"AGI within 18 months" explained with a boatload of papers and projects

hey everyone David Shapiro here with an update sorry it's been a while um I am doing much better thank you for asking and thanks for all the kind words um yeah so a couple days ago I posted a video where I said like we're gonna have AGI within 18 months and that caused a stir on in some corners of the internet um but I wanted to share like why I believe that because maybe not everyone has seen the same information that I have so first Morgan Stanley research on Nvidia um this was really big on Reddit and basically why we are writing this we have seen several reports that in our view incorrectly characterize the direct opportunity for NVIDIA in particular the revenue from chat GPT inference we think that gpt5 is currently being trained on 25 000 gpus or 225 million dollars or so of Nvidia hardware and the inference costs are likely much lower than some of the numbers we have seen further reduce reducing inference costs will be critical in resolving the cost of search debate from cloud Titans so basically if chat GPT becomes much much cheaper then it's actually going to be cheaper than search um is is kind of how I'm interpreting that now this paper goes on to say that like the industry is pivoting so rather than seeing this as a trendy new fad or a shiny new toy they're saying No this actually has serious business implications which people like I have been saying for years but you know the industry is catching up especially when you see like how much revenue Google lost just with the introduction of chat GPT um I like this and we're not trying to be curmudgeons on the opportunity so anyways Morgan Stanley Nvidia and I've been I've been uh on in nvidia's corner for a while saying that like I think they're the underdog they're the unsung hero here so anyways you look at the investment and so this reminds me of the ramp up for solar so 10 to 15 years ago all the debates were like oh solar's not efficient solar isn't helpful it's too expensive blah blah blah and then once you see the business investment going up that's when you know you're at the inflection point so AI is no longer just a bunch of us you know writing papers and tinkering when you see the millions and in this case a quarter of a billion dollars being invested that's when you know that things are changing and so this reminds me of like the 2013 to 2015 uh range maybe actually even like 2017 range for solar where it's like actually no it makes Financial sense um but of course everything with AI is exponentially faster uh so Nvidia is participating they've got the hardware they're building out the big computers so on and so forth the investment is there so the Improvement is coming the exponential ramp up is coming now that's great uh one tool let's take a quick break and um when I when I talked about uh n8n in Nathan or naden I'm not sure how people pronounce it as well as Lang chain people were quick to point out Lang flow which is a graphical interface for Lang chain so this is this fills in a really big gap for Lang chain which is okay how do you see it how are things cross-linked so I wanted to share this this tool it's a github.com logspace dash AI uh Slash Lang flow so you can just look up Lang flow and you'll find it so this is a good uh good chaining tool a nice graphical interface this is exactly the direction that things are going um great Okay so we've got the business investment we've got people creating open source libraries it's going it's advancing so I wanted to share this paper with you uh mm react for uh was it multimodal reasoning and action so this basically makes use of the latest GPT where you've got vision and chat um and it's like it's kind of it's exactly what you what you kind of expect um but this page does a good job of giving you a bunch of different um examples and they're uh I think they're pre-recorded is it playing it looks okay there it goes um so you can check out this paper the full paper is here and there's a live demo up on hugging face so you can try different stuff and then talk about it um which is great like the fact that they're able to share this for free just as a demonstration is just a hint as to what's coming um because imagine when this is commoditized you can do it on your phone right your phone's Hardware will be powerful enough to run some of these models within a few years certainly if it's uh if it's offloaded to the cloud it's powerful enough to do it now um and then uh so you when you stitch together the the rapidly decreasing cost of inference these things are basically going to be free to use pretty soon when you look at the fact that an open source framework like Lang flow and and uh and so on can allow pretty much anyone to create cognitive workflows and all these things it's like okay yeah like we're gonna have really powerful machines soon and so someone asked for clarification when I said okay well what do you mean when you say AGI within 18 months because nobody can agree on the definition and if you watched the Sam Altman Lex Friedman interview he ref he refers to Sam Allman refers to AGI several times but the definition seems to change because early in the interview he talks about like oh you know you put someone in front of gpt4 or chat gpt4 and what's the first thing that they do when when and these are his words when they interact with an AGI is they try and break it or tease it or whatever and then later he says oh well gpt5 that's not even going to be AGI so he keeps like equivocating and bouncing back and forth I think that part of what's going on here is there's no good definition and because later in the conversation they were talking about things that a chat model can do it's not autonomous right um but I'm glad you asked reflection came out an autonomous agent with dynamic memory and self-reflection um so between cognitive workflows and autonomy and the investment coming up in into these models we are far closer to fully autonomous agents than I think many people recognize so the reflection stuff I'm not going to do a full video on reflection there's there's other um ones out there but basically this outperforms humans in in a few tasks and it forms a very very basic kind of cognitive architecture Loop so query action environment reward reflect and then repeat so you just continuously iterate on something in a loop and there you go uh and also for people who keep asking me what I think about um uh what's his name Ben gertzel I'm not sure if I'm saying his name right but I read his seminal paper a couple years ago on general theory on general intelligence and he never mentioned iteration or Loops at least not to the degree that you need to when you're talking about actual intelligence so I personally don't think that he's done anything particularly relevant today I'm not going to comment on his older work because obviously like he's made a name for himself so on and so forth but I don't think that Ben has done anything really pertinent to cognitive architecture which is the direction that things are going um but yeah so when when MIT is doing research on cognitive architecture and autonomous designs when Morgan Stanley and Nvidia are working on investing literally hundreds of millions of dollars to drive down inference cost and when open source uh libraries are creating um the rudiments of cognitive architectures we are ramping up fast and so someone asked what I meant again kind of getting back to that what did I mean by AGI within 18 months I said in 18 months any possible definition of AGI that you have will be satisfied um so it's like I don't care what your definition of AGI is unless like there's still some people out there that like you ask them and it's like oh well once AGI hits like the skies will darken and nuclear weapons will rain down and I'm like that's not AGI that's Ultron that's different that's that's a fantasy um that's probably not going to happen it could if skynet's going to happen it will happen within 18 months um but I don't think it's going to happen Okay so that's section one of the video talking about the news and everything out there so now let me pivot and talk about the work that I've been doing um so I've been making extensive use of chat gbt4 to accelerate my own research um I've been working on a few things many of you are going to be familiar with my work on the heuristic imperatives which is how do you create a fully autonomous machine that is safe and stable ideally for all of eternity um so this is this is is this is probably one of my most important pieces of work and I've put it into all of my books and a lot of other stuff the tldr of heuristic imperatives is it's like it's similar to asimov's three laws of robotics but it is much much more broadly Genera generalized and it is also not um androcentric or anthropocentric and so basically the three rules that if you embed them into your your autonomous AI systems uh reduce suffering in the universe increase prosperity in the universe and increase understanding in the universe this creates a very thoughtful machine and it serves as a really good um reinforcement learning mechanism self-evaluation mechanism that results in a very thoughtful uh machine so that information is all available out here um under uh on my GitHub Dave shop here is to comparatives I've got it published as a word doc and a PDF so I started adopting a more scientific approach um because well there's a reason that the scientific paper format works so if you want to come out here and read it um it's out there it's totally free of course um oh actually that reminds me I need to put a way to cite my work because you can cite GitHub repos but basically this provides uh quite a quite a bit and one thing it to point out is that this paper was almost written entirely word for word by chat gpt4 meaning that all of the reasoning that it does was performed by chat gpt4 and at the very end um I actually had it reflect on its own performance um it looks like it's not going to load that much uh more pages oh there we go examples so anyways uh when you read this and you keep in mind that the the Nuance of it whoops that the Nuance of this was uh within within the capacity of chat gpd4 you will see that these models are already capable of very very nuanced empathetic and moral reasoning and this is one thing that a lot of people complain about they're like oh well it doesn't truly understand anything I always say that humans don't truly understand anything so that's a frivolous argument but um that leads to another area of research which I'll get into in a minute uh but basically keep in mind how nuanced this paper is and keep in mind that chat GPT wrote pretty much the entire thing and I've also got the transcript of the conversation at the end so if you want to if you want to read the whole transcript please feel free to read the whole transcript and you can see um where like we worked through the the whole paper um yeah so that's it so on the topic of uh does the machine truly understand anything that resulted in this transcript which I have yet to format this into a full um uh scientific paper but basically the the tldr here is that I call it the epistemic pragmatic orthogonality which is that the epistemic truth of whether or not a machine truly understands anything is orthogonal or uncorrelated with how useful it is or objectively um correct it is right so if you look basically it doesn't matter if the machine truly understands anything because again that's not really germane to its function as a machine and so this is uh it's a fancy term but it basically says okay and there's there was actually a great Reddit post where it's like can we stop arguing over whether or not it's sentient or conscious or understands anything that doesn't matter um what matters is it's its physical objective measurable impact and it whether it is objectively or measurably correct or useful so I call that the epistemic pragmatic orthogonality principle of artificial intelligence I've got it summarized here so you can just read this is the executive summary um that I actually use Chad gbt to write so again a lot of the work that I'm doing is anchored by chat GPT and the fact that chat GPT was able to have a very nuanced conversation about its own understanding kind of tells you how smart these machines are um yep so that is that paper now moving on back to uh some of the cognitive architecture stuff um one thing that I'm working on is called Remo so the rolling episodic memory organizer for autonomous AI systems I initially called this hmcs which is hierarchical memory consolidation system but that's a mouthful and it doesn't abide by the uh the current Trend where you use an acronym that's easy to say right so Remo rolling episodic memory organizer much easier to say much easier to remember basically what this does is uh it's also not done so I need to add a caveat there I'm working through it here with chat gpt4 where we're working on defining the problem writing the code so on and so forth but basically what this does is rather than just using semantic search because uh a lot of folks have realized that yes semantic search is really great because it allows you to search based on semantic similarity rather than just keywords super powerful super fast uh using stuff like Pinecone still not good enough because it is not organized in the same way that a human memory is so Remo the entire point of Remo is to do two things um the two primary goals is to maintain salience and coherence so Salient memories means that uh what you're looking at is actually Germaine actually relevant to the conversation that you're having which can be more difficult if you just use semantic search the other thing is coherence which is keeping the context of those memories um basically in a coherent narrative so if rather than just focusing on semantic search the two terms that I'm introducing are salience and coherence and of course this is rooted in temporal binding so human memories are temporal and associative so those four Concepts salience and coherence are achieved with temporal and associative or semantic consolidation and so what I mean by uh temporal consolidation is you take clusters of memories that are temporally bounded or temporally nearby and you summarize those so that gives you that gives you temporal consolidation which allows you to take you can compress memories AI memories you know on a factor of five to one uh 10 to 1 20 to 1 depending on how concisely you summarize them so that gives you a lot of consolidation then you use a semantic modeling to create a semantic web or a cluster uh um from the semantic embeddings of those summaries so it's a layered process actually here I think I can just show you here um wait no I've got the paper here let me show you the Remo paper um so this is a work in progress it'll be published soon um but let me show you the diagrams because this will just make it make much more sense oh and chat GPT can make diagrams too you just ask it to Output a mermaid diagram definition and it'll do it so here's here's the tldr the very simple version of the Remo framework it's it's got three layers so there's the raw log layer which is just the chat logs back and forth the temporal consolidation layer which as I just mentioned allows you to compress memories based on temporal grouping and then finally the semantic consolidation layer which allows you to create and extract topics based on semantic similarity so by by having these two these two layers that have different kinds of consolidation you end up with what I call temporally invariant recall so the topics that we that we extract um are going to include all the time uh from beginning to end that is relevant while also having benefited from temporal consolidation I'm going to come up with some better diagrams to to demonstrate this but basically it's like actually I can't think of of a good way to describe it um but anyway so this paper is coming um and I'm I'm actively experimenting with this on a newer version of Raven that uses a lot more implied cognition so I talked about implied cognition in a previous episode but basically implied cognition is when I in using chat gpt4 I realize that it is able to think through stuff without you having to design a more sophisticated cognitive architecture so the cognitive architecture with gpt4 as the cognitive engine actually becomes much simpler and you only have to focus your I don't want to say only but the focus shifts then to memory because once you have the correct Memories the the model becomes much more intelligent so that's up here under Remo framework I'm working on a conversation with Raven to to demonstrate this um and and that's that the paper will be coming too so that this is one big important piece of work the other most important piece of work that I'm working on is the atom framework which this paper is already done um but atom framework let me just load it here there we go so um autonomous task orchestration manager so this is another kind of long-term memory for autonomous AI systems that's basically like the tldr is um it's like jira or Trello but for machines with an API um and so in this case uh you it's inspired by a lot of things one agile two on task by David Bader um Neuroscience for dummies uh jira Trello a whole bunch of other stuff um but basically we talk about cognitive control so I'm introducing a lot of Neuroscience terms to the AI community so cognitive control has to do with task selection task switching task decomposition goal tracking goal States those sorts of things um and then we talk about um you know some of the inspiration agile jira Trello um and then so it's like okay so what are the things that we need to talk or that we need to include in order for an AI system to be fully autonomous and and track tasks over time so you need tools and Tool definitions you need resource management and you need an agent model all these are are full are described later on or in Greater depth um then actually in my conversation with um with chat GPT one of the things that it said is like okay well how do you prioritize stuff and I was like I'm glad you asked um and so I shared my work with the heuristic imperatives and chat GPT agreed like oh yeah this is a really great framework for prioritizing tasks and on and measuring success okay great let's use that um I'll I think let's see is the uh transcript posted I don't know if I posted the transcript I didn't I'll post the full strength transcript of of right making the atom framework um in the repo um so then we get into like okay so now that you have all the background what do we talk about so it's all about tasks and the data that goes into the task so first you need to figure out how to represent a task so there's basic stuff like task ID description type goal State priority dependencies resource time estimates task status assigned agents progress and then the one that is um new is Task impetus so this is something that you might not think of if you think about you know your your jira board or your kanban board or Trello board is the why so the why is implicit in our tasks why am I trying to do this but when we added this um Chad GPT got really excited and it's like oh yeah it's actually really important to record why any autonomous entity is doing a task for a number of reasons one to track priorities or the the the the impetus might be superseded later on any number of things but also you need to justify the use of those resources in that time so this all goes into the representation of a task which you can do in Json yaml flat files Vector databases whatever I don't care like you can figure out how you want to represent it I'm probably just going to do these in text files honestly because that's the easiest thing for an llm to read um and then so talking about the task representation then we move on to the task life cycle task creation decomposition prioritization execution monitoring and updating and then finally completing the task um and then and then you archive it and you save it for later so that you can refer back to it again this is still primarily a long-term memory system for autonomous AI systems um some of the folks that I work with on Discord um and by work with I mean just like I'm in you know the AI communities with them um they all think that the atom framework is pretty cool um so then we talk about task Corpus management which is like okay looking at an individual task is fine but how do you look at your entire body of tasks because in autonomous AI it might have five tasks it might have five thousand tasks and then you see you need some processes to like okay if we're going through these tasks how do we manage a huge volume of tasks and so some ideas about how to do that are here um and then finally uh one of the last sections is some implementation guidelines which is just okay this is this is probably some things that you want to think about when you deploy uh your implementation of the atom framework um yeah so I think that's about it I I'm obviously I'm always working on a few different things but the atom framework and the Remo framework are are the two biggest things that I'm working on in terms of um in terms of autonomous Ai and so yeah all this stuff is coming fast uh I think that's about it so thanks for watching um like And subscribe and support me on patreon if you'd like um for anyone who does uh jump in on patreon I'm happy to answer some questions for you even jump on video calls if you jump in at the high enough tier um I help all kinds of people I do have a few ndas that I have to honor um but those are those are those are pretty narrow and some of them are also expiring um so I've had people ask you know for for help just with writing prompts for chat GPT I've had people ask um simple things like how did you learn what you learned um all kinds of stuff uh but yeah so that's that thanks for watching and cheers everybody

Channel: David Shapiro

Views: 226,698

Keywords: ai, artificial intelligence, python, agi, gpt3, gpt 3, gpt-3, artificial cognition, psychology, philosophy, neuroscience, cognitive neuroscience, futurism, humanity, ethics, alignment, control problem

Id: YXQ6OKSvzfc

Length: 24min 1sec (1441 seconds)

Published: Tue Mar 28 2023