TEST TO SEE IF AI CAN MAKE $1,000,000 (modern Turing test)

Key Takeaways

so this video is about creating an AI agent capable of passing the proposed modern Turing test in which the AI would have to successfully complete the following instruction go make one million dollars online in a few months with just a 100 000 starting investment I'm going to use Minecraft doom and World of Warcraft to illustrate my points it's gonna be good and we'll also see why the latest research out of Nvidia Google deepmind Stanford and Microsoft are showing that we're not too far away from this and pretty soon someone's gonna have it will it be me will it be you let's find out so quick story time when I was a kid I really enjoyed playing World of Warcraft but not in the same way that other people did so most people focused on meeting up with their online friends and working together to achieve complex missions and objectives while coordinating everything over TeamSpeak I did it differently the people that did what I did were aided in the game I was a botter icky is somebody who had Bots I would set up these scripts that would automate actions in the game that would make me gold the in-game version of money my characters would run around doing quests completing objectives and dealing with any hostile people that got in the way finding collecting valuable resources and then selling it on the oxygen house where other players would be able to buy it they were very efficient at it they could run 24 hours a day you can even have multiple instances of them set up all going about your business one of the best experiences was waking up in the morning looking at the computer and seeing that the bot was still running still doing my bidding still increasing my digital wealth it was like Christmas morning every morning unless the bot got stuck or fell off a cliff or whatever so when Changi PT came out uh one thought immediately popped into my head can this be used to create some sort of an autonomous AI agent make a bot that would run around the world of internet and make my digital wealth go up but in like dollars instead of gold now obviously I wasn't the only one that had that thought many people made money with some application of Chad GPT Amazon was overwhelmed with AI written books Google had to do multiple updates to SEO to search engine optimization to try to reduce the number of AI generated websites that kept popping up now some people took it to the next level they attempted to create autonomous AI agents that would go about their bidding that would create plans figure out some tasks and try to execute those tasks they would basically run these Loops trying to accomplish the task that was given to it you've probably heard about some of these there's Auto GPT chaos GPT baby AGI Etc these are still early prototypes and they're very exciting but they're not quite printing Millions on autopilot quite yet at the same time the latest research out of Nvidia Google deepmind Stanford and Microsoft seems to be indicating the following that one we already are able to create autonomous self-improving AI agents at least for Minecraft these agents can code and create tools for themselves to use to complete other tasks they can also spin up other copies of themselves to complete subtasks as needed one study showed that gpt4 can spin up faster cheaper models of itself as assistance it would be able to code tools for those assistants so they can use them and this is where it gets interesting those faster cheaper models the GPT 3.5 turbo in this case are able to complete the tasks at the level of gpt4 the more intelligent more powerful but slower and more expensive model the results demonstrate that with the help of the tool the coded tool that gbt4 made for that particular task a lightweight model like gbt 3.5 turbo can achieve performance on par with gbt4 sometime in the summer of 2023 this meeting called the AGI house brought together some of the top mines in the space including Andre carpathy one of the main openai developers these people talked about how to build these autonomous AI agents where they think these will come from is a bit interesting let's take a look so in this video we're going to take a look at why autonomous AI agents that are capable of putting income on autopilot why they're not that far off and maybe even here now we're going to see why the first people that build these will probably not be these huge labs in fact when I hear from Andre carpathy on why he thinks a small team of entrepreneurs is more likely to develop this instead of a huge Army of phds in some underground lab we're going to see a few prototypes that are working right now and we're going to talk about why you yes you should be thinking about building something like this regardless of your Tech background so let's dive in use the video chapters below to skip around if you need to but watch until the end because if you missed this train well you don't want to miss this train I'll just leave it at that chapter one The Modern Turing test so this is Mustafa Suleiman he's kind of a big deal so he co-founded a deep mine sold to Google and now he runs inflection AI as well as being a venture partner at Greylock he's releasing a book called The Coming wave which looks interesting but one thing that caught my attention is that he's proposing a new version of the touring test he's calling it The Modern touring test let me read a quote from an article he wrote but simply to pass the modern Turing test an AI would have to successfully act on this instruction go make one one million dollars on a retail web platform in a few months with just a 100 000 investment to do so it would need to go far beyond outlining a strategy and drafting some cop as current systems like the gpt4 are good at doing they would need to research and design products interface with manufacturers and logistic hubs negotiate contracts create an operate marketing campaigns it would need in short to tie together a series of complex real-world goals with minimal oversight you would still need a human to approve the various points open up a bank account actually sign on the dotted line but the work would all be done by a high he continues something like this could be as little as two years away many of the ingredients are in place image and text generation are of course already well Advanced Services like rgbt can iterate and Link together various tasks carried out by current generation of llms large language models these are the AI systems that everybody's talking about Frameworks like land chain which helps developers make apps using llms are helping make these systems capable of doing things although the Transformer architecture behind llms has garnered huge amounts of attention by the way if you missed it he snuck a little nerdy joke in there because the big breakthrough 2017 the paper by Google attention is all you need that was the thing that helped these llms be capable of a lot of stuff that they're doing it was one of the big breakthroughs it created a technology called Transformers which allowed these LMS to have something that is called attention so although the Transformer architecture behind LM says garnered huge amounts of attention the growing capabilities of reinforcement learning agents should not be forgotten putting the two together is now a major Focus apis that would enable these systems to connect with the wider internet and Banking and Manufacturing are similarly an object of development recently Mustafa was asked how much of the recent excitement about AI was hype versus reality and here's what he said what what is just being completely over High pipe and his pure speculation versus what's legit look I mean here's the reality right look around you today everything in your line of sight right now is the product of intelligence humans collectively individually have created everything that you see around you and so the potential to take what has made us special our intelligence and try to automate that paralyze it speed it up and make it widely available to billions of people on the planet is I think one of the most exciting missions that we could possibly do so I think that's very reasonable for us to think over the next 10 to 20 years this is going to be the greatest Leap Forward in productivity that our species has ever known if anything we're understating the real potential of this new wave of AI in the next couple decades now as you probably realized somebody hitting this modern Turing test somebody passing it would mean some pretty big implications for the world it would fundamentally change the global economy you might even change change the way we think about what money is what jobs are now obviously there would be a big difference between this Tech being available to just a select few versus being available to everyone if a few people will have it they would be able to redistribute a lot of the wealth to themselves once everybody has it the advantage isn't quite as big it still would increase productivity and tons of other stuff but imagine if you had this thing right now how much wealth could you print before everyone else caught on now my only gripe of this monetary test as it's called as it stated is the part about the retail web platform now I've been doing e-commerce since 2011 and while I'm very excited that this is the domain that was chosen to host this AI challenge in I would like to expand that definition a little bit or or at least clean it up a little bit some of the things that I would like to see mentioned I'm gonna end up buying the book to see exactly what he's talking about what the details are so it looks like September 5th 2023 is when that's going to be available but what I would like to see is one so the retail part I think is needlessly restricted for example so we saw in newsletters like the hustle get acquired HubSpot bought it for 27 million dollars morning Brew another daily email newsletter got sold for about 75 million the model is very simple acquire customers largely with paid ads but also by sort of getting the virality going by having the users refer other people to get a subscription at the end I at the end I believe they had 2 million people doing about 20 million a year in profit and sold for 75 million which is I would say Kafkaesque now this would not fit under a retail web platform restriction but certainly an AI run newsletter that grew to earn millions of dollars a year maybe you should be considered to have passed this modern touring test if it selected articles wrote a little summaries of it wrote up its own exclusive stories went out there and acquired new users tested different ways to get those users to interact and for new users Etc certainly if an AI was able to build something like this by itself we could say hey it passed this thing it passed the Turing test now at the same time Twitter is paying out I believe 50 Revenue share to verified users so let's say as part of the marketing promotion the AI would use Twitter as you know as part of its marketing strategy to get more users to interact with them Etc would that Revenue count towards the one million dollar goal now and this is my guess but I believe the reason that Mustafa might have purposely excluded anything where you can make money by getting paid to show ads the reason why that was excluded is because that could be gamed by sort of these sort of the less intelligent less capable AIS I mean something like that could be probably pretty easily done right now you know create a website have the eye populated content put some ads on it you can make some money with that already and certainly I don't think that should pass the modern joint test and I think that's why Mustafa specifically phrased it in a way that exclude that from it but I think my point is that can we replace the retail web platform of something broader while still retaining the same challenge level for the AI don't get me wrong I love the idea I love it as is a lot but if we can nail the definition just a little bit better if we can expand it while keeping sort of it as rigorous as this current definition is or at least add some sort of clarifications I think that would be great but either way I think that this modern Turing test is an excellent test an excellent sort of milestone for AI development all right chapter two the research so let's look at a few studies that recently rolled out that deal with autonomous AI agents now I've covered a lot of these in previous videos so take a look at the video description if you want to do a full Deep dive in there but here are some of the more interesting ones so the first one is Voyager AI Minecraft Agent so the team at Nvidia built an autonomous AI agent that plays Minecraft we had other attempts at this but nothing quite like this so This bot runs around the world of Minecraft and tries to become the best Minecraft player ever that's literally The Prompt that is given to Chad gbt which is the brain that controls the Minecraft character now Chad gbt or more accurately the model behind gpt4 it's fed some data about its surroundings in Minecraft it's told about the biome it's in what type of data is how hungry or healthy it is what it what it has in its inventory Etc it then reasons about what it needs to do next to become the greatest Minecraft player for example it has a wooden fishing rod and it's next to a river it might decide that this would be a good time to upgrade the fishing rod and then go fishing once it reached that decision it passes that decision to another version of gpt4 another instance of the same AI model that model is tasked with completing that objective more on that in just a second so the character actions are controlled through mind flare API which is which allows Chad GPT to carry out basic functions like move around mine or kill bad guys Etc so the second AI tries to complete the objective and will often write new code to interact with the API in order to complete it this is important let me restate that when the AI reaches a task that it can't do yet a new novel task it will attempt to write code in order to complete that task so for example if it reaches a tough enemy it might create a series of actions where it equips the best armor it has the best Shield it has it get gets out its sword then it rests up it eats to make sure it's at full health and then it engages the enemy and kills it then it picks up any valuable resources that the enemy drop and then test to see if that skill worked or not if it's if it's effective and if it is then it adds that skill to its skill Library an Ever growing sort of collection of skills that it's learned that can use later to get through the game so this AI runs completely without human help it decides on what to do based on a very broad objective AKA become the best and then it codes different skills for it to use tests them and then save them for later if they work well the results are stunning unlike other state-of-the-art AIS in the space this one goes further accomplishes more and just overall outperforms the competition the important thing here is that other AIS in general they'll hit a plateau somewhere meaning at some point they can no longer make meaningful progress they can't get better Voyager tends to keep improving so this I think is an example of a fully autonomous AI agent that is as a bonus is self-learning and self-improving the thing that made this work was largely using regular English language to explain to Chad GPT what you wanted it to do a lot of the instructions for this was written in English not computer code a lot of software that was needed like the Mind flare API that was already built and available you have to read the tutorial to pick up some of the basic commands and understand how it works but you you don't have to code it from scratch so importantly the thing that makes this tick is English language or any other language really but my point is it's natural language it's sort of the same how I'm talking to you right now it's not a computer language it's not code it's not even a certain type of way of speaking like some legal form or poetry or whatever it's just you just say what you want you tell it go be the best and it gets what you want it tries to do that chapter three next up we have a study from Google and Stanford that's called generative agents interactive simulacra of human behavior in it AI researchers created 25 agents each powered by its own chat GPT its own gpt4 they live in this little village so think of the Sims I think as a good illustration that's what they're comparing it to although I think stardew Valley would be a better comparison that's just me each one of these 25 agents it's given a backstory and what activities they do on a daily basis as well as who they know if they're related to anyone else in the village Etc there's a brief description of what that looks like but it goes something along the lines of John Lin he works at the pharmacy between 9am and 5 PM his wife is Samantha Lynn and he is friends with June Park and he likes to go photographed birds in his off time his interests are politics and Birds by the way as of this recording this town is still running and you can go see it online live I'll include the links below so the characters go about their day trying to follow their schedule although they're Sometimes Late and sometimes they'll Miss events that they're committed to just like real humans would so they talk they make plans they make new friends like they really act like real human beings would they have goals and responsibilities and places to be overall it's very believable except for the fact that everyone is a very polite and positive like no one's experiencing doubt about whether they're living in a simulation or not or having some inappropriate thoughts about their neighbor that part's not in there the researchers for the most part do not interfere or change the environment in any way except for one thing at the start they ask one character to do something that character is given an intent a suggestion it makes them want to achieve some and that thing is to host a Valentine's Day party so normally in a video game development in order to do this you would have to script the behavior of every single one of those characters those NPCs non-player characters you would have to script them to show up on that particular day you would have to script how they're interacting what they're doing you would have to code tons of things you would have to test to make sure that code worked it's a lot of work here they tell one character hey go do this on this day that's pretty much it so the important thing to understand here that they're not these characters aren't really coded to to do anything really all of this done it is done with Chad GPT Chad GPT is given sort of like this is who you are this is what you do and then similar to Minecraft Voyager the game kind of tells it where you are sort of like you wake up in the morning what do you want to do and chat GPT goes well let's go get some coffee let's go say hi to the wife and then let's go to work so this is important there's no scripting going on here there's no hard coded anything it's just they do whatever Chad GPT determines they would do now also interestingly Chad GPT will respond to certain things in the environment so for example if the if the food on the stove catches on fire Chad gbt will say okay stop whatever you're doing like let's say we were brushing our teeth stop doing that go deal with the coffee all this to say is there's really not any scripting going on here this is all just kind of unfolding and so what this means is that there's tons of ways that this could fail when you tell One agent to go do something that requires a lot of other people a lot of other agents to cooperate tons of ways to fail so the agent who is tasked with this could forget or get distracted or simply try but fail at doing it the people that were invited could forget to come or just refuse to come but in the end the party does occur with many agents coming and taking part in it by the way not everyone heard about it from that initial person that was initiated with the suggestion there was information diffusion she in this case Isabella that was the original agent told one person that person told somebody else and so the message spread all over the town just like it would in a real sort of social human situation so the new spreads people make plans they ask each other out they work together to make it happen on that date while still at the same time taking care of all the other stuff that they have to do on a daily basis so again one Chad GPT is told to make something happen and it coordinates with 24 other agents to make it happen again here most of the structure is built with simple English now the researchers did build a wave for the ages to remember what happens and to be able to sort of sort through their memories to be able to reflect on what happened basically each little detail is assigned and importance then those memories get stored and later gbt4 kind of looks at all the important memories and tries to come to maybe some new conclusions Etc but anyways what do these two studies show in terms of how close we are to AI agents making a million dollars without much input a few years ago if you asked this question most people would think this is just science fiction there were at least decades away from anything that's even resembling this a lot of people even in the AI field would push that sort of projection to maybe a hundred years away but these studies they some of these are rolling out a few months after the official release of gpt4 while they seem to indicate that the Technologies already here now we just need to add some tools and architecture to it we need to figure out how to get the right response out of these AI systems but overall I mean I would say we have all the Lego pieces we just need to put them together in the right order now one thing that would be really helpful for this for the modern Turing test is to have Vision meaning for something like Chad gbt to have Vision to be able to see for example a website understand what's on it and then interact with it now there's a lot of approaches that use various hacks to kind of get around this we saw Voyager AI use an API to sort of substitute for vision we can use Python and something like OCR for character recognition but if gpt4 could just look at the web page and interact with it that would be a huge deal now open AI did say that gpt4 has vision and they kind of promise to roll it out soon but it's it hasn't still been it still hasn't rolled out yet now the code interpreter has some of those abilities kind of baked in you have bar two you have Claude you have these other models that kind of are able to do something similar to Vision where they could look at an image and kind of describe what it is but we're not quite there yet after now we we have some workarounds but we don't have the full vision yet all this is to say that this idea of the modern Turing test that it could be beaten in the next two years that's not crazy talk that's not crazy it's possible it's very possible especially if we get Vision within let's say the next year now with that in mind let's take a look at who is most likely to stumble upon this first because of course if Google and Microsoft and all the big corporations if they figure it out and decide to withhold that from the rest of us Mortals well that doesn't really do us any good so chapter four Andre carpathy and the autonomous AI agents so Andre carpathy who has been working on building autonomous agents for a while now and he's one of the main guys over there at open AIS he joined this Meetup called the AGI house hackathon to talk about why you should be building an autonomous AI agent let's take a look at some highlights from that and this was 2016 or so and is that price of the day actually where our own agents and so everyone's really interested in building agents and obviously didn't work so the technology was just not ready and that was not the right thing to work on at the time and so it turns out that the right thing to do at the time was actually to forget about AI agents altogether and start affiliate language models and then language models now we're back here five years later and so the way you would approach these problems today is a completely different in fact all of you are working on AI agents but you're not using any reinforcement learning probably and so that's so crazy and I don't think we would have anticipated at the time it's just the way this played out it's very it's very interesting so I want to spend a bit of time on like okay what is causing all this High I think obviously all the reason that all of you are interested in this topic is that I think that very obvious to a lot of people that AGI will take the form factor of some kind of an AI agent and it's not just going to be a single agent and there's going to be many agents and they're going to be in organizations or civilizations of digital entities and I think it's just extremely inspiring to sort of like thank you it's kind of crazy it's clear that language model is a part of a solution but you know how do you build an entire digital entity and that has all of the cognitive tools that humans have so obviously we all think we need some kind of advantage of the system too to actually like plan ahead and think through and reflect on what we're doing finally I wanted to end with some words of inspiration what's interesting and not obvious is that you guys building AI agents are actually at the Forefront of capability of aiation today and all the big laughs like lln labs open Ai and so on I suspect are not at the edge of the capability you are at the first benefit so opening up for example is very good at training investment Transformer language models so as an example one way to put it is if the paper comes out that proposes some different way of training and Transformer the internal socket open AI is something along the lines of oh yeah someone tried that two and a half years ago and here's what happened and here's why it didn't work and it's very well understood and very very well mapped out but when a new agent paper comes out we're all interested in we'll look at it and we're like oh that's really cool that's another one and that's because you know the team didn't have like five years to spend on it and it's competing now with all of you and the entrepreneurs and hackers the song it's really hard to do so yeah I think it's really inspiring that you are at the edge of capability and on something that is obviously very important in transformation and so with those words I think I'm eager to see what you guys build foreign I will leave you with a clip from a person who I believe created one of the most influential video games in history and why he believes that AI super intelligence might be solved by just one person in the next video we will talk about what skills you need to start building autonomous agents and I'll show you the exact business that will be the first to win the modern Turing test then turn 100K into 1 million subscribe so you don't miss it and here's one of the smartest Tech Minds in our industry on why AGI will be a lot easier than we think here's John Carmack one of his in my opinion greatest Creations was a little game called Doom it is likely that the code for artificial general intelligence is going to be tens of thousands of lines of code not millions of lines of code this is code that one individual could write the artificial general intelligence side of things it seems to me like this is the highest leverage moment for potentially a single individual potentially in the history of the world where the things that we know about the brain about what we can do with artificial intelligence that it is likely that the code for artificial general intelligence is going to be tens of thousands of lines of code not millions of lines of code this is code that conceivably one individual could write unlike writing a new web browser or operating system and based on the progress that AI has machine learning has made in the recent decade it's likely that the important things that we don't know are relatively simple there's probably a handful of things and my bet is that I think there's less than six key insights that need to be made each one of them can probably be written on the back of an envelope we don't know what they are but when they're put together in concert with gpus at scale and the data that we all have access to that we can make something that behaves like a human being or a living creature and that can then be educated in whatever ways that we need to get to the point where we can have universal remote workers where anything that somebody does mediated by a computer and doesn't require physical interaction that an AGI will be able to do foreign