Mark Zuckerberg on AI with Lex Fridman

Mark Zuckerberg’ thoughts on AI Agents ecosystem, Meta’s AI products, Timeline for AGI, Open source, and Existential threat of AI with the Lex Fridman podcast. A deeply insightful and nuanced perspective on what the world could look like with the new wave of AI breakthroughs.

AI Agents Ecosystem:

Mark's view is that people will want to talk to multiple agents vs a single agent like chatGPT or Bard or Bing. Just like people want to use different? apps for different purposes or have different people play different roles in their life, they will also have a number of AI agents that they interact with for different purposes. Some example directions:

?? Social Assistants: that can serve different purposes, help you connect with your friends & family better by reminding you and providing you with all the context of other people’s lives.

?? Creator Assistants: that can embody the personality of the creators and engage / connect with their fans more deeply. Help creators negotiate and engage with partners.

?? Business Assistants: that help businesses sell their products, answer customer questions and provide the right level of support.

Mentors or life coaches: that can give you advice, cheer you up, and help you through the challenges that people face on a daily basis

?? Tutors: that are available all the time and can help kids & adults get unstuck in a non-judgemental manner.

?? Coding assistants: that can accelerate and improve the experience of building software

?? Other functional relationships we have in our lives. Number of companies will take a crack at these.

Meta’s approach

?? Empower people to create these agents easily whether it for creators to engage with fans or businesses to engage with potential customers.?

?? This is a unique product direction from the other big companies and also aligns well with the open source approach, more community oriented, more democratic approach to building out the products and technology around AI that Meta wants to take.

??In addition, Generative AI will be deeply integrated into all of Meta products: (i) Creating/editing photos using natural language, (ii) Automatically generate creatives for A/B testing ads and (iii) Building metaverse worlds using natural language

Timeline for AGI

?? It is clear that we have had a big breakthrough with transformer and diffusion based models. We can do some pretty interesting things with them.

?? On one side there is the argument that we just had a big breakthrough, we are maybe only a few more big breakthroughs from something really crazy, something that looks like general intelligence. We could stack breakthroughs on top of each other every 6 months.?

?? On the other side, we have seen that a breakthrough typically leads to the Gartner hype cycle: hype followed by the trough of disillusionment. The next big breakthrough could take 5, 10, 15 years and we have to sit with this one for a while.

?? There are very large error bars on what happens next, when the next big breakthrough will be.

?? Mark’s view is that the next breakthrough may take longer than people’s expectations.?

?? He closes with “I don’t have any particular insights on when a singular AI system with general intelligence will be created. I’m not going to opine on what or how many breakthroughs are going to be needed to get to super intelligence because no one knows. :-)”

?? Yet, we can build a lot of cool products with this breakthrough!

Open Sourcing LLaMA foundation

?? Meta approach to fundamental research has been to bring in the best people in the world, give them access to industry scale infrastructure and then share their work with the world. Ethos is to share what’s invented broadly. Done that with a lot of AI tools like Pytorch to a number of models including LLaMA.

?? LLaMA was released to researchers and the community has exploded with fast iterations building on top of it: running on edge devices, making inference faster, finetuning using RLHF etc.

Value & Risks of Open Source

?? Open source approaches to Foundation Models could have the same benefits of open source software, more secure due to the security of the community vs being developed in a silo to get security through obscurity.

?? AI development is still in early stages the benefits outweigh the risks. For example the LLaMA model is 10x smaller than the ones Google and OpenAI are using right now. Approach could change if we get closer to super intelligence. But we have time to think these issues through.

Existential threat of AI

?? Anytime where there's a number of serious people who are raising concern that is that existential about something that you're involved with? I think you have to think about it. So has thought about it quite a bit.

?? Mark thinks we are quite a few steps away from super intelligence so the existential threat is much further out. There are more near term risks of people using AI to do harmful things like fraud or scams etc that need to be tackled now. These are going to be a pretty big set of challenges to tackle. So he is worried that people are focused on the tail risk of existential thread instead of the doing a good job with the risks that are more certain near term.

?? In addition, Marks talks about separating the concepts of intelligence and autonomy gets conflated in discussions of risk. He uses the analogy of neocortex benign the powerful machinery for intelligence benign subservient to the basic impulse mechanisms in the brain. Applying the same logic we could build highly intelligence systems that have the capacity to think but not have autonomy. Mark’s view is that we need to think carefully about the development of autonomy. Even simple less intelligent s/w like viruses can do a lot of damage when they have runaway autonomy.

?? So when we think about governance and policy we need to distinguish between intelligence and autonomy. Policy makers should focus on autonomy that is given to AI systems.?

Podcast link: https://www.youtube.com/watch?v=Ff4fRgnuFgQ