I just started a gaming channel: ru-vid.com/group/PL7Z4ZOfI3Ap-sMfzcBUloGlkWXbfxJlrJ If you enjoy gaming, I hope you check it out. I'm VERY early in that progress, so don't expect pro level gaming :) But, the goal is to eventually add various AI concepts and showcase AI games and progress. It's a long road ahead, but check it out if it sounds interesting.
You really jumped right in and I felt like I had missed a part 1 or something. I would love to see a longer video with some more preamble and context. Great work.
Thanks! I would look at this one: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-oAsJVZlDOgQ.html This is a similar software, made by Microsoft, but functions in a similar way. I did a full deep dive into how to set it up and get it running.
Yeah the first chapter of the video is called "Cold Open" and is actually an excerpt from later on the video. The real start of the video is at 0:38 where it says "Let me show you what AGI is gonna look like". I understand the premise of a cold open but this one was rather confusing!
Would be interesting to mix different agents like ChatGPT 3, 3.5, 4 bart, Wolfram alpha... and give the HR department different prices for different kind of employees and some "budget" it would be interesting to see what kind of management strategies emerge what kind of "employees" end up in what kind of job
As a programmer, and after utilizing LLMs to write better code than I ever could (with 20+ years of experience)... I needed that positive take. Thanks :)
I find it gives an answer, which is often subtly wrong, and can't handle complexities. I have 40+ years of experience. Do you find the same thing? It is *amazing* when I'm, eg: working with a new language, it saves a ton of time.
@@jeremykothe2847I find it is directly related to your prompt, and the iterative process of prompting repeatedly. So, if I see a simple problem with the output, I will fix it simply by re-prompting, with an added part about avoiding the issue in question. This might be something simple like a syntax problem, or complicated like "I don't want to use that library". On the other hand, taking the above library choice example, I will often solve those problems by iterative prompting. It also seems to help if you back away farther from the metal. By this, I mean act like middle management. Literally, over a long-ish series of back and forth 'conversation', I have been doing my best to let the LLM do the specifications, planning, testing, etc. as well. Oh! Also, always tell the LLM to restate/summarize the current progress on the project every time it responds. And if it doesn't summarize correctly, tell it so, and make it respond again correctly, preserving the info/progress appropriately. Also, if you can get a good response chain going, where it is clear the LLM 'understood' you when you fixed a problem with it's previous response, save it. You can 'replay' it later (for example, the whole 'summarize your progress' spiel). This is important, because sometimes you will explain requirements in just the right way, and the LLM works perfectly, and you want to capture it, for lack of a better way of explaining. Oh! Also, the way you define the LLMs 'role' from the beginning is very very very important. This can also be fine-tuned throughout the conversation by simply telling the LLM and guiding responses. Oh! Also, be certain to use good grammar and spelling, and be polite but demanding. I believe it aligns you with the best training data that was created by intelligent and competent users. *shrug*
I never thought all those hours spent playing 'Donkey 'Kong' and 'Zelda' was in preparation for my future as an AI development Agent! What a time to be alive!!
Next step would be to simulate the market by creating agents that buys and use the products created by these virtual companies. In this way, you can predict how successful the product will be.
That is very fitting. In a video of Mr.Sci-Fi, he elaborated that he thinks that large companies are already a form of superintelligence. Performing tasks that no individual could at that rate or at all. Also, alignment with human values is already a problem. So, if companies can be viewed as a form of intelligent agent, it makes sense to me that visualizing AIs as companies is a good analogy.
True almost by definition - even ant colonies are examples of superintelligence from some definitions. Hopefully coprorations are at least that sophisticated :-)
I just found you in my AI journey and this video is amazing. I have been trying to cobble different tools together to do this myself, but I love this. This reminds me a lot of Game Dev Tycoon game that I love. Imagine creating a social media brand and you have personafied every one of your workers.
Truly incredible tech. Image the possibilities when these AI agents are trained to use software like Unreal engine 5, Unity, Z brush and all the other major game development software. Game creation will become truly democratized and non-coders can become game creators. What a time to be alive!
i think agi will be something like open interpreter running in a loop, querying its local llm what to do next. it records its own thoughts and saves it for later reference (database) we are almost there
I don’t think we’re “almost there.” What you described has already been built and it certainly didn’t give us AGI. Don’t underestimate what true AGI requires. I do believe we’ll get there within the next 10-30 years but we still likely need quite a few more breakthroughs.
As karpathy says it could be more than that, humans only need languages to code. Machine can take it to the next level write in even in machine code or llm acting as tools replacing conventional systems. When power or RL with better chain of thoughts and fine tunings in place. Focus is all it needs
Don’t limit to local. Imagine if each agent could make requests to specialized DNNs. This could be useful for multi sensory content. I could be wrong since multimodal models seem pretty awesome.
Thanks! , you made me understand that what I am building with my multiple API calls with multiple personalities with presite expertise is the right path! It’s more costly API wise but you can get really fantastic results with little input for execution!
The best user interface, in terms of efficiency, is undoubtedly a BMI (brain-machine-interface). This video does an excellent job of keeping us up-to-date on agent development. Thanks for the insightful content!
What about the standard one where you just talk to it, and it talks back? At least then you can just say no when it hallucinates or gets hacked. The 5 senses are the human interface with the machine of reality. Bypassing that is a downgrade, so you lose control over your responses. It's not like we think that much faster than we talk or listen, as we talk and listen at the speed we think, not limited by how fast we can move our lips etc. If someone talks too fast we don't understand, not because our ears limit us, only our brain speed. Imagine sowing your arm to your car's steering wheel for the added benefit of not having to hold on. What could go wrong? Mind-reading apps might be useful, but not mind-writing ones, at least not direct to your brain tissue. If it were beneficial to let stimuli write to brain tissue without an intermediary organ, like eyes and ears, we would have evolved like that. Mind-reading apps would be the apps evolving better, not us. Plus, you'd have to end up having an electrode in every braincell to avoid the brain itself acting like the intermediary organ, replacing the ears and the nerves that connect them to all the relevant parts of the brain that the info needs to reach. Further, you'd bypass all the biological tools that clean the signal etc, and the machine would have to be able to do all that work instead, however, the signal from the machine is not as wholesome as the one from nature, so it would be underwhelming anyway, like watching TV vs real in person viewing. BMI is about downgrading you to bypass your safety and control mechanisms.
@@Sulayman.786 I fully understand your points and concerns. It's true that the brain would just be another LLM for the agent, which is meant ironically. Current implementations of agents emulate the human as a user through another tool, but one could also use the LLM interface for this. With appropriate filters, the speed would automatically adjust to the brain. However, I agree with you that all of this is very dangerous. My original comment was more from a technical perspective. Otherwise, we could of course say, "Welcome to the Matrix". It's important that we are aware of the potential risks and ensure that we take appropriate safety measures to minimize them. Thank you for your insightful response and interesting discussion.
The best quote from this is “AGI will be a game”. You are quite right sir. Game theory will drive AGI just as it drives cryptocurrency. People need to change the way they see the world…
All very mindblowing and I love it - my big concern short term is we see what happened with SEO and Social Media both in terms of spam, propaganda, rabble raising etc. Spinning up massive bot farm factories becomes trivial.
Your videos always leave me conflicted. Like, on the one hand, there goes my job, but on the other hand, here comes my new better job (hopefully). Becoming a solopronuer is becoming ever more possible, yet so will competition. Exciting times. Truly every passing day I think to myself, I should quit school and start building like mad.
This and most people who start, or think of starting never clear the first hurdles. Its easy to get to the edge of the abyss and look into depths, its a whole lot harder to take the leap.
Looks fun. Id imagine we would have multiple different types of interfaces for different people but the gaming one looks appealing to me. The idea of having an army of NPC's is cool.
Mind blown! Running a business would be like playing sims. but i guess when it gets to that point running a business would be completely different from what it is now and the current way will be irrelevant
Thinking about the multi agent framework: wouldn't that be a great idea to solve memory issues. Have a dedicated agent whose sole purpose it is to record and condense the whole conversation so far and then another agent that takes this condensed info to reapply it to the original conversation. And depending on limitations you could split that process up between many agents
You could take that further and have the documenter know how to query a database where you store pieces of info based on some system of key words or something.
From a neuroscientist perspective this seems reasonable. Everything digital that we interact with can be seen as a neural prosthesis adapted to our common sense understanding of and interaction with our natural environment. Complexity is hidden behind easily understandable metaphors or actions. Building custom general AI models might just be best represented with natural language and video game characters.
When the agents start interacting with every major platform; Meta, X, TT, YT, Reddit, LinkedIn, and maybe any social platform you can think of; then this will be a powerhouse that can't be stopped. I'm looking for agent developers, but will only work with people I can talk with face to face. People I can feed, laugh with and really build a visible community with. The power & secrecy of the corporate and political arena to take over basic living, control populations, and extort the wealth out of us has to be phased out.
no developer like that will sit with you face to face! the agent network structure is already set up with payment protocols and a Marketplace for agents. Time is money to developers and when there's a marketplace for AI agents.... you'll just have to put in your bid like everyone else.
"The power & secrecy of the corporate and political arena...has to be phased out" And AGI/AI is going to maximize the elites power consolidation way faster and more completely than what they will ever allow you to do with it lmaooooo Use your brain and look at the writing on the wall, bud. No wonder why this world is totally done for...geniuses like you walking around lool
Just last week I was discussing with ChatGPT how I wanted a virtual production movie studio that responded better to the director's vision and in the conversation I pitched to it that it should use agents modeled after traditional roles in a film production. I thought I had invented this concept but clearly I wasn't the first one to think of this concept.
Unfortunately, in my experience these things do get stuck (with me being in the loop and just posting successive tracebacks). GUI bugs are a big problem, because there's no traceback for it. You really need a way to make sense of the visual feedback by expressing it in a verbal description somehow. One unsticking strategy I found that GPT uses when coding shell scripts is it falls back to different approaches when you tell it that it failed, for example, I told it to write a shell script to recursively copy folders and contents but leave out files > 1M. I told it that it didn't work, so its solution was to rewrite it as an rsync command, which then worked perfectly the first time. I really think until we get feedback from visual parsing, GUI's are not going to work very well.
Great overview of Society of Mind Wes! I like the shout out to Marvin Minsky's book - will have to dust that off. I love the approach of a MoE multi agent system with LLM IO and a Code/Test loop. You are staying always on the edge of what is AGI capable - much appreciated!
This could likely be used to better mimic the processes of human thought. i.e. how we break up a problem and process it through the different stages to reach a certain goal. This reminds me of a video on "split-brain", where it's used to theorize that consciousness could possibly be split up into multiple parts.
If AI get's a memory system/ management upgrade it could do wonders. If it could increase the potency and range of it's input and outputs it could increase the range and potency of its abilities and what it can achieve. If it learns to build these studios dynamically it could achieve a scary amount of things.
Do you know that episode of Rick and Morty where there's a whole universe inside Rick's battery? You reminded me of that. In the future, we're gonna have a whole universe of 'minions' working for us to deal with our shit basically
Dude, how can you make a video with such gigantic implications only 9min long???😮😮 Edit: I'm starting a student assistant job at one of my profs chair. I'll have to work with pytorch to develop a neural net for one of their research projects. Considering the costs mentioned in the video, I might try implementing that
Absolutely phenomenal. “We live in the best possible timeline of all time” Agreed 🥳 the fact that AGI will be like a video game is actually quite mind blowing. I am impressed with the state of technology and so excited about all of our futures!
I have used automated AI agents before ... i didnt get it on a recording but my test lasted about 24 hours on a fully automated mode. I asked it to design me a shooter game with aspects of a diablo game. It learned how a shooter game is built and it learned how diablo is built. Figured out the features to pull from each ... found a way to download blender.... downloaded unreal engine 5.3 ... was teaching itself how to design stuff in those. If i would of let it go on its journey to a complete project i have no idea what i would of ended up with. I mean its cool and i didt this maybe 7 months ago.
It's like from the movie The Matrix, where you have the Architect, the Oracle, agent Smith, other agents,... Each one performs their job with a common goal. It looks like some sort of precursor to The Matrix.
Great info....got me thinking...I can build my own super biz with my ai super agents working for me.... this advaned tool can be THE tool to help the most people.....
I think at this point you should be as immersed as possible if you are trying to learn and understand a complete character, teacher, I guess understand a different experience. I think that could be a huge problem solving agent for future generations.
remember when ai was just weighted nodes running mario over and over? and now ai agents will be working full agencies in games. is it still a game mom?
aw dude, this is such a smart implementation of a neural network... a self learning model based on driving sales or production in a business! damn bro! lets see how it goes!
ok yeah i spoke to soon but hey its DEVELOPING INNIT ... looks like me using gpt to learn pygame/tkinter but the way you talk about the ideas bouncing back bewtween testing and code because it doesnt work at least automates the manual part of using GPT( eg i have this error, can you fix it: *copies and pastes code, waits for 2 mins for it to regen all the code..*)
ChatGPT 3.5 can a write winning tic-tac-toe game in just 2 iterations, no team needed. The first iteration was a working game but no CPU opponent, the send prompt got it to beat me.
Think of your own body. Your mind is the executive function of the rest of it. Without it, your body does nothing useful. When your employer pays you, he is paying you first and foremost for your CEO: your brain. Of course you are also getting paid for your arms and feet and strength and skill, but all of those are trained by and powered by your executive function. CEOs are supposed to get paid the most. Maybe more people should become CEOs :)
@@mami_tomoe The C suite is the brain. The CEO is the ultimate decision maker. But yes, the CTO is the ultimate “thinker”, and if the CEO is any good, he will make decisions very much with the CTO, the COO, the CFO and others in mind. But it is still the CEO making the decisions. Including when to hire any of those other guys
Using a game like environment for these AI agents to interact will also open up countless possibilities for the different ways we interact with them. Imagine if you could watch the agents in real time and see which "room" (i.e. which conversation and with which other agents they were with) they were in, then add or remove agents from those rooms as you see fit, experimenting with different agents with slightly different directives to see which ones produce the best end results. You could even jump into a given room yourself and take the place of an agent for micro-managing aspects of the process. You could add or remove rooms if you felt like the process needed more or less steps for a better end result. You could move agents to a different room to have them brief that group of agents with information about the inner workings of the field they were specialised in. Each room could also have its own database accessible by the agents in it, where they could store information as a form of long term memory. Or they could query databases of other rooms if they were missing some information. You could have a scoring system for agents where they would receive points for completing tasks well or innovatively, and where you could manually assign bonus points when you liked what one was doing. The possibilities are endless and I can't think of a time I've been more excited for a piece of technology to take off.
I want a model trained to find the best teams structure (tasks, responsibilities, governance, meetings) for any given entreprise. (Let’s not think about the fact that our world might be a simulation to find the best conditions to perpetuate consciousness)
Watching AI development over the last few years has been like watching kids grow up. Although I am reminded of Elon's quote, "We're going to need a UBI for no other reason than we won't have a choice, AI will be able to take everyone's job."
I'm usually against UBI, but even I have to admit AI might make UBI a serious thing to consider... out of necessity, like Elon said. Only thing I'd want to point out is that only solves financial issues... it won't solve the issue of what people will have to do with their lives. I think that's the bigger crisis here. People are already depressed AF, nihilistic, and feel aimless. I think UBI, work being unnecessary, and being able to do basically whatever whenever will only make it worse.
@@urphakeandgey6308 I mean we could do what we do now, increasing burocracy, Compliance, due diligence and consulting. Highly paid positions with no productive output.
@@urphakeandgey6308 I think with some humility, we could look into the issue and find that humanity already has all the answers to the problem you are addressing. It's a complete school re-haul with focus on spirituality (lol), philosophy (eastern Buddhist, Nietzsche, etc...). This way we can focus on creating a true culture of peace and prosperity, where everyone will have a chance at finding their own meaning of life. It's very narrow-minded, in my opinion, to view work (in the state it's in right now), as the primary catalyst for meaning.
@@MultiGoban The meaning of life has meaning when your path is not forged. In a world where everything is at your fingertips, that cannot exist. We are not talking about a world where everyone can do whatever they want. We are talking about a world where people can do whatever they want and everything is done at once. I feel bad just thinking about it. You don't have to go far to know first-hand that this approach leads nowhere. I recommend you read about the universe 25 experiment, perfect life. The results of this experiment have similarities with the unnatural behavior of the individual in Western society. Elon can say whatever he wants. He is not an expert in economics, sociology or artificial intelligence. It's just a person who was in the right place, at the right time and with the necessary financial resources. And, of course, innate talent.
@@jmw1500Same thing really. UBI is a physical impossibility without population control. You can't create something from nothing, and constant expansion means constantly expanding demand. When UBI becomes a thing, so will government-sanctioned eugenics.
still has a long way to go, also is this limited to the openAI api, which then if no special account; will be limited by the amount of tokens, and considering you are making code, your probable length would be 300 lines max for the code considering it goes both ways with the input and output from the model? possible to use other models(llama2,falcon,etc), didnt go thru the project yet.
Inspection of the code reveals that openai is imported in the codebase; thus this app is hardwired to use openai's LLM models exclusively. But, given that "import openai" is a library, an enterprising programmer might conceivably write a replacement library (taking care to expose the same methods/functions as the openai library, i.e. has the same api interface - once this is done, that library can be a drop-in replacement for the openai library - hence the rest of the code doesn't change)
that's a great point. The documentation (manuals, code comments etc) is where gpt3.5 works just fine. But, how much does a gpt4 CTO change things vs the coder, for example? Very interesting thing to test!
I would definitely customize some of the roles. I think I’d implement this tool more as a reviewer or editor for scripts that I’m already working on. This is pretty shocking though.
We always talk that maybe we are in a simulation. What if we are that video game? What if we are the AGI? And now, we are at the point of creating our own AGI?
Why would you simulate so many unnecessary thoughts/parameters (90% of the populations' non-technologically-advancing actions)? And how would a simulation that runs so profoundly slow (think of current time, current advancement, vs. instantaneous speed of AGI/AI systems in general) be BETTER than purely soulless, emotionless AI-based logic/knowledge exchange? Those two questions essentially refute any theory relating to "simulation for the cause of X Y Z development." (If you extrapolate the depth from those questions)
imo, it's clear that we're created (programmed through frequencies, physics, math, sonics, etc.) We're created in the image of God, and this reality is essentially a spiritual filtration system (separating good spirits from evil). The conclusion of this limited-time filtration (test; through free will) follows the "unseen code; universal code; universal (moral) law"... Similar to how objects follow laws of physics, math, wireless connectivity, etc... Is that our energy (soul) is transferred (you can't destroy energy, this is a fact) to the spiritually correct plane of eternal frequency/existence (imagine 1+1.. 2 is the mathematically correct answer... imagine dropping 25kg object.. it falls at a velocity of X ft per sec which is gravitationally correct...) ---> same way, there is a "spiritually correct" law/code that is invisible to our eyes (especially in this reality). Math "code" is technically invisible, yet the entire "Universe" operates on it. Why wouldn't there be a spiritual "code" which directly correlates to the universally present aspect of "duality"? Positive / Negative Up / Down Left / Right Life / Death Growth / Decay etc.. ... Good / Evil If we know energy can't be destroyed, only transferred, (and every human body loses 125g after losing consciousness/dying.. publicly researchable), then why wouldn't there also be a subsequent "direction"..? The next plane of reality? The next realm of life? (present in most duality of life; everything is moving, because everything is simply frequencies... objects/matter are nothing but a bunch of atoms vibrating, nothing is truly solid. This is basic physics)
🎯 Key Takeaways for quick navigation: 00:28 🧠 AGI, or Artificial General Intelligence, is expected to have a user interface that plays a crucial role in its adoption into mainstream use. 01:21 🎮 The ideal user interface for AGI is suggested to be a video game format, as studies show that AI agents perform better when they collaborate and debate like a team in a game-like environment. 02:01 💼 A project called ChatDev demonstrates how AI agents can power various departments, similar to separate AI entities, working together to develop software applications and solve problems. 04:46 💰 Using GPT-3.5 Turbo for such applications can be cost-effective, with individual apps costing as low as a few cents to create. 05:42 📄 The completion process includes a documentation team that creates manuals explaining step-by-step how to use the applications developed by AI agents. 06:50 📢 The concept of autonomous departments powered by AI agents can be extended to content marketing and other tasks, allowing businesses to automate various functions. Made with HARPA AI
Oh my God this reminds me so much of that Rick and Morty episode where there was a little people living inside the battery and then they had a battery that there was a little civilization of people living inside of….so funny
This is so cool! We could have entirely non human run companies this way- even if they perform only at say the median level of human businesses, that's still about half of human businesses that this kind of setup could outperform.
