I disagree. I don’t think LLMs are a pathway to AGI. I think LLMs will lead to incredibly powerful game-changing tools and will drive changes that affect the course of humanity, but this technology won’t lead to AGI directly.

I think AGI is going to arrive via a different technology, many years in the future still.

LLMs will get to the point where they appear to be AGI, but only in the same way the latest 3D rendering technology can create images that appear to be real.

I'm not saying LLMs are. LLMs are not the only thing going on right now. But they do enable a powerful tool.

I think the path to AGI is: embodiment. Give it a body, let it explore a world, fight to survive, learn action and consequence. Then AGI you will have.

Note that embodiment doesn't mean in anyway human or animal like.

For example, you're limited to one body, but an A,G|S,I could have thousands of different bodies feeding back data to a processing facility, learning from billions of different sensors.

Well, I disagree with you on embodiment, but on the thousands? Right that’s another part: evolution. Spread your bets.

But I disagree about a human or animal body not being required.

I think we have to take the world as we see it and appreciate our own limitations in that what we think of intelligence fundamentally arises out of our evolution in this world; our embodiment and response to this world.

so I think we do need to give it a body and let it explore this world.

I don’t think the virtual bodies thing is gonna work. I don’t think letting it explore the Internet is gonna work. you have to give it a body multiple senses let it survive. That’s how you get AGI, not not virtual embodiment. Which I never meant, but thought it was obvious given the term embody minute self strongly, suggesting something that’s not virtual! Hahaha ! :)

Also continuous learning. The training step is currently separate from the inference step, so new generations have to get trained instead of learning continuously. Of course continuous larningin a chatbot runs into the Microsoft Tay problem where people train it to respond offensively.

Yeah, evolution multiple generations. Necessary for sure. Things have to die. Otherwise, there’s no risk. without risk, There’s no real motivation to live and without that there’s no emotion no motivation to learn and without that there’s no AGI.

I'd say you need it to learn within each generation as well. Human & animal learning doesn't require evolution, we learn without replacing the whole. Connections within brains change while the organism is still alive, without even individual neurons dying off. The current process is purely iterative; GPT 1 was "done" and GPT2 got made. Then GPT 3, etc. If ChatGPT4 gets corrected about some particular thing by many users it won't learn that for the next user to come along.

Of course learning has an inherent trust issue; if the inputs one is learning from are incorrect then incorrect information will be learned, and incorrect output will result. Microsoft experienced that with Tay, which quickly learned how to respond to people via training from 4chan users. That was predictably disastrous, and what OpenAI are hoping to avoid with the generational learning strategy.

My embodiment is my PC environment. I interact with the world through computer displays.

There is no reason embodiment for AGI should need to be physical or mammalian-like in any way.

Strong disagree. But it may take me a while to elucidate and enumerate the reasons.

I'm disabled and have had a computer in front of me since I was 2. I'm rarely not in front of a screen except to shower and sleep.

Plenty of very intelligent people are completely paralyzed. Sensations of physical embodiment is highly overrated and is surely not necessary for intelligence.

Nice try SkyNet

Hahaha! :) thank you. That is such a compliment hahaha! :)

> LLMs will get to the point where they appear to be AGI, but only in the same way the latest 3D rendering technology can create images that appear to be real.

Distinction without difference

Disagree. Huge difference. In our tech power society, we often mistakenly think that we can describe everything about the world and equally falsely only what we can consciously describes exists.

But there is so much more than what we can consciously describe, to reality, like 10,000 to 1 — and none of that is captured by any of these synthetic representations.

so far. and yet all of that is or a lot of that is understood, responded to and dealt with by the intelligence that resides within our bodies and in our subconscious.

And our own intelligence arises out of that, you cannot have general intelligence without reality. No matter how much data you train it on from the Internet. It’s never gonna be as rich as for the same as putting it in a body in the real world, and letting them grow learn experience and evolve. And so any air quotes intelligence you get out of this virtual synthetic training is never going to be real. Itis always gonna be a poor copy of intelligence and is not gonna be an AGI.

Developing an AGI is not the same as developing an artificial human. The former is achievable, the latter is not. The problem is many of the gnostics today believe that giving the appearance of AGI (ie having all the utility that a general mechanical intelligence would have to a human being) somehow instills humanity into the system. It does not

Intelligence is not the defining characteristic of humanity, which is what you're getting at here. But it is something that can be automated.

nothing I’ve seen from OpenAI is any indication that they’re close to AGI. gpt models are basically a special matrix transformation on top of a traditional neural network running on extremely powerful hardware trained on a massive dataset. this is possibly more like “thinking” than a lot of people give it credit for, but it’s not an AGI, and it’s not an AGI precursor either. it’s just the best applied neural networks that we currently have

As a layperson, what does the special matrix transformation do? Is that the embeddings thing, or something else entirely? Something about transformer architecture I guess?

I'm not saying OpenAI is close. Collectively we are tho. The train is rolling, unstoppable momentum. We just have to wait.

Science fiction theory: Ilya has built an AGI and asked it what the best strategic move would be to ensure the mission of OpenAI, and it told him to fire Sam.

Haha, yeah! See my closely related but more 90s-action-moviey: https://news.ycombinator.com/item?id=38317887