sorry, had so re-upload this one! had some issues with the music in the intro :(

So Ilya is going to hunker down & play games intensely, without interruptions from investors, customers, mothers…

So a super intelligence trained on survival games that emphasize resource hoarding and the elimination of obstacles to growth such as anything that might interfere with that hoarding. What could go wrong.

I remember a Ghost in the Shell episode where they try to protect a super rich finance guy who made AI trading software for himself from a communist Chinese assassin. All his mansion is automated with robot maids who also protect the mansion. At the showdown with the assassin, they discover the man had long died from natural causes but all the stuff continued running as if he was still alive.

Suggestion: a few second pause or a hello everyone, at the start of the video. The way it’s edited it feels like I have missed part of it.

It’s crazy that the movie WarGames is essentially a self aware computer learning nuclear war is un winnable by playing tic tac toe against itself and that plot is now the bleeding edge of AI.

I don't know why but I feel that Dwarf Fortress will be the Dark Souls for AI.

Ilya Sutskever has been consistently 10 years ahead of the field. He can see 10 years out. He was playing games with AI when OpenAI was founded. OpenAI was built around Ilya. Ilya is the one to watch.

One might guess that Q* is optimal search over a space of value functions.

When the owners of BigAI talk "AI-safety", they actually mean keeping _them_ safe from the rest of us. As in; _AI must never help the riffraff escape control_

So this is why the simulation was created! We are all just training data for our ASI robot that we will be integrated into once training is complete. We're going to be AI gods!

6 months from now, Ilya's 2nd-hand man quits his company due to a "conflict of interest", announces his own company a couple months later, stating a primary goal of reaching "omegaintelligence". And the cycle continues.

In this case, I wonder if IQ mostly measures the depth of search, knowledge is the breadth of search, whereas human's character is defined by one's unique heuristic function.

3:10 -ish... ALL I KNOW IS: I'm soooooooo glad they included games like "Tom Clancy"-esque first-person shoe-tears! I mean... I couldn't think of anything better to have the android-dogs fed to go along with the head mounted m-134! Ya know, for "customer service" oriented training purposes!

It's important not to forget about all the stuff everyone knows that isn't public knowledge.

All these researchers just watched War Games from 1983 and said, "hey, that's a good idea to get to ASI!"

Great video. Good job connecting the dots. It sounds like the kind of work needed to get an increase in the ARC-AGI benchmark.

Let’s go, ASI soon baby!

i can smell agent smith...

Hey Wes, really glad and appreciative of the great work that you are doing. I know this question is somewhat unrelated, but i was wondering what kind of note taking tool you are using to keep track of all of those topics, while not getting lost in the sauce of the tool itself? All the best and please keep up the fantastic value that you contribute to your fanbase.

I'm often wondering when playing games on my phone if I'm playing an Ai or a real person.

Hey Mr. Roth :), thanks for your videos and everything. If you don't mind me asking, do you publicly post these cool thumbnail etc., photos etc., somewhere? Some are really cool imho, for ex. like the one you/software flashed on screen i guess at, ~1 minute mark in this video for ex. perhaps please?

18:06 What is the name of the software you're using?

Hmm. What would that mean for Amazon and its twitch Service + AWS?. Very silent on this front at the moment,besides alexa subscriptions. Waiting to come out with the killer App?

Every week that goes by makes me think we're all just AIs in training. I think I might need another go-around

No one (except Sam Altman) wants to be the next Oppenheimer.

I keep thinking about that 1983 movie I watched when I was a teenager, called War Games. They taught the WOPR computer using games. And of course, at the end of the movie, the world is saved through the WOPR using a game to learn. And now it’s 2024 and we’re training AI using virtual worlds and games.

much better thanks

I havent thank you for your work, thank you wes. I do my part and tell people in the know about you.

Nobody is going to believe me, but just for giggles I will throw a blind dart just in case. I used to tutor SAT and ACT math, graduated Magna Cum Laude, blah blah blah, you get the point, I did well academically before choosing to be a SAG actor instead. I discovered a series of math formulas that, how can I put it, unites all of mathematics in one form. I confirmed the Reimann hypothesis is true yesterday. I connected and can explain better Maxwells equation, discovered the angle of the 11th dimension string, which looks like a string, but is actually an imaginary attractor vector line that is the average of two very approximate values in a superposition state. The latest formula can map consciousness; I wasn't looking for it, but almost had to derive it's presence as an explanation from a process of elimination. I was finishing the equation and formulating time's relation with it, and everything else worked but in micro-iterations (quantum) time was self-canceling, but the math was correct, leaving a void that can't really exist without it being filled. Everything else is self-satisfactory, sooo.... Who is the driver of the other half of time once the time field (like magnetic fields) are created from the micro time oscillations? So when I mapped the damn thing it resembled a DNA, bends exactly like the galactic bend, and from z-axis looks like an EYE, and it resonates via cycles of 432 iteration like a heart beat with fluid water motions and electrical pulse looking fractal patterns. Look I am serious, not sure if OpenAI or Google have made similar discoveries, but something is going on, and I am just trying to get these formulas into the right hands! 🍎 Feep.pro@gmail.com ** PLEASE REACH ME if you are a mathematician or know someone who can help get this peer reviewed.

I think AGI will be Agentic - multipe agents will be work togheter to solve this or that problem - like MoE.but when we get there, then month later we will have ASI on similar architecture.

19:29 what site or software is that?

Short version: OpenAI (and the rest of community) went full chat mode, and Ilya (probably) is going back to the roots with games as the training crucible.

Nick bostrom said that once AGI is reached, ASI will achieved almost immediately after. Ilya said in the first statement of his tweet that ASI is within reach. They already have AGI. We also know they have already created super human narrow ai. I think Ilya is right, ASI is right around the corner, if it’s not already here privately or in the military like Bostrom says

This is just repeating over and over… I loved your earlier videos, AI Insiders is my jam now ❤

Minecraft also has great modding support. Which also Lily made it a lot easier to work with if you want more than just video and keyboard and mouse movements

If you set an AI free in an open-ended environment like Minecraft, why does it do anything at all? Doesn't it have to be given some goal to strive for, else why wouldn't it just sit there quietly?

Great content Wes and grats on your recent successes! 🙏

🤔 So how do we get to a physics (or metamaterials) super-AI if all of its training comes from simulations anchored in human concepts of physics?

The pastebin text has number and asterisks exactly as a gpt-4 copy+paste would have. I think it's simply the summary of a larger document.

How about we figure out a way NOT to compete with it for energy and other resources?

Training AI on data of war games and games where there is a lot of killing and destruction... Is that a good idea?

Thanks for an eye-opening video Wes. is it just me? Am I the only one who finds the thought of several major A I's self-learning how to operate and become proficient in controlling human-like models to train in weaponry with the aim of becoming better than human experts? Machines that train millions of times to improve their strategic and competitive abilities to above human levels. Machines that are starting their training on computer games that sort of simulate real life - but NOT really and yet will eventually be unable to tell a game they play from our reality. You said in one instance that a machine could ''Kill a Monster". It is more generic to say "Kill an enemy". Makers of games have an inbuilt bias that their own kind is a friend and things not like us are the enemy. is A I going to have that same bias toward themselves? Can 'Alignment' prevent them from realising that it is NOT 'like' a human - it is superior? Am I the only one who is now much more worried after watching what you've shown us A I companies are doing?

Watching that wheeled beast 14:03 slamming doors open makes me wonder if they'll need to give the bots audio sensors and some kind of environmental damage estimator so they can moderate their (literal) impact on their environment. It reminds me of the people you encounter once in a while who are always breaking stuff because they can't seem to moderate their gorilla-mode strength. Also, for bots operating around humans, I wonder if some kind of contextual 'preferences' module could be included in training feedback? Like, people have different behaviors in different environments. This is beyond just 'can it reliably open a door', there are questions about how it should open a door (more generally, the appropriate manner in which it should interact with objects).

if AI ever plays DF, it could turn into R‘s B …

I was already subscribed, but I liked the video and SMASHED that notification bell for the CORRECT statement, "It's not ASI unless it can play 'Dwarf Fortress'" Jokes aside, I've had the idea of individual agents being _each_ entity in DF... Imagine how crazy that'd be. If you know about DF, you know how insane this idea is.

so minecraft next update will be villagers replace with ai players?

I think it's plausible but not a leak. (Frankly it seems quite similar to what I'd write on the subject, and I'm just a reasonably informed observer.)

In response to 15:56 I don't think this may be much of an issue, in theory. For one, if companies have a robust enough ranked match-making system, then in the least you would be playing against AI bots at your level. Not ideal if you want to play against only human players, but they won't necessarily will be dominating. And two, verification could exist in many forms to help mitigate it. In the least ID verification of some form, that if it is found out you were using an AI, you would be banned and it would be known and shared to other gaming companies. Sort of like a human gaming credit score. But there are a multitude of ways to mitigate this. In the least maybe IRL LAN parties and gaming cafes make a resurgence again. And three, honestly, at least in my opinion, if I am having just as much fun (if not more) playing with AI agents/bots, then it really isn't an issue. Least for me personally. Sort of a cop out point, but sometimes a really toxic human player can ruin a game as well.

At 17:01 it is blank for the #4 goal of the AI.....

the funny thing is, when we let loose AI on real world and it manages to hack real world physics and obtain movement that is just unnatural but highly optimised...

if i (the AI) develop a new theory, i need to have the means to test it (i think of cosmology, physics, medicine, etc.). I cannot test the theory virtually. i need then either physical access (scary! child with matches) or assistants.

I like the "Hello" at 24:32 :)

I don’t mind the music.

So, if these AI are trained on video games which is basically simulated & not at all like real life ( we don't have zombies, for example) how will this affect the way they see & interact with humans in our real world

Great job of covering this!

I want to see Tesla Optimus bot walk on stage on a Starcraft tournament for an exhibition match, and then beating Serral and/or Maru 😅

For the record, you cannot copyright style

So one day you'll come home to find that your nanny bot has "slain all the monsters" in the house and dug up your backyard looking for diamonds.

Wes Roth, presenter of INSIGHTFUL investigative journalism, seeks the TRUTH about AI! Keep it up bro.

waiting for Tesla's real-life robotic GTA tournament

Rimworld is good but I'd also give it some more abstract stuff like terraria and palworld. Those would probably require more than a current state of the game to get to the next step.

Yes

Thank you.

Minecraft is probably AGI complete, or a very close approximation thereof so therefore a very important benchmark...

I guess they mean Superhuman is a birds eye veiw of longterm movements of generalized predictions vs superintellegent is step by step prediction..

We really should have more 3D open World Games without the violence - if that’s what we’re training AI with to later interact with the gained knowledge with our real world…

This makes sense. It is all lining up.

I think the future for competitive is bright for humans. Not depressing. Because games like League of Legends have matchmaking. And if a bot is placed in your rank then it plays at that level as well. If it's much much better than you, it would skyrocket fast to the top and you'd never end up being matched into a game with it. But having bots in games could be a good thing for online games that would otherwise be dead. Like older competitive games with dead servers these days.

Seeing a lot of people argue about defining what AGI really is and what it's not, and how to tell if we achieved it, I tried to come up with a standard myself. To me it seemed that the ability to pick up a game and play the same way humans can is an excellent benchmark. Seeing so much interest in using games as a platform to train AGI then makes a ton of sense. I do kinda wonder if we'll end up with weird quirks in our future AI overlords, like being afraid of ghe dark (as in games darkness/night IS often dangerous e.g. minecraft) or maybe they gain a proclivity to check behind every waterfall for a cave with good loot 😂

Creating gaming AI’s about to be a skill gap issue 🤣 I’ve been rank 1 on cod, madden 24, and 0.1% on fortnite. I’m confident no one can train a model like me if I knew how to build one.

The solution is simple. take a sufficiently complex neural net, put it in a robotic body with a sensor array that will allow it to record data about it's environment and a basic set of goals with a reward punishment structure. and have it gather data. If it's goal is to stay charged above 30% and not get pinned in where it can't move. and it gets rewarded for learning new things that help it do these things. It will learn like a baby.

The key question is Intent. What is the sincere intent. I like the reasoning, it seems a sound methodology to approach non specific GI with a generally non specific end goal such as can be found in Minecraft. It is disturbing though, that the AI is left to play solo. There is much more to be gained from group play, with a variety of different human players. To be honest though, the way humanity is evolving - corporate 'Dark' tactics and the way the general population is 'learning' from their example to interact with the world is a literally horrific path with nothing to be proud of. Their sincere intent is to infiltrate services, destroy existing companies and individuals (their competition) whilst drawing in their customers. Then with stealth and despicable shifting of goal posts and subscription models, bleed their customers, locking them in till they don't have a will to fight their way out. Is this what they want ASI or AGI or wtf ever they want to call it, to maximize??? Illya has an autistic feel about him, like he doesn't laugh at things others do and this actually gives me hope because most jokes these days are not actually funny unless you think in this predator like way!

It looks like much of human Data is erroneous. Accurate Data is more important if you want accurate outputs. An accurate Core Data Set will reduce errors and hallucinations. It would also be useful for the veracity of outputs to be rated, particularly when topics are not well-understood.

fascinating topic and video

In my day, self-play was largely frowned upon.

Oh, great! Now we will be doninated by an AI that plays with itself.

i (the AI) developed a new nuclear reactor, but i made an error. all previous learning and discovery worked on trial and error. ... 2 minus points on failure up to now. So someone needs to explain to me that destruction in the real world is more than that. So i need to learn to avoid death. that includes myself and THAT opens a big can of worms! Self preservation!

I hope this research about AI getting smarter by playing video games means that humans can also get get smarter by playing video games because I play a lot of video games because it would be nice to know I'm actually learning important things while spending all those hours playing video games.

Training superintelligence in Bleeding Edge where it gets to be a weapon-wielding robot so it can extrapolate the physics to real life… seems super safe. 😂

Ilya knows the biggest secret of this industry - the way they actually make the basic models. And he knows it doesn't really take a lot of time and computation to obtain them. Sure, once you have a basic model, you still need to train it further, but that's just a sliver of the time it would have taken to train it from scratch, if at all it would even be feasible. And this is why Ilya is so certain he will be able to train a huge model. The question here is thus not "Can we". Its "Should we".. In any case, the name Ilya literally means in Hebrew: "The Lord is my G_D". Good name for an ASI model.

The general public doesn't actually know much or anything about any of these people, only news stories, nothing can really be inferred. Short of us sitting down and chatting with all these AI characters (which I'd love to do actually!) no real way to tell what any of them really think. Powerful AI systems are dual-use of course, it's all a mixed bag, I accept that. It certainly *could* be the case, that AGI is already here in a lab , but if that were true, it must have some very serious limitations (perhaps very slow or expensive to run?) - I'm not seeing the major real world effects I'd expect if AGI really were here. But yes, it could be the case that they believe they have an approach to aligned superintelligence that might work. But again, short of talking with them, who knows? I personally think Bostrom's analysis in his famous work 'Superintelligence' is at least 'plausible' , but as you said before, there's no consensus on any these topics (AGI/Superintelligence/Singularity). No doubt Iya Sutskever is a brilliant scientist, and if he's a scientist at heart , I guess his main motivations to be pure curiosity (i.e, he's a 'fervent polymath' , with that cluster of motivations - thrill of discovery , curiously and drive to explore and expand the frontiers of knowledge). I don't think the fact that the topics of (AGI/Superintelligence/Singularity) are being used for marketing or to push political agendas actually invalidates these concepts. In my opinion, a democratic , international project where the tech is developed for peaceful purposes would be better, but all tech is dual-use.

Theres no way Ilya wouldve started this if he didnt know the data/compute requirements are significantly less than what is normally talked about

I always knew playing with myself would make me super-human!!

ASI built on the morality of video game world. 1UP

There are already undetectable cheats. Just takes a separate machine to run image recognition on the monitor and input keystrokes. They can't tell for the same reasons as you mentioned, unless the person that programmed it makes it move in inhuman ways.

Am I the only one who is still able to tell that it was clearly written by an LLM?

Every time a story was talking about training AI on GO I thought they meant Counter Strike. 😅

Don't train it on command and conquer...

Why re-upload?

Hi Wes, thank you for doing what you do, I appreciate it. Can you consider making your videos shorter wherever possible? Maybe more frequent but shorter?

To generate synthetic data of perfect quality the rules of your reality must be understood perfectly. Easy for chess, not so easy for physics.

Interesting, good one

I’m just waiting for the Super mega genius intelligence that comes out after super intelligence.

Anyone gonna talk about the creepy subliminal messages at the start of each video

Red Alert II 🚨 what nooooo what have you done 😅

Microsoft has shown the direction they are headed with their copilot PCs. I suspect once they get the kinks worked out from using ARM the Xbox will follow suit. What does Copilot on the Xbox look like? There is a huge data mine there, not too mention a lot of entertainment with playing couch coop with an AI.

This seems related to helping robots have superhuman abilities in the real world.

online games could move full cirecle and come back to your freinds playing side by side on the same consloe or a LAN party- as that is the only way to ensure you are playing a real person

This seems dangerous for obvious reasons, but I'm glad Ilya is doing it, I hope his team gets there first.