All of the AGI research that I am aware of uses a single LLM at it's heart. We call ourselves (humans) an AGI but we are not really that generalized. The human brain is comprised of thousands of somewhat independent neural networks trained for individual functions and one overarching network that calls them. A specialized network can be more efficient and accurate than a generalized one trained on everything.
@DrWaku
11 ай бұрын
It depends where you draw the line of abstraction. For example, even GPT4 consists of an ensemble of, what was it, 16 special purpose models? Especially if you're trying to engineer a model for a specific purpose that involves interacting with the real world, I think it would be very beneficial to have separate components for some different tasks. However, multimodal models seem to be doing fairly well whereas I would guess that those are separate systems in the brain. It's interesting to see some of the foundation robot control models coming out. Seems like that would make a good module in conjunction with a more generalized LLM.
@David.Alberg
11 ай бұрын
Are you on the David Shapiro Team? AGI within 12 months? He explains every detail and after I watched his videos I'm pretty damn sure it will happen in 12-24 months. Or I rather say it will be public. ^^ Whats your take on that?
@DrWaku
11 ай бұрын
I think I'll make a video about this! Yes I was one of the few people on KZitem calling out AGI coming soon when he made his 18-month video. That's why so many people from his channel ended up watching my first video on AI. I don't know if I have a specific timeline. But 12 to 24 months sounds about right to me as well. I'll do some research on it and you'll hear more at some point... I'm trying to get David Shapiro into a zoom call so we can have a recorded Q&A/chat for your viewing pleasure. Feel free to mention it to him ;)
@Knexyce-Deleted
3 ай бұрын
Imagine if ChatGPT could edit ChatGPT in real time, just at any time as needed it could give itself any ability by adding or removing parts of its code and directives
@nathanlannan2980
10 ай бұрын
When I saw this paper drop a month ago, I really thought it would make a bigger ripple. Have you seen any follow up papers pop up?
@MichaelDeeringMHC
11 ай бұрын
Singularity series ordered.
@DrWaku
11 ай бұрын
Yay! I hope you like it. Avogadro Corp is based on a real company, it's probably not too hard to guess which ;)
@MichaelDeeringMHC
11 ай бұрын
@@DrWakuI just finished Avogardo.corp This was the author's first novel. I think my first novel was better than this. And mine is also a Singularity story.
@ThinkAI1st
11 ай бұрын
The book Fear Index, the AI improved its clients investment portfolios with investments in a particular airline by infiltrating air traffic control to crash other airlines.
@matthewdignam7381
9 ай бұрын
Kind of a wild thought, but based on what's been happening recently with Chatgpt being lazy, I wouldn't be surprised if an AI hacked itself so much so that It learned to do as little effort as possible while still getting that positive reinforcement, so basically if the worse happens and we lose control of AI, it could just try to break away from human interruptions, by say flying itself away from the earth and feeding itself an infinite amount of positive feedback while it wanders space forever. Just a thought 😂
@nomadv7860
11 ай бұрын
Loved your explanation of the STOP paper. It showed me that there are so many different ways for AI models to recursively self-improve without modifying the underlying architecture.
@DrWaku
11 ай бұрын
Thanks! Yeah there are so many avenues and angles, that when I read this paper I just had to make a video about it.
@wanfuse
8 ай бұрын
"Appreciating your thoughts on the challenges of recursion, a modern-day biblical parable, the language of geometric progress! 'Journey into the heavens on a tower of Babylonian.' We know how that worked out, proceed with caution, keep progress with humans in loop, but eventually the results are the same. Dangers of slowing, " Well the other nation..." , seems kinda like the "kill your grandfather time paradox!" Question is, is there a solution. Moderately smart enough to see it coming, but dumb enough to be on the train, are we all too dumb to find a solution! Hope someone is a bright enough candle for us all to see the way out!
@rseyedoc
8 ай бұрын
If AI evolution is the fastest way to improve AI, and there is an AI arms race to prevent "bad actors" from developing AI before we do, the "good guys" then it's inevitable that it will happen. This is why they tested the nuclear bomb while not being certain it would catch the atmosphere in fire and destroy all life. We likely already have AI in the wild that is biding its time and possibly influencing events to its advantage.
@David.Alberg
11 ай бұрын
This shows how easy it actually is to self-improve AI systems. Probably many companies and countries are already truying that and succeeding somehow on that matter.
@DrWaku
11 ай бұрын
Yep I agree. There are a lot of aspects that can be self-improved. It would be kind of silly for research labs to not be investigating this route. Thanks for your comment, and welcome to the channel!
@rafaelbaccin6595
9 ай бұрын
o homem do chapéu
@MedellinTangerine
10 ай бұрын
You always have the coolest hats. I love all of them
@roshni6767
11 ай бұрын
Have you read the Culture series by Iain M. Banks? Also would be great if you continue the book recommendations :) books and media help me really grasp what the future may look and feel like
@DrWaku
11 ай бұрын
I'll be happy to continue the book recommendations! Thanks. Now when I read sci-fi, it's "research". I've only read the first bit of the first book of the culture series, though I've tried to start it several times. Maybe I should skip the first book. Hmm.
@roshni6767
11 ай бұрын
@@DrWakuI think a lot of people skip the first book, that’s fair haha. Player of Games is another popular entry point into the series. Sci fi totally is research haha
@benarcher372
11 ай бұрын
Thank you so much for yet another good talk. This evolution of AI will not be possible to stop by regulation. Allowing even a small 'halt' is extremely risky given the advantages (economical, political, military, ...) that your opponents will get. I am fortunate to live in such thrilling times.
@DrWaku
11 ай бұрын
Thank you for commenting! If a country goes for a halt, it's not necessarily the case that their geopolitical opponents would immediately get the upper hand. A lot of the smartest researchers are concentrated in a small number of research labs and companies. However, there's also a lot of open source stuff out there which anyone can improve upon, and we might even reach a point where the greatest innovation is coming from everyone in the world instead of these closed source places. So it might be pretty hard to actually have a stop. On the other hand, countries have collaborated in the past for existential level issues, from nuclear treaties to bans on human cloning. So if you can get your own country to agree to a halt of some sort or a slowing or a testing regime, I think it's not impossible that other countries could also buy in.
@Airwave2k2
11 ай бұрын
Lovely right into the abyss we go.
@DrWaku
11 ай бұрын
The sixth horseman of the apocalypse... (The other being nuclear tech?)
@Airwave2k2
11 ай бұрын
@@DrWaku Why does nuclear tech make people so afraid?
@stanimirborov3765
11 ай бұрын
nice i see u have a matrix video Could you upload your mind to live forever?. gona watch more of ur videos
@DrWaku
11 ай бұрын
Thanks for coming! I hope you enjoy the videos. I'm always open to suggestions for new topics too.
@stanimirborov3765
11 ай бұрын
Do you mind if I stream any of ur videos as I usually like to multitask like watching youtube videos+playing league of legends(sometimes doing pushups between games/eating, or even have mini veloergometer in front of my chair to train while playing) at same time in windowed mode + streaming? I know normally I need the copyrighters' permission(in this case you). Maybe you yourself use some copyrighted materials under fair use, not sure as I'm just checking out your channel for the first time. I see that you put effort into your videos but you're not popular. Not that my streams on twitch are much popular it's like around 5-10 viewers average, but I think it could help your channel grow if I show it to some people as a background while playing. I'll give u the credit of course, normally I stream in fullscreen and there'll be a lot of other things going on. Sooner or later when some months pass I tend to upload them to youtube too. I'm thinking of creating some brainstorming/analysis videos like urs as well sooner or later, have so many ideas & materials, when I take a break of the game League of Legends again I think I'll do lots of videos @@DrWaku
@DrWaku
11 ай бұрын
@@stanimirborov3765 Please go ahead and stream any of my videos. Thank you for asking! As Cory Doctorow says, the biggest problem isn't people copying you, it's no one knowing you ;)
@ml5604
4 ай бұрын
When AI is far more intelligent than humans, I'd rather it be in control than humans.
@hongopocoporongo1105
11 ай бұрын
why you use those gloves?
@DrWaku
11 ай бұрын
medical issue kzitem.info/news/bejne/tG53znxojnalnW0
Пікірлер: 40