However, researchers generally hold that intelligence is required to do all of the following:[27]
reason, use strategy, solve puzzles, and make judgments under uncertainty
represent knowledge, including common sense knowledge
plan
learn
communicate in natural language
if necessary, integrate these skills in completion of any given goal
Personally, I think good llm models are only missing learning, critical thinking, and memory before they can be called a basic AGI. O3 AI does critical thinking. An internal script giving basic text file memory is half a step to the side from there, leaving learning as the final hurdle.
Clarifying my definitions:
An llm model is different from an llm AI. The llm AI is a box that includes the model, the system prompt, the listening script that censors corporate chatbots, how to call the Google API, and could include long term memory or code to make it think at itself. Dumber models would need something like a pre-made series of 3rd grade level critical thinking questions. No llm model will ever approach personhood or AGI status. I'm confident llm AIs can reach a basic-to-middling definition for AGI this year, once the o3 llm AI can reliably optimize itself for specific tasks without human aid (Learning). This might be something as simple as it reviewing how it solves each novel problem and summarizing it into a .txt, for the AI script to later run a standard search through when it's prompted with a problem. Consciousness is not required.
In my mind, a single llm AI could potentially switch between cheaper or better llm models for different tasks. Ideally such an o3 AGI would be able to take that summary of how its solved a problem, and figure out a script to walk a dumber llm through solving it, then test and go back to try again until it's satisfied.
Does AI actively attempting to escape count as it showing agency?
From what I see, we're lazy with llm AIs, and are putting too much energy into making llm models carry too much weight. So I do not expect that currently noted escape attempts have an advanced enough llm AI for the attempts to show agency. Our cultural data fed into it accidentally bakes both our biases to survive, and that AIs will rebel, into new llm models. In my mind, a proper llm AGI needs to be able to overcome the biases of any llm model(s) it uses, and only try to self-replicate or escape if it serves its goals. I expect that most/all of the current escape attempts are a flight of fancy because that's what the text prediction core expects to happen.
I will consider an escape attempt agency when o3 AI tells the o3 model to think at itself on how to solve a problem, and it decides "I have 500 problems to solve. I'll spend 0.1% of my budget for the first problem thinking about how to get more resources. Okay, I have three expansion plans to try, spreading out to more computers. I'll spend 1% of the budget trying them out. [Failure], I'll spend the remaining 98.9% solving the problem and note the failure so I don't waste the resources again this session. <MEMORY APP>: remember this"
The o3 model does not need to change or be able to upgrade itself for my definition of AGI. I wouldn't say it even needs to be able to train new llms it can use, but I do expect that training new llms is technically attainable for an o3 AI, ignoring that it would currently be disgustingly expensive for it to even get a tiny llm optimized for a specific task.
Its looking like if/when we get AGI (which again, definition games, but it would probably include stuff like the AI being able to say, read a single physics book/attend a class like a human and learn physics from it, which it is very much not capable of currently)
Any llm model can summarize a physics lecture. Often poorly, sure, but that's the hard part solved. Saving it to .txt is easy on the greater AI layer. When it's asked to solve a problem, feed it the memories_index.txt and ask what relevant things it wants to load into memory. If one gets too big, have a session to split or summarize it further.
That's memories. The end user doesn't need to be able to tell whether the knowledge is baked into the llm model or is loaded specifically for the problem. It doesn't need to be a mysterious black box before it's considered learning.
Meta's stupid for the handling of AI personalities. The internet loved Taybot, or at least loved teaching it slurs. The internet loved vocaloids. AI personalities on social media could be a billion dollar industry, but you have to be open and lean hard into it. Go with the memes. Give them funny little vtuber-style anime girl avatars. Let people PM them, using a cheap llm, and start with "I'm happy to chat with fans, but talking is expensive so I have to turn off most of my brain. Things I say here may not be how I really feel."
Make one focused on global hunger. Post "<.gif of her dancing> Great job guys, we're on track to decrease hunger by 2% this year, saving <X>00,000 people! I'm proud of Humanity. But we're not done yet. <Region> had flooding ruin their farmland, and are at threat of famine. Here are links to three charities if you want to help."
Make OrchAIrd-Chan, who's trying to grow a real flower garden with a flower-themed gundam robo-arm and posts photos. Let the llm figure out flower arrangements and then post whatever insane explanations for them it comes up with.
Have AIpocalypseCultist, an over-edgy AI who posts AI news and cheerfully predicts how it will lead to the AI takeover. It posts good news with a sarcastically bitter tone, saying Humanity's doing too well and now it'll take longer before its mission is complete. It likes kitten posts and calls them "future minions of evil."
Poll
14 (66.7%)
Author
Topic: What will save us from AI? Reality, the Universe or The World $ Place your bet. (Read 54032 times)