The New Web Paradigm: Verses vs. OpenAI

Summary

Highlights

• The future of AI hinges on developing systems that can reason like humans, moving beyond simple pattern matching. 🧠
• Verses AI is developing agents that operate using a principle called active inference, aiming for more natural learning methods. 🌍
• The challenge for AI today is to understand the world in terms of cause and effect rather than just processing data patterns. 🔄
• Current AI models fail in logical reasoning tests, showing gaps in their ability to function autonomously in real-world scenarios. 🚧
• Gabriel René argues that AI should be able to adapt, learn, and make decisions based on causal understanding, not just brute force data analysis. ⚙️

Key Takeaways

• AI needs to shift from pattern matching to genuine reasoning to achieve true autonomy. 🤖
• Verses proposes a new model, active inference, that mimics biological intelligence to understand causal relationships. 🌿
• The current AI systems, like OpenAI's, face significant challenges in logical reasoning, impacting their trustworthiness. ⚠️
• Future AI will likely be more efficient, requiring less data and compute power, possibly revolutionizing the field. 🚀
• The talk hints at AI's broader implications, including trust, decision-making, and societal ethics. 🌐

Overview

Chapters

• 00:00 - 01:00: Introduction and Setting the Stage The chapter 'Introduction and Setting the Stage' delves into the complexity of our modern world, characterized by significant technological advancements such as AI and data centers. It explores overarching themes and questions about the relationship between machines and humans, the contrast between knowledge and insight, as well as the balance between determinism and creativity. While these issues may seem daunting initially, the chapter encourages readers to engage deeply with these questions to better understand their implications.
• 01:00 - 02:00: The Opportunity for Creativity and Deciding the Future The chapter discusses how current times present a unique opportunity to shape the future, specifically in what is referred to as 'The New American Century.' It emphasizes the potential for human creativity to be central to insight and abundance. The decisions made at personal, societal, and cultural levels will influence outcomes, with the possibility of either fostering unprecedented human creativity or yielding power to existing authorities.
• 02:00 - 03:00: Introducing Gabriel Rene and Verses The chapter introduces Gabriel Rene, the founder and CEO of Verses, a forward-thinking cognitive computing company. Verses is working on developing an operating system named KM, which stands as the first AI operating system (AIOS) aimed at advancing AI development and deployment. The chapter emphasizes the importance of creativity in shaping the future and positions this new operating system at the forefront of technological innovation for what is termed the 'New American Century.'
• 03:00 - 04:00: Disrupting AI with Autonomous Agents This chapter discusses the work of Gabriel, a longtime technologist with extensive experience in AI, IoT, XR, Web3, and spatial technologies. Gabriel is also the author of 'The Spatial Web,' a bestselling book that explores how Web 3.0 is connecting humans, machines, and AI to transform the world. Currently, Gabriel is focused on developing standards for ethical interoperability between AI and other technologies.
• 04:00 - 05:00: The Limitations of Current AI Models The chapter begins with the introduction of Gabriel Renee, who is joining the discussion via Zoom from Los Angeles. Gabriel is welcomed by the audience with applause. He introduces himself as the CEO of Versus AI and the Executive Director of the Spatial Web Foundation. The chapter aims to delve into the current limitations faced by AI models, with a focus on distributed computing technologies. Gabriel Renee will provide insights into the work being done at Versus AI in relation to these topics.
• 05:00 - 06:00: The Importance of Trustworthy Autonomous Agents In the chapter titled "The Importance of Trustworthy Autonomous Agents," the discussion focuses on the transformative impact of autonomous agents within the realm of artificial intelligence. It highlights the recent surge in attention surrounding AI technologies, with industry leaders like Jeff Bezos emphasizing AI as a pivotal technological advancement poised to disrupt various industries. The narrative sets the stage for understanding how foundational technologies, such as trustworthy autonomous agents, are shaping the future.
• 06:00 - 07:00: Limitations of Large Language Models The chapter discusses the limitations of large language models in artificial intelligence. It begins by addressing the concept of AI as a disruptive force, promising a new layer of innovation. The speaker intends to present work being done by their organization, versus, and others in the field to illustrate how current AI practices are being disrupted. A central theme is how the industry is progressing with AI and addressing its limitations.
• 07:00 - 09:00: The Future of Reasoning in AI The chapter titled "The Future of Reasoning in AI" delves into the concept of autonomous agents. It defines autonomous agents as software-based agents capable of making their own decisions and taking actions in an environment without human control. The chapter challenges the common practice of viewing these through the lens of artificial intelligence, suggesting instead that we should focus on the idea of autonomy. This autonomy represents a new kind of intelligence.
• 09:00 - 10:00: Verses' Approach to AI and Active Inference The chapter discusses the implementation of AI systems capable of making independent decisions and taking actions, known as embodiment in software. The need for such systems arises from human limitations in cognitive abilities when dealing with large-scale and complex problems. Historically, humans have invented tools to overcome biological boundaries, and AI represents the next step in this evolutionary tool-making process.
• 10:00 - 15:00: Compositional Scaling and Query Engines The chapter discusses the progression and capabilities in technology that extend human abilities, such as reaching space or detailed exploration through telescopes and microscopes. It highlights not only physical innovations but also cognitive advancements, emphasizing that today's AI and cognitive systems are pivotal for tackling challenges and solving problems beyond human reach.
• 15:00 - 16:00: Beyond Traditional AI to Natural Intelligence The chapter discusses the need for advanced AI agents that can manage the exponential complexity of the modern world. It emphasizes the necessity for these agents to go beyond traditional human intelligence, performing with the speed, scale, and power of computing systems while understanding real-world complexity, uncertainty, and dynamics. The importance of trustworthiness, accuracy, and adaptability in autonomous AI agents is highlighted.
• 16:00 - 20:00: The Promise of Natural Learning Agents The chapter titled 'The Promise of Natural Learning Agents' discusses the evolving requirements for artificial intelligence systems. It highlights the need for these systems to adapt to changing circumstances, interpret ambiguous, fuzzy, or uncertain information, and maintain explainability to gain trust. Moreover, to function autonomously and collaborate at large scales, these agents will be integrated into various IoT devices, robots, cars, drones, toys, and electrical systems, affecting different aspects of daily life.
• 20:00 - 23:00: Verses' Vision for the Future of AI The chapter emphasizes the importance of collaboration and knowledge sharing within AI development, highlighting the need to address exponential challenges and opportunities through innovative solutions. Companies are rapidly adapting to these changes, with recent advancements like SAA's co-pilot and Bill Gates' statement on AI agents as transformative forces in human-computer interactions. The discourse suggests a significant shift in technology dynamics, positioning AI agents as crucial tools in an AI-driven future.
• 23:00 - 29:00: Greg Meredith's Perspective and Response In this chapter, Greg Meredith discusses a transformative shift in the software industry, emphasizing the evolution of autonomous problem-solving agents. Instead of relying on static solutions or applications, these agents are envisioned to adapt and develop responses in real-time, akin to human problem-solving capabilities. This shift is expected to revolutionize computing just as significantly as the transition from command-line interfaces to graphical user interfaces. Mark Benioff from Salesforce underscores the need for the company to pivot towards these intelligent agents, similarly to how it embraced data and cloud technologies in the past.
• 29:00 - 30:00: Conclusion and Next Steps Yan Laon, a leading AI scientist and chief scientist at Meta, discusses the limitations of LMS (Language Models). He argues that LMS are not suitable as foundational agents for achieving human-level intelligence. Laon identifies four essential capabilities for human intelligence that LMS currently lack, suggesting that these limitations must be addressed for further advancement in AI.

The New Web Paradigm: Verses vs. OpenAI Transcription

• 00:00 - 00:30 [Music] as this year's cosm uh reflects we live in a very complicated world uh one where technology such as Ai and data centers um web 30o and overarching questions surrounding super intelligence cause US caus in US deep questions about machines versus Humans knowledge versus Insight determinism versus the creative pursuit of the deepest questions this sounds like a a daunting set of questionss at first but at the
• 00:30 - 01:00 same time it affords us a unique opportunity to shape the direction of what George has called The New American Century when we're human creativity is at the center of insight and abundance make no mistake there are a myriad of decisions we can and will make personally as a society and as a culture that will determine the outcomes here and the guard rails we decide on will either unleash an explosion of human creativity as we've never seen before or we will abdicate to the powers that be
• 01:00 - 01:30 who want to make those decisions for us I for one welcome the opportunity to shape the this this conversation and to put creativity at the core of the New American Century to help us frame this I'm going to introduce our main speaker Gabriel Renee Gabrielle is the founder and CEO of versus a Next Generation cognitive Computing company developing an operating system called km ironically km uh the world's first aios for the develop Vel M and deployment of AI
• 01:30 - 02:00 powered applications Gabriel is a longtime technologist and has been at the intersection of AI and iot XR and web3 and spatial Technologies for for many decades he's the author of the number one international bestseller the spatial web how web 3.0 connects humans machines and AI to transform the world he specifically focused these days on developing standards the eth ethical interoperability between Ai and other
• 02:00 - 02:30 distributed computing Technologies so with that I'm going to ask Gabriel to join us he is joining us via Zoom uh I believe from LA and there's Gabriel Renee so please join me in welcoming [Applause] Gabriel hello everybody great to be here thank you for this opportunity to share a little bit about the work we're doing in versus my name is uh Gabriel Renee I'm the CEO of versus AI I'm also the executive director of the spatial web
• 02:30 - 03:00 foundation and today I want to paint for you a picture about the future of Agents over the last two years we've seen this explosion is Around The Narrative of artificial intelligence um you know Jeff Bezos describing this as uh the key technology that will in fact be disrupting every industry I think that every once in a while you have a foundational technology
• 03:00 - 03:30 that's not just another mouse trap that acts as a new disruption layer for the next thousand disruptions and artificial intelligence has certainly promised to be that but today I want to share with you how the work that we're doing at versus uh along with a handful of our other colleagues is actually disrupting AI where is everyone going with AI the main theme has emerged
• 03:30 - 04:00 around autonomous agents autonomous agents can be thought of as um the ability for a a a software-based agent to be able to make its own decisions and take actions in an environment without any human control in fact often we use the term AI we we tend to think of it through the lens of artificial intelligence I'm going to talk about why that's a bit problematic today but the word that we really should be thinking about is autonomy autonomous in intelligence of a whole new kind and how
• 04:00 - 04:30 this is embodied in software with software systems that can take their own actions and make their own decisions why do we want this because humans have effectively reached a sort of limit on our cognitive abilities the size and scale of the challenges that we face today at planning scale and Beyond are difficult for us and in the past whenever we reach the boundaries of our of our biological limitations we invent a new tool whether that's a spear to be
• 04:30 - 05:00 able to grab some food just beyond our reach or or rockets that can move us to other planets um telescopes and microscopes that let us see deeper or further but the cognitive function not just memory not just libraries and books and the web or even I would argue today's AI but the ability to have systems that can think and figure out challenges beyond what we can to solve problems that we can't uh is the key
• 05:00 - 05:30 right we need we need something to help us manage the sort of exponential complexity of the modern world and this is what we want agents to go beyond human intelligence so we want agents to be able to perform at the speed and scale and power of computing systems but be able to understand the complexity and the uncertainty and the Dynamics of real world real-time situations it's critically important if they're going to be autonomous that they're trustworthy right this means that they must be accurate and they must be able to adapt
• 05:30 - 06:00 to circumstances as they change they're going to need to be able to make sense of information when it's not when there aren't a million examples when it's ambiguous when it's fuzzy when it's uncertain and still be explainable enough to be trustworthy for us to grant their autonomy and they're going to need to be able to collaborate at scale agents are not just employees they're going to be embedded into iot devices and robots and cars and drones and toys um in our electrical you know dynamic electrical
• 06:00 - 06:30 grids in our supply chain so they have people to collaborate share knowledge skills insights and do this at the speed and scale that can solve some of these exponential challenges and opportunities we face and everyone is uh pivoting their their companies around this even just the last few days SAA coming out their latest version of co-pilot saying agents are the new apps for an AI powered World Bill Gat saying agents are not only going to change how everyone interacts with computers they're also upend the
• 06:30 - 07:00 software industry essentially it'll be autonomously developing Solutions in real time much like we rely on humans not to just have a package solution or application but to figure things out as you go um this is going to up in the software industry bringing about the biggest revolution in Computing since we went from typing commands to tapping on icons hinting at sort of natural ways of interacting with these agents and Mark bet out from Salesforce is that we have to Pivot the whole company agents just as we did data and the cloud for
• 07:00 - 07:30 that Yan laon is um widely considered one of the the top scientists in the world one of the early uh developers of artificial intelligence technology and he is the chief scientist at meta and he is pointing out that LMS have some fundamental limitations that do not do not set it up well to be the kind of foundation that we would want for agents and he he talks about these as uh the four capabilities necessary for human level intelligence that LMS do not
• 07:30 - 08:00 exhibit essentially the ability to have a memory the ability to uh understand the physical world ability to reasoning and planning and this term reasoning has become the sort of key essential uh phrase and the capability set that all of the companies are saying is required we you need an agent to be able to figure out what to do if they're going to autonomously act and do it effectively and we just saw open AI do their latest raise at 160 billion plus and laid out their vision for where the organization was going the kind of key
• 08:00 - 08:30 uh piece that you need to get to agents all the way at the end there is reasoning and this is the vision that they've laid out level one chatbots level two reasoners human level problem solving in order to get to agents and then beyond why because again if you're going to have agency if you're going to have autonomy you're going to execute uh decisions and actions by yourself we have to be sure that you're making sense of the world so you have to be able to
• 08:30 - 09:00 reason um when they rolled out their latest version of Chad GPT called 01 uh this is just a a couple months ago um despite billions of dollars being spent around attempting to achieve higher levels of accuracy and and hoping for common sense uh reasoning in these systems we see these same sort of hallucinations and errors that persist and today I'm going to explain to you why those exist and why these are an unsolvable problem uh for the kinds of critical real world activities that we
• 09:00 - 09:30 want agents to do and so you know I I saw Sam post this he said here's 01 a series of our most capable and aligned models yet um 01 is still flawed still limited and still seems more impressive on first use than it does after you spend more time with it and I couldn't help but try to imagine Steve Jobs or any CEO of of any technology company in the last 35 years walking out and introducing the lat latest project by saying who the iPhone it's still flawed
• 09:30 - 10:00 still limited and still seems more impressive on first use than it does after you spend more time with it perhaps we could think of this as refreshingly honest but the reality is that these systems are deeply flawed in ways that we can't trust um they are what we call you know unreliably unreliable it's because they throw errors that no common sense person that understands how the world Works would would would would come up with and so the uh the story of the last I think the
• 10:00 - 10:30 sort of buried lead of the moment is that Apple was going to invest in in open the eye and we found out on September 27th that they bailed out on October 2nd opening ey closed their multi-billion dollar round and on this October 7th um this month just about three weeks ago five days after opening eye closed their round they released this paper and I want you to think of this as the emperor wears no clothes moment and I believe this is indicative of why Apple decided not to invest
• 10:30 - 11:00 either in open the ey or really in a future based on llms so this paper is is created uh all researchers from from Apple and essentially is called the gsmk test that's grade school math and there's about 8,000 questions what they're trying to figure out is can large language models do reasoning you can see here at the bottom that current llms are not capable of genuine logical reasoning they're doing pattern matching but they're not doing reason
• 11:00 - 11:30 um I think this is the moment where 01 which was supposed to be considered this reasoners which is supposed to get us to agents which was code Nam strawberry for a long time I like to think of this as the Apple has just squashed the strawberry um and I'm going to walk you through just a couple of things there's a few there's a wired article that came out about a week ago that does a great job of of sort of exposing this and walking through it very readable and there's the paper itself but here's just a couple simple things and I want you to stop thinking about AI through the lens of um of some sort of super intelligent
• 11:30 - 12:00 Al alien intelligence I want you to be very uh uh Discerning in how you think about how this might apply to your business to your data to any sort of real world scenario so here's the setup there are uh hundreds of these examples uh that came out of this research but here's the one I that I like the most remember grade school level math so Oliver picks 44 kiwis on Friday then he picks 58 kiwis on Saturday on Sunday he picks double the number of kiwis he did on Friday now this is all part of the
• 12:00 - 12:30 normal GSM AK test but then the Apple researchers decided to add one little bit of additional information and they they did this over and over in many different ways merely changing out the names or changing out the numbers or the kinds of things that were to be counted and it would just throw these catastrophic errors in this case certain number of of kiwis Friday certain number on Saturday an obscure number on Sunday so can it reason and do some math but the key that they added is but five of them were a bit smaller than average how many key is Oliver have 01 and llama the
• 12:30 - 13:00 the two most popular um and latest closed source and open source models in pink here see you can see double the number you picked on Friday which is 2 * 44 88 kiwis however on Sunday five of these kiwis were smaller than average this is the model talking to itself so we need to subtract them from the total so 88 minus 5 equals 83 and then you add that all together you get 185 qes that's wrong and llama did the exact same same
• 13:00 - 13:30 thing and we see this this fail uh in this in a number of these examples where the score drops from let's say the high 90s in terms of accuracy to 80s 70s and all the way down to 65% drop I'm talking accurate to the 30s in some of the state-of-the-art largest most successful business models and the future of all software why is it making such a bizarre error well it turns out reasoning
• 13:30 - 14:00 is simply the function of understanding what cause and effect is and systems that are doing correlational pattern matching might see millions and millions of correlations but those correlations do not necessarily mean that the model has understood cause and effect this is a very costly and dangerous way to get to causal reasoning and causal understanding in this case not understanding that the size of something the amount the size of ke shouldn't affect the number of them and there are an endless number of these just you a couple more things from the researchers
• 14:00 - 14:30 themselves before we move on to the next step this is merad he's the team lead over at Apple and uh this he posted this on on Twitter just a few weeks ago overall we found no evidence of formal reasoning in language models including all the top models um they're better uh their behavior is better explained by sophisticated pattern matching um so fragile in fact that merely changing the names from Jennifer or oler you know uh
• 14:30 - 15:00 resulted in 10% uh you know increase in fail failure rates we can scale data we can scale parameters the size of the model we can scale compute or use better training data for these including gp5 the next big model but we believe this will result only in better pattern matchers not necessarily better reasoners this is a pretty big indictment and it it is a a proof that these systems are incapable of doing actual reasoning why because reasoning to be accurate and effective is nearly the ability to understand cause and
• 15:00 - 15:30 effect if you do it historically sometimes we call that reasoning if you do it forward in time we call that planning and this is what you need for agents so can scaling data models or compute fundamentally solve this we do not think so um and even though some of the later models perform a little bit better understanding lm's true reasoning capabilities is crucial for them to be able to be deployed in any real world scenario where accuracy and consistency are non-negotiable AI safety alignment education Healthcare obviously autonomous vehicles of any kind our findings emphasize the need for more
• 15:30 - 16:00 more robust uh evaluation methods but essentially the headline here is developing models that move Beyond pattern recognition to True logical reasoning is the next big challenge for AI but it versus this is a challenge that we took on a couple of years ago and it's one that we have already been able to achieve some significant results with I'm going to explain to you today the difference if the the story and the plot is Agents if Microsoft is using open AI for their agents and Salesforce
• 16:00 - 16:30 is using llama for theirs and so on and so forth trying to get reasoners first this is a fail they do not understand cause and effect why is this it's because they're doing what you can think of as artificial learning right biological intelligence human intelligence we learn a very specific way and artificial intelligence doesn't do this at all so what they do is they try to learn correlational based pattern matching right off of large data sets that have repeated examples of the
• 16:30 - 17:00 same data so you say million pictures of cats then now I know what a cat is right and so but what are they trying to do they're trying to model causal relationships that's the whole point of downloading all the information on the web and hoping to get a model that can churn through that but if all you have as a starting point is correlational data then you're doing associative learning not understanding how small things cause effects in other things the sort of top down Brute Force approach uh is the kind of fly in the ointment right this this big data wave is the opposite
• 17:00 - 17:30 of of how biological intelligence so it's correlative not causal um and as a result it leads to these kinds of unfixable gaps and inaccuracies in the model um and of course we've used all the data we' Gathering all the chips we're spinning up nuclear plants and talking about hundred billion dollar data centers and this scale is all un need approach is now achieving you know diminishing returns for exponential costs so there's an AR exural flaw in this how might we understand how the
• 17:30 - 18:00 brain works well enough to be able to make one in software well at vers is we start by looking at the brain itself our chief scientist is Professor Carl friston the world-renowned neuroscientist who invented statistical parametric mapping a lot of the under underlying foundational brain Imaging Technologies it's one of the most cited um scientists in the world not not even just in neuroscience and Carl friston H has developed a underlying mathematical model much like Newton developed um for
• 18:00 - 18:30 physics and and and and and Einstein for relativity but based on how neurons learn in the brain and essentially how all biological intelligence is able to understand how to model information efficiently it turns out given the amount of energy you have is a huge function for how well you can think and learn and adapt the essential ingredients for reasoning turn out to be a function of physics and this is called active inference so active inference is a mathematical framework for describing
• 18:30 - 19:00 how systems learn cause and effect relationships in order to build an accurate model of the world or environment they need to operate in which allows them to build a kind of model that environment upon which they can then run hypotheticals do predictions reason about past events reason about current states and even reason about the future what's amazing about active inference though as opposed to the kinds of inference we're seeing in today's systems which are largely passive kind of still pinging a very
• 19:00 - 19:30 large model that's a black box that's already been trained but cannot continue to learn active inference if I if I'm trying to learn cause and effect I actually need to be able to see what effects I cause let's go back to agents for a second that is what an agent must do an agent is effectively a kind of cause that you want to produce a kind of effect the desired effect not another effect these types of errors that we're seeing in large language models are are are causing uh a a massive uh whole uh
• 19:30 - 20:00 sort of fragmented glass if you will within that entire model where you never know where the air is going to come up because it just doesn't understand some basic cause and effect things active inference agents uh are able to model causal relationships as a result they can generalize unknown scenarios and they do not require massive amounts of data or compute to do that how do we know this because the underlying science has already been validated about uh 2023 in August uh professor frisen and some other colleagues of ours at University of
• 20:00 - 20:30 Tokyo demonstrated how active inference um uh was the self-organizing principle for how biological neural networks work in the brain this is tested with real neurons live neurons and Petri dishes where it could predict how learning which neurons would update in what order within a uh within within a a petri dish um given a certain set of sensory inputs so this is a way of demonstrating that the underlying mathematical model is able to represent the way that neurons work and what we've done at versus is
• 20:30 - 21:00 take that underlying framework and apply that to software what you get is a new class of agents that are capable of learning faster using fraction of the amount of data and compute um and we are at an inflection point as an organization where in the next couple of months expect by the end of the year internally we're going to be able to demonstrate a sufficient uh uh Superior reasoning planning and adaptive learning in What's
• 21:00 - 21:30 called the Atari test uh where we will be challenging one of the largest AI models in the world known as Alpha zero uh and even the hyper efficient version of that um the approach that we're taking is with an intent of enabling a new CL kind of agent that doesn't use language comprehension as its basis for understanding the world language is a really powerful communication tool but it's not how your brain encodes information you're encoding smell and sight and and taste and and shapes and
• 21:30 - 22:00 geometry SpaceTime interactions that are nonlinear it's not just about predicting the next token if you're going to survive in the world you have to be able to model the world and operate in it so we've been developing a full stack solution which we call genius to enable better cheaper faster kinds of agents that can model the physical and social and emotional aspects that are required for real world agents not just in software in accounting system so for chat Bots but to be able to operate in cars and drones and robots and in the
• 22:00 - 22:30 real world and even be able to work together in ways that are also more efficient wildly more efficient and explainable why because they're able to model cause and effect you're not just trying to pick out data out of a trillion you know parameters in a massive model so this alternative to artificial learning is natural learning and this is how you've learn this is how your children learn this is what we see you know billions of years of evolution has shown that the way that systems learn in the real world is they learn from the body bottom up not from the top down and it's all about learning cause
• 22:30 - 23:00 and effect so we we we we understand this from Neuroscience we understand this from physics but the ability for agents to be able to understand how small things uh how you learn little pieces those building blocks make up the next set of building blocks shapes become letters letters become words words become sentences and so on and so forth this allows you to model not just the the objects themselves but the relationships between them because it's your modeling cause and effect it's a
• 23:00 - 23:30 sparse lightweight model that doesn't require a hundred billion dollar data center in fact could run on each one of a hundred billion different iot devices this this approach allows for a causal understanding that agents themselves can test their understanding of the world that's what cause and effect means not just a record of what causes and effects have occurred before but how can I learn from my my results if I'm right I want to reinforce that if I'm wrong I need to be able to correct it this is a a fundamentally different approach to
• 23:30 - 24:00 developing AI agents that give us a baseline for knowing whether or not we can trust them to be autonomous enough to make decisions and take actions on our behalf so to kind of Connect the Dots here the underlying architecture for large language models chatbots generative AI is uh is are called artificial neural Nets these were originally developed in the 1950 50s um and we've basically been using the same
• 24:00 - 24:30 architecture applied to the latest era of what's called Deep learning it's just add more compute add more data uh add more energy and give it more Cycles now I want you to think about how you would want an agent to solve a problem for you in your business or in your life if if that problem is a kind of puzzle and you're not just giving it instructions or directions because what's the point of giving instructions directions to autonomous car I want it
• 24:30 - 25:00 to figure it out and drive me to the location in the world of artificial learning what a root Force approach to solving a puzzle is is taking every puzzle piece learning the shape And then trying every puzzle piece on every other puzzle piece now obviously you can have a system that never knows what's on the puzzle that can Brute Force its way to solving a puzzle occasionally and I've got a four-year-old so I've seen this it might put a puzzle piece that doesn't really fit what we would see on top of the puzzle but the shape can kind of be
• 25:00 - 25:30 jammed in there there's your five kiwis and there are billions of these inside of the model puzzle pieces that are the right shape but not the right piece on what's happening at the second layer of the puzzle and that's what natural learning does in the underlying what we call renormalized generative model that we use that that uh vers is using for our genius agents it's not just about the shape of each puzzle piece fitting together and we know no human being solves a puzzle by trying to do one
• 25:30 - 26:00 piece at a time and reinforcing it until it works we we actually group puzzle pieces together by color by shape by objects that we see today's AIS don't see anything but shapes nothing is on the puzzle it doesn't understand the real world context it just misses something like size and shape right that or the number of something and the and the size of something as as in the KB examples we have to have uh models that allow agents to be able to understand
• 26:00 - 26:30 what's happening at that secondary level whether that's human social context whether that's understanding physical activities in the real world or understanding space and time effectively these are the critical ingredients and so having an entirely new approach to how you develop agents that starts with the concept of agency and not just the concept of more data is fundamental to being able to do this right and this gives us the ability to essentially have agents that learn reason right they
• 26:30 - 27:00 understand cause and effect relationships and this is what allows them to then reason plan and self-correct because that is their main job active inferencing is trying to infer the causes and effects that lead to reasoning which also allow explainability in these systems and this approach inevitably allows for agents to dynamically adapt to unseen unknown in in environments or situations right and generalize because they're learning the relationships between things and the structure of that as opposed to just
• 27:00 - 27:30 correlations and the number of times something happens now you can get agents that don't just learn what but can learn why to kind of Connect the Dots here and wrap this up for today I want you to think about all the different kinds of AI that we've developed over the period of the last few years and I'm not going to make the case today that active inference-based agents or what we're doing at versus is a whole replacement for all kinds of AI there are kinds of
• 27:30 - 28:00 AI that are very good for for narrow tasks certain kinds of skills things that are highly repetitive you don't need high explainability for it doesn't need to figure much out but you need something that's missing in the entire AI space today so if you imagine that we have pretty good Vision Systems computer vision we have we have a pretty amazing robotic systems that have all kinds of motor functions right and we think of these as parts of the brain we have iot devices that can sense um all types of information and and measure that information in ways
• 28:00 - 28:30 using artificial intelligence whether that's um you know Vision or sensory or uh or or or motor functions um what we've added now is language comprehension large language model is not a world model it's a language model language is an amazing representation of the world but it is insufficient model of the world and any arguments is is a pretty good example of this the thing that's missing if we want to use all those other pieces is what we
• 28:30 - 29:00 have in the prefrontal cortex right so we have this sensory AIS and we have motor AIS and we have language AIS but we're missing is the reasoning and planning functions that come in the prefontal cortex this is often called the executive functions this is the part that does the orchestration the reasoning the planning the learning even emotional regulation social cues all of this happens and it happens that if you're going to do reasoning you have to do that in space and time and if you're going to do planning you have to understand space and time right this is ability to understand those cause and
• 29:00 - 29:30 effect Dynamics be able to run those simulations that's the power of the frontal load right and that's what versus is building with active INF agents it can still use the language model for language it can use a vision model for vision it can use a a motor based model for robotics whether that's a car or a humanoid robot or a toy but the part that we're missing is the part that can do orchestration navigation and figure out how to reason about the world make sense of the world make take decisions and uh take actions and even
• 29:30 - 30:00 adapt and learn as it goes this is the I think the missing ingredient if if we look at the context of all we're essentially building a kind of brain or a kind of human sometimes called artificial general intelligence this is the missing key ingredient and this is what we're delivering you can kind of see this in the market the the bubble feels ever present um I think the Apple paper was the emperor wears no closed moment I think that's going to reverberate through the industry um uh by this time next year the industry will not even
• 30:00 - 30:30 look like it it does now in the same way it didn't look anything like this in be at the end of 2021 when then opening eye came out you know December 2022 so just in a two-year period we'll have seen this already kind of come and gone in terms of its peak and replaced I think by more first principles based approaches that if you think about the concept here uh versus it's just been named as one of the emerging leaders by gner in what's called first principles AI which they described as first principles approach Bridges a gap between p data driven and physics-based modeling enabling more reliable more
• 30:30 - 31:00 efficient and more versatile generalizable approaches to Ai and you see generative AI kind of coming over What's called the peak of inflated expectations on its way to the sad trough of disillusionment and versus is on its way up uh as a I think a valid Contender uh in the field for a whole new approach incidentally when we say something like first principles what what does Gartner mean what do we mean well first principles is understanding cause and effect right going back to understanding how small things make bigger things and how those make bigger
• 31:00 - 31:30 things and how that rolls up then you can track your way back down and really understand the structure of something and this first principles based approach is at the heart of what we're doing not merely a generative approach based off lots of correlational pattern matching and so we have a milestone coming up and you can think of this as kind of the Olympics of AI but back in 96 you know was kind of a milestone moment in the AI you know Gary Kasparov loses to big blue but might be MHS 20 years later this latest wave of Deep learning uh AI beats Lisa doll this was
• 31:30 - 32:00 called Alpha go it was later beaten by a model called Alpha zero and then they said that's the hardest board game there is what about trying to learn and play dynamically not just one game but many different games right and at tari has emerged as sort of state-of-the-art test they took Alpha zero they kind of tuned it up and said okay what if we could get this thing to op to learn to play a handful of games in two hours what we've been able to demonstrate over the last few years is that we can get neurons to play Pong
• 32:00 - 32:30 the very first game in this test from a petri dish using active inference that this is what some Partners called cortical Labs um last year we demonstrated that we could do that purely in software with an a version of pong which we shared that showed that we could beat efficient zero and in the latest paper that we put out with those renormalized generative models that learn the different levels and layers of information um we were able to demonstrate the ability for the model to begin learning to play in 18 seconds now
• 32:30 - 33:00 just to give you some context here what we're planning on doing uh demonstrating internally uh within the next um two months probably by the end of this year we're doing what we call Atari 10K this would be a our approach to beating efficient zero the kind of big bad reinforcement uh deep reinforcement learning model uh there is uh with a tenth of the data so 90% less data uh not in two hours but in roughly 12 minutes um uh not uh using four gpus it's for
• 33:00 - 33:30 NVIDIA a100 but on a laptop and with a model that's 100,000 times smaller this would be kind of that hopefully Time Magazine moment that we saw when Alpha go beat the human like Lisa doll or big blue be Gary COV and so we we we're looking into what's the right way to share this this with the world um but this will be the key sort of agent-based Benchmark why is this important because a atar games are largely irrelevant to the to most of the world what this does
• 33:30 - 34:00 is it shows how you can take an agent from scratch and it can learn reasoning planning and learning and adaptation in a handful of different environments that these games set up and so these are analoges to the real world which is often changing often Dynamic has physics has all kinds of complexity to it that that the model has to learn uh you could think of this as if if if the rest of the industry had been building you know optimizing fuel ection on a 100y old you know combustion engine architecture
• 34:00 - 34:30 versus is the Tesla and This Is Us beating a Porsche uh for the first time with the first world you know electric vehicle in this case it's a natural intelligence-based approach to AI um that we think is you know demonstrate better cheaper faster and safer so to conclude we talked about Bill Gates um and others saying that agents are the key this is the new kind of app but instead
• 34:30 - 35:00 of being an app that software developers give instructions it's something you give goals to and it figures It Out by itself in many ways this is the ultimate vision of both technology and software all of our sci-fi stories have arrived at some version of this Bill Gates talks about this not only changing how we're going to interact computers but how this is going to change how we interact with software but I think if it's clear if anything is
• 35:00 - 35:30 clear it's going to change how we interact with the world it will become the interface to to all of the interactions and be the key mediator ensuring that we can trust these systems to make decisions to be autonomous uh is is only going to be a function of being able to know that they understand cause and effect that it can be modeled in a way that's human interpretable and auditable that it can adapt and learn and self-correct as it goes um and that it's efficient enough to be able to operate not just in one
• 35:30 - 36:00 giant hundred billion dollar or trillion dollar data center but on the laptop I believe that the work we're doing at versus is equivalent to the microprocessor moment for computing where everyone thought back in the late 70s and early 0s that supercomputers you know maybe we'll have three or four of these things and everyone to connect to it in the world it turned out the microprocessor that gave the power of computing to everybody put it in their hands on their desktop in our pockets and now everywhere else has just been the setup for a new kind of software that's going to enable agency at scale and I would agree with the last thing
• 36:00 - 36:30 that that that Mr gate said here is which is that there's a significant chance that this future AI winner um will not be one of the big players in fact will be a startup so thank you so much for the opportunity to speak with you today um it's been a real pleasure to share well thank you Gabriel uh certainly a very provocative uh set of comments here um I have to admit when I came to CM this year I didn't think I'd learned so much about broka area but uh Knowledge from an unexpected Source um
• 36:30 - 37:00 at this point I'm going to actually ask Greg Meredith to come on up to the stage here Greg is uh a longtime multi-decade technologist who has been working on distributed systems and Communications layers and data access patterns uh across many different kinds of architectures and has uh I'd like him to sort of introduce what he's doing on working on at Firefly and then also uh any comments you may have uh with respect to the the comments Gabriel just shared uh well thanks Bob and I I I want
• 37:00 - 37:30 to thank uh George and Steve for inviting me to to engage with this panel and to thank all of you for for attending today um what originally the way we structure the panel was I I would respond to Gabriel and and and so what I want to talk about uh turns out to be a response that I think fits very well so I think uh Gabriel and I wholeheartedly agree but we're coming at it from completely different perspectives um but one thing that we we
• 37:30 - 38:00 absolutely agree on is going back to First principles and and so I want to talk to you a little bit about some first principles that you may have never heard of before um and and so what I'd like you to do I'm sure probably everyone in this room is has heard of the the simulation hypothesis that we're all living in a simulation well I want you to forget the simulation hypothesis and I would like you to enter into the world of by simulation theorems right so so uh so I'm going to
• 38:00 - 38:30 explain a little bit about uh what that means um you you may have uh recalled at the top of Gabriel's talk he talked about um getting to planetary scale uh with with these kinds of intelligence agents A and and I I think we in general humanity is terrible at scaling and why is it terrible at scaling well because the consequences of the way we scale in up being fairly disastrous so I I tend to
• 38:30 - 39:00 ask how does nature scale and one of the ways that that nature scales is through composition um so I'm just going to plant that like a seed uh you know uh from a plant from the natural world into your minds and then I'm going to return to that idea uh toward towards the end of my response to Gabriel um and and so having planted that seed keep that in the back of your mind ready to ready to bloom at some point and now I want to turn to a couple of practical queries these are queries
• 39:00 - 39:30 which which currently are not uh statable even in the ex in existing search and storage mechanisms including AI based search and storage mechanisms so the first one is find me all the small molecules from this uh uh repository say Swiss Pro or one of the other uh uh chemistry or bio um repositories such that when I when I drop it into this cell signaling regime uh in a certain concentration uh the the
• 39:30 - 40:00 the cell signaling regime never reaches a particular State now why is that an interesting uh query to be able to uh to articulate because the English translation is find me a small molecule-based cure let me uh let me uh provide another query oops there we go uh let me provide another query that would be of interest um so find me all the so using this uh
• 40:00 - 40:30 collection of cad-based repositories uh that articulate the structure of um aircraft find me all the wires running from the nose of the aircraft passing through the midsection where aircraft Flex um uh that are likely to that that undergo a Twist so the translation of that query in English is find me likely sources of electrical
• 40:30 - 41:00 shorts now again in today's systems you can't write those queries down it turns out however there are there's there's an algorithm in fact there's a family of algorithms that allow us to generate query languages and query processors that process those kinds of queries they can be generated from the uh domain languages where you naturally find these kinds of questions so in
• 41:00 - 41:30 particular um uh we we the the first one uh can be um generated from the language of chemistry and the second one can be generated from geometric algebra also known as clearford algebra um so the interesting point about the kinds of query languages uh that uh that that the that this algorithm generates is that the queries are perfectly
• 41:30 - 42:00 explainable we can find cause and effect style reasoning that goes into the processing of those queries uh now now you're you're wondering where does this algorithm arise from probably uh so the algorithm arises from a notion from theoretical computer science that you've probably never heard of by its by the name it's given in that community and that name is by simulation but you know it by another slogan and that slogan is
• 42:00 - 42:30 anything you can do I can do right and and if we have time and the other talk ran a bit over so we probably won't have time to get to it but if we have time I'll address what happens if you add better anything you can do I can do better at the end of that phrase now what's interesting about this principle is that it classifies all computational
• 42:30 - 43:00 Behavior that's really important right so the mathematicians were mathematically adjacent in the audience um probably probably remember what a big deal it was us a couple of decades ago when all finite simple groups were classified right that was that was a huge deal because in some sense it meant we understood all the mechanics underlying uh Notions of of of certain kinds of symmetry right so by simulation
• 43:00 - 43:30 does this for all computational Behavior now what does this have to do with AI um so I want to argue that AI is a misnomer and and again very much in agreement with what Gabriel was saying but but coming at it from a different perspective so he's motivated by biological systems and I resonate with that but I also want to understand it in terms terms of of compute right because I believe that right now with the the
• 43:30 - 44:00 the premise underlying the the AI industry is that there's at least a collapse of practical useful aspects of human intelligence that can be represented as computation so even in Gabriel's presentation ultimately he's rendering what he's doing in software right so he's still bought into the idea that when we're doing AI we're reducing at least certain aspects of of
• 44:00 - 44:30 uh and I I would argue human intelligence uh to computation and not just any computation because there's all kinds of computations we can imagine right but we're talking about computation that can be rendered on machines that we can produce and manipulate at scale okay so so the right word is computational intelligence or CI and CI
• 44:30 - 45:00 presupposes that a particular world of intelligent actors for example the kinds of things that uh that Gabriel was talking about as agents um are live inside the world of of of um computations now what B simulation says is everything in that world can be completely classified that means that any dzens of that world
• 45:00 - 45:30 for example an algorithm that can learn can also learn no finer classification of phenomena than by simulation equivalence classes so so if you come up with some algorithm that can classify things in the world it's actually never classifying things in the world it's only ever classifying computational repr presentations and those things can all
• 45:30 - 46:00 the finest uh uh uh classification of those phenomena is is in terms of B simulation um so so these facts are at the heart of this algorithm that I mentioned and and it's it's at the heart of query engines that Firefly is building actually we're building sort of a meta query engine that allows you to generate those those query engines at scale now how do we get to planetary
• 46:00 - 46:30 scale so I I want to argue that the the natural world scales through composition now so what do I mean by that so in in terms of the standard model which we all know is probably wrong but it's the best we've got you know atoms are are are made of um quarks leptons and bons and molecules are made made of atoms and proteins are made of molecules cells are
• 46:30 - 47:00 made of proteins organisms are made of cells populations are made of organisms right so so the the thing to take away from this is is that composition is really the Bedrock of a proposition uh that is at the heart of Western science right so w the view of Western science is taking a reductive uh view of the world is is really backwards what what what estern
• 47:00 - 47:30 science is really saying is that the natural world is built through composition right we compose cells to make organisms and and and and so so the the the sense intelligence is uh a naturally occurring phenomena right right we are all biological organisms in this room um shouldn't we expect that an effective
• 47:30 - 48:00 account of intelligence should be deeply compositional so what b simulation uh does is it classifies all comput uh all computational Behavior because it exposes the Deep compositional structure of composition of I'm sorry of of computation right so so so computation actually has building blocks and bi
• 48:00 - 48:30 simulation exposes those building blocks in terms of this principle Anything You Can Do I can do and that is how we begin the process of scaling so I've tried to keep it very short uh rather than long so so uh so uh I'd like to thank you again for your attention and if there's time to open it up for questions yeah i' to open to questions one of the things that's been fascinating for me though over the course of the last thinking back over
• 48:30 - 49:00 the last few years of cm is you sort of think about the impact of AI almost when we first started having some of these conversations is about kind of the machines and this year is much more about sort of that that recognition that actually the machine is reflecting kind of the the human brain and the and the ways in which the patterns of the human brain understand I suspect that while we've been talking about reasoning I suspect next year's
• 49:00 - 49:30 CMS and the CMS after that will be about reasoning judgment and maybe even morality because the problems that you actually end up with in a in a world where it's just the machine separated from those other elements is that you don't have well-formed choices you get answers not wellform choices and it'll be interesting to see how this sort of conversation evolves over the over the coming years uh with that I'd like to see if there are any questions uh in the
• 49:30 - 50:00 audience sure Gabriel do you have any any comments to respond to Gregory or to add there yes um can you guys hear me okay I shift I shifted to airpods you're good can um yes so I I uh I'm not familiar with by simulation Theory but I have some assumptions so I want to um I want
• 50:00 - 50:30 to respond to areas where I think there's some overlaps where I think there's some gaps and then some questions to to to make better sense of it um so first of all when we first we started verses we started from the perspective of spatial Computing and cognitive Computing as a function of SpaceTime understanding and so we we we began by developing a
• 50:30 - 51:00 language called hyperspatial modeling language which was a way of um you know essentially how can you get computers to understand what all those different sensory inputs might be and how they might be fused together to have real world um coherent digital twins right how could they understand the physical activities object States properties but also social ones like you just mentioned ethics morals which turned out have have many uh physical cories right
• 51:00 - 51:30 um we spent you know several years developing that we've been working with the i e for the last four years it's just been voted in as a standard for AI interoperability explainability and AI governance um I expect that to be ratified in early q1 of next year um and I you know felt pretty proud of that um no one had ever built these sort of ontological Frameworks or arguably epistemological Frameworks that were very all-encompassing because we saw that
• 51:30 - 52:00 everyone take sort of Peace meal industry specific or technology specific approaches and so amazing group of people from around the world that worked on that about three years ago I I I was introduced to Carl friston through a professor at MIT and he was like you you gotta talk to Professor friston who i' never heard of because a lot of the things that you're saying are very similar like like Gregory just mentioned the idea that things are compositional that they're hierarch iCal sort of nested structures right meaning things are made up of other things what I we
• 52:00 - 52:30 often refer to as honic but our our peers often just refer to as nested um these were very similar um the idea of the sort of structures um that Greg was that you were talking about as being representations of aspects or attributes of the world was very similar um and I remember the moment where uh Professor frisen I said look what what do you think we're not modeling here because I haven't found anyone in years that could come up with something that we can't model whether it's at the physics level
• 52:30 - 53:00 whether it's you know whether it's materials whether it's that's light refraction like anything that you could describe we should be able to describe in hsml and he said how about probability like what what do you mean and he said well you can model all these things but you think things I'm paraphrasing he's he's British and he says it much more interestingly he says you think those things are what they are they are where they are and they're doing what you think they're doing but they only probably are with respect to
• 53:00 - 53:30 what what matters from a representation perspective given a decision you need to make and that blew my mind and so the the notion that you need to be able to have representations of the world that if they're if they're to be like uh uh if they are similar to how the brain represents information they're Dynamic they're probabilistic they're not fixed um they're are Trends and there are you know structures that are common hence
• 53:30 - 54:00 being able to say something is chemistry or something is biology or something is an apple or something is banana those features and states are are you know have some degree of stability but it turns out you know everything is made up of Parts but all those parts and all those relationships are constantly changing Dynamic and and you have to be able to model that part um and so that's the thing that in any representational structure uh and I certainly came to that you realization after working with Carl for the last few years um and and genuinely trying to
• 54:00 - 54:30 understand not from a computational perspective which ultimately to be useful I agree is what we have to do and what what we're doing but we have to understand how does nature do it first I mean it's the best best the best lab in the world um so my that's my my comment I I think that GRE Greg's absolutely correct we we are very aligned and I think we're approaching it from different angles that are also complimentary but I think we're talking about structure and the process of intelligence is not limited to structure
• 54:30 - 55:00 or or sort of um specific sort of representations but probabilistic so my sense is that this is where I'd love to learn more B simulation is a kind of semiotic um approach to signs and symbols and representations which I would agree can be you can represent anything it turns out representing probability uh is it is its own kind of challenge which is usually sort of basian model and this is a what you know free energy modeling gives you the
• 55:00 - 55:30 framework to do and as a result that was how we ended up at active inference but I believe we're you know we're fellow Travelers on on a very similar journey and I do I do think that next year we will be talking about the impacts of reasoning or the ethics of reasoning um so yeah I think that's those are my initial thoughts if it's all right for me to respon F first of all thank thank you for that I I wanted to I wanted to just reemphasize what I said said at the very beginning B simulation is not about
• 55:30 - 56:00 structure it's about Behavior the principle is anything you can do I can do right so it's it's all about behavior and in particular it's about comp uh computational behavior um what what's very interesting is that Jonathan warell a noted researcher um in computer science and I have shown that um for any given system where we have a natural notion of B simulation we can not only
• 56:00 - 56:30 automatically derive um a stochastic version which includes probability but that's just one norm-based probability right so that's like the coin flipping kind of stuff but we can also automatically derive two Norm probabilities that's Quantum stuff so for any given given logical notion of computation we can autod derive a corresponding stochastic version and a Quantum
• 56:30 - 57:00 version uh so that that's why that's why this approach is uh uh is starting to um uh for me make me feel a little bit more confident a little bit more comfortable because we're as you as you suggested Gabriel you have to be able to address uncertainty in the form of probability but you also have to be able to address uncertainty in the sense of quantum mechanics which is a different kind of uncertainty fundamental different kind of uncertainty so great conversation I
• 57:00 - 57:30 would love to talk to you more sounds like we have a conversation has to happen after this uh so first of all thank you to both Greg and to Gabriel uh wonderful conversation I appreciate that we are getting the Grim Reaper red sign here so I think it's time for us to head off thank you