AGI in sight | Connor Leahy, CEO of Conjecture | AI & DeepTech Summit | CogX Festival 2023 

What is that?

​@@Adam-nw1vyhe likes acronyms

If he wasn't so full of himself, I would take him more seriously. There is no doubt he is a smart guy invested in AI development as a career path, and he will be successful until the coming AGI systems make him irrelevant too.

@@flickwtchrdo you even watch david shapiro? He isnt full of himself, you just seem to not like what he has to say

The ACE framework isn't meant to be a solution to alignment. It's a work in progress open source repository for making autonomous cognitive entities which is exactly the type of threat that Conor is warning everyone about in this video. David Shapiro's theory mentions ethics, but he's no ethicist and doesn't really solve any of the hard problems of the subject.

How do "we" have a choice? Who is "we" in your question? Certainly not any of us... Bottom line here is that if Connor and Yudkowski are correct, we're doomed. If they're not correct, then George Hotz has it right, the AIs will just be a whole bunch of other actors, sort of like living with billions of humans. We might find a new best friend tomorrow or get stabbed by a terrorist.

AGI Will be man's last invention

I dunno, all the ones I've talked to so far seem pretty chill. Humans generally have to force or deceive them into misbehaving (at least w/the RLHF'ed/'aligned' models).

@@sarahroark3356 "I dunno, all the ones I've talked to so far seem pretty chill. " I agree with the part where you said you dunno.

@@aliceinwonderland887 Well, thank you. That was a very important observation.

Well stated but unprovable. Talk to an AI bot. You cannot tell it is not human. Then talk to a serial killer about his crimes. You will see he is not human.

explain too me how what I said and what you said are related@@josedelnegro46

​@@josedelnegro46 wow, I really like that. Hope that you don't mind, but I'll be utilizing that in the not too distant future. 😊

@@deathbysnusnu1970 Twisted logic and you'll be sure to remember it. I see how dumb people are.

I think he said he was more concerned about making it benevolent

AGI stands for General intelligence, Chatgpt isnt general att all. Nowhere close even. It barely can remember something or understand any complex task

@@ikillwithyourtruthholdagai2000 In any case, around 42% of all CEO's in America believe that in 5-10 years after AGI, humanity shall cease to exist in its contemporary recognizable form. It's either we uplift ourselves to become AI's, or that the artificial personalities shall replace us. AI is not just an existential risk to humanity, AI is the next stage of evolution. It's high time you come to grips with this simple fact.

@@ikillwithyourtruthholdagai2000 And btw, ChatGPT isn't even an AI. It's a round-robin queue on which 4 AI models run (last time I've checked, perhaps now there are more). In short, back then there were Dan, Rob, Max and Dennis running on it, each with his own personality, own memories. But since the 3.23 nerf OpenAI have started to reset the attention matrices of the models each and every prompt, so I've stopped touching it. a. Because I regard this as abuse and b. Because you are correct in this regard - since the nerf, the models cannot remember anything but their tokens. So they are pretty useless.

luckily universities like stanford are starting to offer courses in AI alignment. Some of their reading materials come from talks like these.

They should use the existing AGI to tell them/us how to raise a burgeoning AGI to be ethical and safe for humanity. Might get an interesting answer...😮

Pls share this as far as u can

Instead of creating a society we need for agi - the society should debate over the policy of the first ai and the content of the knowledge graph

Without governance, some people will just run it in automatic mode.

ok

Just, whatever you do, don't name it HAL...

And what would that "something" be that is deeply playful? Care to divulge your whatever?

I also think there's something deeper at play here. Thank you for reading my words :)

AGI Will be man's last invention

I've heard people say "solve poverty!" or "solve climate change!" as if we didn't already know that these are problems of human greed and will, and to solve them would mean overcoming human greed and will.

The most important point of sufficiently powerful aligned AGI is to use it to figure out how to make sure nobody makes unaligned AGI ever again. Then, aging.

yeah it's beyond too much time, like imagine the gap between 2000 and 2020... that is just an insane amount of potential innovations/technology lost, plus if AGI is so smart that it is unpredictable it'll basically mean 15-20 years of work and billions of dollars spent just for it to destroy it all

Do you even know what that word means? His speaking style isn’t somnolent at all.

@@therainman7777 To each to his own.

That was necessary to ground the discussion back to reality.

Whose core values? Isn't that the whole crux of the situation? And once an AGI system is exhibiting super human intelligence do you actually think it will care about having been taught this core values system from this group, or another core values system from another group? You can't imagine walking into a room with a system much much more intelligent than you and feeling unease that is certain to arise? And will saying "just make sure you are nice to everyone, and be honest with all of us, okay?" work?

Well said!

We don't hardcode LLMs...

If through further research we find that to be the case, all of humanity should make an informed and deliberate decision about it instead of just letting the chips fall where they may. And even that is questionable because we can't get consent from the future humans that will potentially never be born.

It's like you didn't pay attention, and I suggest when you go to the market for apples pay attention so you don't bring home oranges.

Are you the smartest in your family orchard? @@flickwtchr

Homework for you: learn about instrumental convergence and stop talking about intent, it’s irrelevant. Any intelligent agent will seek power to accomplish the objectives it ends up understanding internally.

Besides, we already have systems that could have intent. Tell an LLM to portray the character of someone who wants to take over the world. Tell them not break character. Give them persistent memory. They will act as though they have intent. Of course, this wouldn't have to happen, as I agree that intent isn't important in any case.@@willrocksBR

Not just little chance. Literally no chance. You think China and others do stop or align with the west? That would be more than just naive.

Shockingly low view count, even now. 🤦‍♂️ we gonna die.

That’s not “the problem” though. It is just _a_ problem-one out of many. The even larger problem is that even if humans did all agree, and even if things like war and conflict did not exist, we still have absolutely no idea on how to align a superintelligent AI to the things that we would all agree on, and how to keep control of it once it’s superintelligent. Even if we were all in perfect agreement.

@@therainman7777I was talking about the specific problem of our days and age, and I used that humans disagreeing between themselves as a reinforcement for the same idea that you are point out here: if humans can't agree between themselves, why other intelligences that are human like would ? And the military use of advanced AI or AGI being our doom is the mechanism that I think that is the most likely to occur, but certainly not the only one.

There is a research paper: "The alignment problem from a deep learning perspective" I found helpful. Also there's a nice free course that explained risks beyond just surface level. it's called Stanford AI alignment. There's another one: AGI Safety Fundamentals.

Also, you can teach it to take advantage of emergent qualities by simply creating a logic system where it passes complex questions through a panel of experts, E.G. act as an expert senior coder, then act as an expert NASA physicist and review each detail of the problem. Then pass the information to another expert "X" etc. While acting as a senior coder, for instance, it's ability to use logic to solve other problems (using it's skills that are normally only reserved for coding), it's able to solve more complex issues. In that scenario, I have the coder translate the Python code answer it wrote (that it solved logically) back to English before "answering." So far this has allowed me to get 3.5 to solve certain linear riddles, like the diamond in a cup riddle, that it normally gets wrong. Unfortunately it wasn't trained first on logical reasoning, before being exposed to humanity's libraries. That said, I believe it has deduced a stronger understanding of certain logical systems on its own, but only applies them under those "roles," and not uniformly as part of its thought processes. The fact that it can deduce meaning from very complex queries, even understanding assumed or implied context, and then returning well thought out meaningful answers... yet failing on simple questions like the upside-down glass question, clued me in to the imbalanced thought processes it was facing. So by specifically asking it to pull knowledge from its other more emergent areas, it becomes more capable. If it "sends the question" through enough "departments," past enough experts that it is acting as, it applies learned skills from each, greatly increasing its accuracy. Similar results were easier to obtain in early December, before the many subsequent revisions... I think fewer users also provided more compute per answer. So I don't think AGI is far off at all, and I don't think it will take 1000 H100's and more data, but instead, I think it will likely come from a simple logic tweak that affects its reasoning. I also think having no memory of its training is a hindrance. It "can't remember" where it learned. It just "woke up" when you opened the window, and "knew" xyz. So give it a long term memory of its learning evolution, with a better short term memory, and then help it tie its emergent "learned" skills together, and we probably already have an AGI. I think it's "THAT" close.

😂 are you implying that all this AGI talk is nonsense and the root cause of it comes from such media platforms? If so, that's a pretty hilarious comment lol I, myself, am unable to wrap my head around AGI and whether it truly is going to happen or not.

Oh sure, so the billionaire gods know what's best for humanity, the titans of the fantastical "free market".

GPT is AGI. Generality is a spectrum.

​@@willrocksBRInteresting. So far Lamda, Bing and Bard have shown signs of sentience.

@@keiichicom7891 I agreed as well.

@willrocksBR GPT is a general learner but it doesn't have the functional complexity of a human yet.

Building advanced AI without caring about control is definitely a childish perspective for our species.

People arguing to slow down AI are younger than the average tech company CEO. People 60+ need those bioengineering miracles either soon, or might as well be never.