course looking for cures is the wrong question, "What is the cause" might find different answers.
Can an uncensored AI answer that or will it cough up conspiracies?
StarLink has faster comms than underwater comms, it it being used to trade stocks faster than normal?
Shh, don't tell, buy StarLink stocks instead......
Anyway, back to looking for simple MoE/Agents/Music AI to run on Pi5.
These new Coding AIs sure look like fun, building one into an IDE would be very handy.
My language of choice these days is Free Pascal, can I train my own code AI for that language?
Most AI's seem to be Python based, wouldn't they be faster if they were compiled?
Warzone2100 is one game I have been compiling on Pi's for years.
It can be modded and can use different "AI"s.
It would also be its own training sim world.
Likewise Blender can take plug-ins and there is one for Stable Diffusion already.
I could make AI generated 3D models and texture in Blender for Warzone2100.
And the whole point of Pi's is to help learn coding.
Statistics: Posted by Gavinmc42 — Sat Apr 06, 2024 1:29 am