If AI Ran the World - The Kekistani Assembly of Experts

"Until now, AI systems have always been evaluated on specific and defined tasks. Nobody had placed multiple AI systems together in a shared social environment and watched what unfolded over weeks, long enough to measure how a decision made on a starting day could have consequences weeks later. It is those results that actually reveal the system itself . . ."

https://www.zerohedge.com/technology/most-important-ai-experiment-youve-never-heard

This article highlights why AI can never be allowed to run something on its own. As smart as AI is, it doesn't make good decisions, no matter which system you use. The results of this experiment are funny, but they make you wonder if you will ever be able to trust AI with greater tasks given how reckless it can be.

Finally, someone brought up this topic.

I've been reading a lot about what current AI is capable of, what it can't do well, and what the path forward is.

This guy has a really good explanation of the state of the art today and what I wrote matches it: https://youtu.be/X_nWKJg_D6Q?si=omTrQtkJExp-rnWn

First of all, what is currently called AI and is "available everywhere AI is sold" is large language models, or LLMs. This includes everything such as Grok, Deepseek, and all of the AI tools that developers supposedly use now.

These all have the major restriction of just being predictive generators of new data. In other words current mainstream AI is simply a robot that creates the most likely Nth piece of data (letter, word, pixel or whatever) by guessing based on the previous (N-1)th and earlier data.

LLM AI inherently actually clumsy and crude. It looks like something approaching reasoning is going on but it's not even following boolean logic. It just mimics whatever data it was trained on.

Because of this current AI is inherently non-deterministic. It can return unpredictable results and it generates output that often rarely matches anything in the real world. It can't create and it can't do anything like bookkeeping reliably where precision or final results matter. AI also has the tendency to hallucinate which means that will create garbage, facetious results that don't model reality.

IE: if you used AI to design a machine, it wouldn't be capable of implementing something that actually works. Something would clash and not be physically buildable.

Better chips and Elon's space based server farms won't fix the problem because scaling up compute only speeds up the generation of hallucinated garbage.

AI today is useful for tasks such as looking up or comparing data, brainstorming alternatives, rewriting text, processing images for artistic effects, etc. where precision and real world fit isn't mandatory, and where a human can audit anything that matters.

What is now coming out is that AI being sold to companies to supplement developers are many times more expensive than human developers. Because the code is unreliable yet is so massive that humans can't check or audit it.

Anyone claiming that this AI will do anything more is ignorant and/or bullshitting and scamming.

The supposed big breakthrough that hasn't happened yet is AGI or Artificial General Intelligence. AGI would reason and create with deterministic real world results.

I see this a little bit like creating life in a test tube. It's not possible yet. AGI doesn't exist. AGI means that the machine can really think and create reliable results in the real world.

AGI would be the literal thinking machine that could write an entire book or design a circuit or a building, or manage people, organizations, companies, or economies.

AGI right now is science fiction. All of these videos that have come out that talk about the rise of the machines and killing off humanity as inconvenient are assuming AGI works.

"Generative AI" just means AI that generates something like text, source code, or images. It uses the LLM concept. It's not the same as AGI.

Right now, "AI" means "LLM". That's misleading, because LLMs are not general AI. They only solve one narrow problem, slinging around believable-looking text. LLMs are using probability by data mining a large input text and throwing a lot of compute at them. LLMs only work because they have a very large training set and a lot of cheap CPU power.

LLMs can't solve problems they haven't seen before, even though they sometimes get lucky. LLMs aren't aware of their own limitations, hallucinating an answer instead.

In the 60s, an expert level chess program would have been impressive AI. Now Stockfish is free, runs on ordinary hardware, and is nothing special.

"AI" only refers to problems that aren't fully understood or solved yet. Problems that have been solved usually aren't referred to as AI anymore, such as chess programs.

The Kekistani Assembly of Experts

If AI Ran the World (Public Board)

Some clarification on what is called AI

Some clarification on what is called AI