Written by: Techub News Compilation

Introduction

At the recently held ChatGPT Talks 2023 conference, Turing Award winner, New York University professor, and Meta’s Chief AI Scientist Yann LeCun delivered a thought-provoking keynote speech. As a pioneer in deep learning and convolutional neural networks, LeCun leads cutting-edge AI research at Meta and has recently pushed for the release of the open-source large model Llama 2. In this speech, he did not excessively praise the current hot large language models (LLM), but sharply pointed out their core limitations and outlined his vision for the next generation of AI systems—"goal-driven AI." This talk is not only a technical foresight but also a profound reflection on the development path and open ecology of AI.

Summary

• Current autoregressive large language models have "fundamental flaws," unable to truly understand the world, perform reasoning and planning, and the hallucination problem cannot be eradicated through existing architectures.
• LeCun predicts that existing large models will disappear within 3 to 5 years, replaced by "goal-driven AI" systems that can perform internal planning and reasoning.
• Achieving human-level AI requires overcoming three major challenges: learning world models, learning reasoning, and learning planning complex action sequences.
• Open source is the inevitable future of AI infrastructure; like the foundational software of the internet, closed proprietary systems cannot be accepted globally.
• There is a need to abandon current popular generative models, probabilistic models, contrastive methods, and reinforcement learning, and switch to self-supervised learning based on "joint embedding prediction architecture" to build world models.

The Glory and Fundamental Flaw of Autoregressive Large Models

Yann LeCun explicitly stated that, despite the impressive achievements of autoregressive large language models represented by ChatGPT, they are "really not that great" in the eyes of the AI research community. These models are trained solely on massive amounts of text, lacking direct experience of the physical world, and thereforelack common sense, with an extremely limited understanding of reality. They are essentially a "divergent process": the model predicts the next token autoregressively, and any slight prediction error at any step can be exponentially amplified in subsequent steps, leading to answers that may diverge completely from facts in longer texts. This is an inherent defect of their architecture that cannot be easily fixed.

LeCun compared the abilities of large language models to the "System 1" thinking described by psychologist Daniel Kahneman, which is fast, intuitive, and requires no deep thought. However, the core of human intelligence—"System 2" thinking, which involves solving problems through deliberation, reasoning, and planning—is completely missing in current large models. They cannot perform complex multi-step planning, such as organizing a trip from New York to Paris, breaking it down into tasks like booking tickets, going to the airport, boarding, and so on.

More critically, LeCun emphasized thatmost of human and animal knowledge is non-verbal. Babies master a wealth of intuitive knowledge about the physical world (e.g., object permanence, gravity) through observation and interaction before they learn to speak. This common-sense knowledge is something large language models can never truly acquire from text alone. Therefore, he believes that AI based solely on text training can never achieve human-level capabilities.

Open Source AI: An Inevitable Future Infrastructure

Before delving into technical critiques, LeCun first discussed the vision behind Meta's recent release of the open-source large model Llama 2. He believes that AI, especially the future AI assistants that will act as intermediaries between humans and the digital world, will become therepository and infrastructure of all human knowledge. Looking back at internet history, all its foundational infrastructures (such as Linux, Apache, HTTP) have been open source. Proprietary commercial solutions have ultimately lost to open ecosystems.

LeCun asserts that the AI field will replay this history. Future AI infrastructures must be open-source for three reasons: firstly, from a security perspective, open-source software undergoes broader scrutiny and is often more secure than closed-source software; secondly, building systems that encompass all human knowledge requires contributions from millions of people globally, similar to the Wikipedia model, and cannot be completed by a few proprietary companies; thirdly, many countries and regions in the world will not accept that all information about their citizens comes from closed-source systems controlled by a few tech companies on the U.S. West Coast.

Therefore, Meta firmly stands on the side of open-source research. LeCun envisions that in the future only a few open-source foundational large models will be needed for the entire ecosystem to fine-tune and develop various application products, greatly promoting innovation and popularization.

Goal-Driven AI: Blueprint for the Next Generation System

What kind of AI systems should be built to overcome the flaws of existing large models? LeCun proposed a cognitive architecture of "goal-driven AI." At its core is a"world model" that predicts outcomes. This system includes several key modules: perception module (estimates the state of the world), world model (predicts action consequences), actor module (imagines action sequences), and cost module (evaluates the quality of outcomes).

The operating principle of the system is: given a goal (for example, "go to Paris"), the actor module imagines a series of actions, the world model predicts the results of these actions, and the cost module assesses whether these results meet the goal and satisfy a series of "guardrail" goals (such as safety, non-harmfulness). The entire system searches for an action sequence that minimizes the total cost through an optimization (reasoning) process. This process itself is planning and reasoning.

LeCun particularly pointed out that this architecture isinherently safe. As long as the guardrail goals (such as toxicity detection) are hard-coded into the cost module, the system will automatically optimize to avoid generating harmful outputs during reasoning, without needing complex "human feedback-based reinforcement learning" to fine-tune the entire model, which is also difficult to prevent "jailbreak" attacks.

The real challenge lies in achievinghierarchical planning. Humans can easily decompose "go to Paris" into high-level goals like "go to the airport" and "board the plane," and then further decompose into specific actions such as "stand up," "open the door," "take a taxi," down to millisecond-level commands to control muscles. Current AI is completely unable to spontaneously learn this hierarchical representation and planning ability, which is one of the core challenges that must be confronted to achieve human-level AI.

Abandon Generative Models: Learning World Models with Joint Embedding Architectures

How can we construct world models that can predict accurately? LeCun provided an answer that might surprise the current AI community: Abandon generative models. The currently popular generative AI (including autoregressive large models) belongs to generative models, which attempt to directly predict future specific data (such as the next token or the next frame of video). However, when faced with the highly uncertain real world (such as predicting the next frame of a video), generative models often can only produce vague average results because they are forced to assign probabilities to all possibilities.

LeCun advocates for the "joint embedding prediction architecture." The core idea is: instead of directly predicting the raw data Y, it predicts theabstract representation of data Y. The system uses an encoder to map the input X (such as the current video segment) and the target Y to be predicted (such as the next video segment) into representation space, and then makes predictions in this abstract space. The key to the encoder is that it can selectivelydiscard details that are hard to predict but are irrelevant to the current task (such as the random shaking of leaves in the wind), retaining only features that are easy to predict and relevant (such as the movement of cars on the road). This makes the prediction task feasible and efficient.

However, training such architectures faces the risk of "collapse": the system may learn to output a constant representation to easily minimize prediction errors. To this end, LeCun's team developed methods like VICReg, employing regularization techniques to ensure that the representation space has enough variance and information. Experiments have shown that these methods significantly outperform generative models based on reconstruction in self-supervised learning for complex modalities like images and videos.

Conclusion: The Challenges and Inevitability Towards Human-Level AI

Yann LeCun concluded that to achieve human and even animal-level AI, three major challenges must be solved: 1. Learning representations and predictive models of the world; 2. Learning reasoning; 3. Learning to plan complex action sequences to meet multiple objectives. His envisioned "goal-driven AI" architecture combines hierarchical planning, world model-based reasoning, and energy minimization optimization, representing a potential direction to address these challenges.

He boldly predicts that the current forms of autoregressive large language models will disappear within 3 to 5 years, replaced by systems capable of internally planning answers, reasoning based on world models, and driven by goals. Such systems will simultaneously ensure factual accuracy, fluency, and safety.

Finally, LeCun discussed his views on superintelligence. He believes there is no such thing as "general artificial intelligence" because even human intelligence is highly specialized. He is confident that in the future, machines will surpass humans in all areas of human intelligence, but this should not be feared. Humans will always maintain control over these smarter machines, just as our cerebral cortex has not escaped our control. The future landscape is that everyone will be assisted by AI systems that are smarter than themselves, jointly promoting the advancement of knowledge and civilization. The cornerstone of achieving this future will be open, collaborative, and continuous foundational research.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。