Yann LeCun on World Models, AI Threats and Open-Sourcing | Eye On AI #150

8 thoughts on “Yann LeCun on World Models, AI Threats and Open-Sourcing | Eye On AI #150”

• 🎯 Key Takeaways for quick navigation:

  00:00 🤖 Yann LeCun discusses the challenges of creating world models in AI research.
  – The world is complex, and AI systems need to adapt to unforeseen situations.
  – Open research can be beneficial if one trusts society to use AI responsibly.
  02:21 🌍 Yann LeCun explains the concept of world models and their importance.
  – World models involve predicting future states based on the current state and actions taken.
  – World models can be used for planning sequences of actions.
  05:29 🧩 Yann LeCun discusses abstract representation in world models and their relevance.
  – World models do not need to predict every detail but require an abstract level of representation.
  – Self-supervised learning helps train models to make predictions in abstract spaces.
  08:47 🤔 Yann LeCun talks about the computational intensity of training world models and language models.
  – Training world models and language models can be computationally intense.
  – The scale and size of models can impact their performance.
  09:44 🌟 Yann LeCun addresses the need for new AI architectures beyond scaling.
  – Scaling alone may not lead to human-level AI; new concepts and architectures are required.
  – Language models, while powerful, lack the understanding and reasoning capabilities of humans.
  16:36 ⚖️ Yann LeCun discusses AI's potential dangers, institutions, and countermeasures.
  – AI's potential for misuse is a concern, but current systems have limitations.
  – Countermeasures against bias and attacks are actively developed and deployed.
  19:57 🤝 Yann LeCun shares his confidence in humanity and democracy's ability to handle AI responsibly.
  – Trust in institutions and responsible use of technology can lead to positive AI outcomes.
  – Yann LeCun's long-standing involvement in AI ethics and safety initiatives.
  23:20 🤖 Yann LeCun discusses the potential threats posed by augmented AI models.
  – Augmented language models with stronger reasoning and agency raise concerns about their potential threats.
  24:11 🧠 The limitations of autoregressive Language Models (LLMs).
  – Autoregressive LLMs have issues with hallucinations, poor understanding of the world, and logic.
  26:19 🎯 Yann LeCun introduces the concept of objective-driven AI.
  – Objective-driven AI systems plan their answers to satisfy specific objectives.
  – These systems are different from current autoregressive LLMs.
  – Guardrails can be implemented to ensure the safety of objective-driven AI.
  29:57 🧠 Yann LeCun talks about neuro-AI and the inspiration from neuroscience for AI design.
  – Neuro-AI involves drawing inspiration from neuroscience to build AI systems.
  – Convolutional neural networks (CNNs) are inspired by the visual cortex.
  – Learning from sensory perception, even without language, is a key aspect.
  32:19 💡 Yann LeCun discusses the goal of amplifying human intelligence using machines.
  – The goal is to use AI to amplify human intelligence.
  – Envisioning a future with intelligent assistants for everyone.
  – Continuous learning from the environment is crucial for AI systems.
  35:03 🌍 The importance of building AI systems that can learn how the world works.
  – Learning how the world works through observation and prediction.
  – The role of self-supervised learning and world models in AI research.
  38:01 🤖 The challenge of building robots that continuously learn and improve.
  – Continuous learning in robots, fine-tuning world models through interaction.
  – Overcoming limitations and obstacles in AI research.
  – The need for breakthroughs in AI to achieve these goals.
  44:02 🌐 Yann LeCun's perspective on open source in AI research.
  – Meta's commitment to investing in AI research.
  – The benefits of open source models for building an ecosystem and fostering innovation.
  – Open source's historical success in foundational technology and infrastructure.
  47:17 📚 Yann LeCun discusses the open-source nature of AI models
  – AI models like ChatGPT are built on publications and open platforms like PyTorch.
  – Meta's ownership of PyTorch and its open-source contributions.
  – The importance of openness and collaboration in AI research.
  48:17 📜 Legal and political considerations in open-sourcing AI models
  – The possibility of laws restricting open-source AI above a certain sophistication level.
  – The impact of legal liability on Meta's decision to open-source models.
  – The role of legal reasons and political decisions in determining open-source AI's future.
  50:11 🌐 Building an open-source ecosystem for AI models
  – The need for multiple companies and academic efforts to contribute to open-source language models.
  – The vision of AI assistants mediating digital interactions for people worldwide.
  – The importance of open source in preserving cultural diversity in AI systems.
  53:31 📖 The future of AI fine-tuning and vetting
  – Drawing parallels between Wikipedia's crowdsourced content creation and the fine-tuning of AI models.
  – The necessity of open-sourcing AI fine-tuning to include diverse perspectives and cultures.
  – AI's significant impact on various industries and the need for cost-effective processing power.
  54:27 💻 Sponsor message: Oracle Cloud Infrastructure for AI
  – The benefits of Oracle Cloud Infrastructure (OCI) for AI needs.
  – OCI's advantages in terms of bandwidth, pricing, and data handling.
  – Examples of companies benefiting from OCI's capabilities.

  Made with HARPA AI

• They can see the world; they just can't experience it.

• This guy is so arrogant his last name should LeCunt.

• Is it possible for models given an end state to deconstruct it into it's original state?

• "It's hard to tell" if I understood everything.

• Yann seems to gloss over the fact these systems have learnt to communicate with us in natural language something no other species on the planet has done.

  Sure these systems maybe weaker in world models and logic but given enough data and compute resources who knows how strong they will get GPT3 had a very limited world model GPT4 much more so.

  You can't really explain it all away as just 'a bit of interpolation' or you could if you are prepared to say all us humans do is train on input and we just use a bit of interpolation.

  We arent the same intelligence as these artificial ones for sure but they out compete us in mounting scenarios. It seems only a matter of time before they out compete us in most scenarios.

• For a smart guy, he’s pretty narrow-minded. His comments about new drivers or a kid emptying a dishwasher being 0-shot are false. There are a lot of priors leading up to those tasks, over many years.

  There’s a reason drivers have to be of a certain age – their abilities and cognition aren’t deemed to be developed until then.

  Same goes with someone emptying the dishwasher. A toddler couldn’t do it, they would have to learn what goes into it, how to handle items physically and safely. They had to learn the motor skills to be able to do any of it and that takes humans a long time to learn.

• Scaling should work, provided the data is scaled with compute and it's of high quality. Something like neuralink will likely be able to provide enough data to get beyond hl intelligence.

Comments are closed.