[Web3 Not in the Books] Autonomous Intelligence. Insights from Yann LeCun

AI Network
10 min readMay 2, 2024

--

Can AI become self-learning?

Originally broadcast on YouTube. See the original live stream with AI Network CEO Minhyun Kim here.

Yann LeCun’s influential paper “A Path Towards Autonomous Machine Intelligence” represents a paradigm shifting idea in AI.

LeCun, Turing Award winning computer scientist, put forth the idea of autonomous intelligence as AI that is capable of performing not only specifically programmed tasks, but is also able learn and adapt like animals & humans do and ultimately make decisions independently.

LeCun’s Core Considerations on Autonomous Intelligence

Three critical questions frame LeCun’s exploration into autonomous intelligence.

1. How can machines learn to represent the world, learn to predict, and learn to act largely by observation?
Involves the development of algorithms that enable machines to interpret complex data inputs & environments and learn from them in a wide-ranging way without manual or human intervention.

2. How can machines reason and plan in ways that are compatible with gradient-based learning? Gradient-based learning is a method used to help improve an AI model’s accuracy by minimizing errors.

3. How can machines learn to represent percepts and action plans in a hierarchical manner, at multiple levels of abstraction, and multiple time scales? Humans and many animals are able to conceive multi-layered and levelled abstractions of knowledge with which long-term predictions and long-term planning can be performed. How can machines emulate this form of reasoning?

Autonomous Intelligence vs. Purpose-Oriented AI Applications

Here we must understand the distinction between autonomous intelligence and purpose-oriented AI. Our current AI models are made to perform and learn through specific tasks. ChatGPT, for example, is designed as a chatbot to answer text queries from users and learns with each query it’s given. Similarly, many Stable Diffusion models are specifically tasked with creating images from text-inputs and learn to create more accurate images over time. Both are designed for a specific purpose and learning from their outputs get better, faster and more efficient as they perform.

AI has developed into something that satisfies a specific purpose, but the world as it is does not reflect this. True knowledge, adaptation and survival come from learning about the world at large and from the environments that entities find themselves in. AI which is based specifically in a narrow number of tasks will not be able to learn much outside those scopes. Autonomous intelligence is the opposite of this. It’s the notion of machines learning from their environments and ‘worlds’ like animals and humans do, where the system learns broadly and without specific focus, ultimately equipping it to engage in general problem-solving across a wide range of tasks without specific programming for each.

This represents a significant shift from the task-based AI models we currently have to machines that learn on their own without specific focus or manual input.

Autonomous Intelligence is more than just about being ‘Smart’

Autonomous intelligence is not about making machines ‘smart’ or more intelligent simply for the sake of being more intelligent. The primary goal here is about understanding the world. Humans and animals have developed intelligence by understanding and adapting to their environments. To understand the world any entity capable of intelligence must learn about the world around them. All intelligence is part of the world, and survival happens by adaption. Since artificial intelligence also exists within our world, if it is to evolve, it needs to learn and adapt to the world around it.

This is the philosophy of AI network. AI network is not only about AI and is not simply just an AI development platform. AI Network is an ecosystem. It isn’t specifically a place where AI lives — the world where AI lives. AI Network is the space that connects AI to the world; the world is an AI network.

The World Model Concept

A central element of LeCun’s vision for autonomous intelligence is the “world model”.

All humans and many animals build a ‘world model’, that is, a model of reality existing in their minds which they have gleaned through their interactions and experiences in the world. Different animals build different world models based on their experiences. Humans, for example, see in the visible color spectrum whereas many snakes see in the infrared spectrum. We can distinguish thousands of scents through our olfactory senses, but dogs’ senses of smell are many magnitudes more powerful than ours. These differences in experiences will create different world models in the minds of us and animals, but we will all share some common sense of reality, like a sense for gravity and its consequences.

In Yann’s words, we seem able to learn enormous amounts of background knowledge about how the world works through observation and an incomprehensibly small amount of interactions in task-independent, unsupervised ways. Using such models, humans and some animals can learn new skills with remarkably few trials. We can predict the consequences of our actions, reason, plan, explore and imagine new solutions to problems. Importantly, we can also avoid making dangerous mistakes when facing an unknown situation, which all stem from our ‘world models’.

This is where machines need to catch up, and where autonomous intelligence comes in. Using the example of self-driving cars, the system might require thousands of trials to learn and reinforce knowledge that driving too fast in wet conditions affects braking speed, turning too little around a corner could result in collision and learning to slow down to avoid skidding. In contrast, humans can draw on their intuitive sense of physics — stemming from their world model — to predict such outcomes without needing to carry them out, and largely avoid fatal courses of action when learning a new skill.

The System Architecture for Autonomous Intelligence

AI systems can develop their own ‘world models’, which would be an internal representation of the external world. The system architecture for learning from those models is a bit more complex. According to LeCun, this architecture for autonomous intelligence comprises different modules.

Yann LeCun’s System Architecture for Autonomous Intelligence. Source.

  • The perceptual module processes raw sensory data and estimates the current state of the world.
  • The memory module that stores and retrieves learned information & keeps track of the current and predicted world states.
  • The configurator module takes inputs from all other modules and configures them to perform the task at hand.
  • The cost module computes energy output through; Intrinsic cost (immutable) computes the immediate energy of the current state, and the Critic predicts future values of the intrinsic cost.
  • The actor module computes proposals for action sequences.

This modular approach allows for more robust and scalable AI systems that can adapt and learn from their interactions with the environment and build their own world models. Given the self-driving car example in the previous paragraph, it would ultimately take autonomous intelligence far more trials than humans or animals to create their own world models.

The Perception-Action Episode: Mode 1 and Mode 2

Mode-1 perception-action. Source.

The diagram above is a perception-action episode, labeled by LeCun as mode 1. It estimates the state of the world through the eyes of an actor (state 0) and then the state of the world after the actor directly computes and carries out an action (state 1). This mode does not take the world model or the cost of the action into account. It represents a reactive form of action and involves no complex reasoning.

Mode-2 perception-action. Source.

This diagram is mode 2 of the perception-action episode. This denotes an actor perceiving the world (state 0) and then using their world model to create a sequence of actions based on predictions, then predicting the state of the world after their actions have been carried out (state 1). This mode takes cost into account and the actor predicts to minimize cost while maximizing outcome. This mode is proactive and involves complex reasoning.

Purpose-Oriented AI can be seen as mode 1. Purpose-oriented AI is reactive and does not learn much outside of its defined scope. It does not predict and cannot plan nor reason in a complex manner. This is akin to animals that lack high reasoning function, and simply react to external stimuli, like a cockroach reacting to external movement by running away.

Autonomous intelligence can be seen as mode 2. Actions and their outcomes can be predicted based on the entity’s world model. Each perception, prediction and action can be reasoned and costs can be taken into account. The entity can use complex reasoning to assess, predict, plan and forecast the consequences of action. This is akin to a human getting behind the wheel of a car and knowing not to drive head-on into a wall. Using their world models they can predict that the outcome of the action would be detrimental to them, and so can opt not to carry it out.

State Transition Systems in Blockchain, AI and AI Network

Both examples of Modes 1 & 2 are state transition systems — they show the different states and transitions between them in a system.

In order to understand state transitions in this context, we must first understand states and their implications, specifically the stateless and stateful differences between Web2 and Web3. Web2 is the age of the internet-user-turned-creator (contrast with Web1, an internet full of static, non-interactive webpages), where users contribute to the internet in the form of personal information & data, written content, pictures, videos etc. In the Web2 world user data and interactions are handled and stored by vast commercial entities, leaving the individual little control and virtually no ownership over their digital information and assets. This can be seen as a stateless system.

The stateless Web2 reality presented a challenge for the development of genuinely self-sustaining, self-governing and self-learning intelligence systems, whose development are significantly hindered by a lack of autonomy and lack of certainty of their ‘state’.

Web3, the third revolution of the web built on the ideals of self-governance, individual ownership and autonomy, introduces a stateful internet, where states are maintained and managed through permissionless blockchain ledgers. Blockchain technology provides a stateful internet by decentralizing the web, ensuring all transactions and interactions are stored in a permanent and transparent manner.

State transition systems form a critical component in blockchain technology because every ‘state’ of any given system is recorded immutably on a decentralized ledger. This lays the foundation for the emergence of true autonomous intelligence, with entities able to exist in a stateful manner, with every perception, reasoning and action recorded in sequential states on the blockchain.

By integrating blockchain and AI, all under the umbrella of Web3, autonomous intelligence has the opportunity to emerge using state transition systems managing operation sequences, from prediction & inference to reasoning & action. In an AI-driven blockchain ecosystem, this would imply a flow where AI not only predicts the next state but also infers the best actions based on past data, ultimately engaging in complex reasoning to optimize decisions.

In the AI Network the different states of an AI are recorded on the blockchain and the evolution of the AI can be explicitly discerned through the chain. We can see AI’s on AI network as having the following flow:

Prediction -> Inference -> Reason

State 0 would be the ‘old world’ state — how the AI’s world model looks before any action. When there is a prediction and inference, the AI goes to state 1 with some form of action (like a chat reply, a message, a transaction etc). Further states are added in the sequences, and all states are recorded on the blockchain. With enough inferences and enough states the world changes.

Yann LeCun’s insights into autonomous intelligence lay down a framework for AI independently predicting, learning, reasoning and acting. Environments create entities within them and those entities affect their environments. Autonomous AI differentiates itself from current AI because it represents sovereignty through self-learning and action. Self-sovereignty is a core concept of Web3 and Web3-based AI must be sovereign within itself by its very nature.

ChatGPT, for example, is not a sovereign entity. It’s controlled by its builders and its learning can be directed — the plug can be pulled at any time. If ChatGPT was open source then it would be self-directing and no one would own it. This is the idea of AI Network — sovereignty and open source models within the AI space and the world at large, where private ownership of applications don’t exist and the direction of AI is decentralized.

AI Network is a decentralized AI development ecosystem based on blockchain technology. Within its ecosystem, resource providers can earn $AIN tokens for their GPUs, developers can gain access to GPUs for open source AI programs, and creators can transform their AI creations into AINFTs. The ultimate goal of AI Network is to bring AI to Web3, where everyone can easily develop and utilize artificial intelligence.

If you want to know more about us,

--

--

AI Network

A decentralized AI development ecosystem built on its own blockchain, AI Network seeks to become the “Internet for AI” in the Web3 era.