In the era of pre-training, we were writing one very large exponential curve towards improving the performance of the models, and that curve was very predictable. It had clear characteristics where the more resources you put in, the more results you would get out. There were a lot of questions at the time of whether or not this curve would continue, and even to the degree it did continue, diminishing marginal returns. I think this was one of the more recent eras of AI skepticism.
Then in 2024, towards the end of 2024, began the era of reinforcement learning. This was the next major wave that stacked on top of the previous wave, and this is when we saw the models begin to learn to reason. There's always this meme of what did I see? Well, Ilya saw reasoning from the models, and it was very clearly this next mega wave of the technology. And even reinforcement learning had clearly some limitations, and there were all these questions around how long that era was going to last and maybe there were limitations in that era.
And then at the end of 2025, so just a few months ago, I think we've really begun an entirely new paradigm or era of the technology, which is really beginning the era of recursive self-improvement. And I think we're really starting to see this in the development of our models. I think most of the major labs are seeing this in the development of their models, which is that the models themselves have now become instrumental in accelerating the process of producing the next AIs.
And so from the outside, what you see is maybe an acceleration of overall development velocity, an acceleration in the speed at which the new models are being released. Internally, what we see is just dramatic speedups. I think the productivity of an individual researcher has grown dramatically, and we don't expect that to stop. In fact, we expect it to kind of smoothly continue accelerating, where the output of an individual researcher and individual engineer will itself continue growing exponentially as the technology improves.
And the other mega wave that is somewhat coincident with this new paradigm within the development of the technology of recursive self-improvement is really the arrival of agents in earnest. This is something that has been talked about since I honestly want to say 2023 with a lot of fanfare, and for a long, long time it was one of these hype buzzwords that never actually lived up to the expectations.
And I think also at the end of 2025, and toward the middle to end of 2025, we started to see agents actually work. These are where the models can actually start taking actions and start automating entire workflows and being dramatically more productive as actual entities unto themselves.
And this era started with coding agents. We're starting to see it really percolate through personal agents. This is a big part of what we believe in at Meta, and I think over the course of 2026 we will see large-scale agent deployments in many areas of the economy, many areas of the world, and the GDP of AI, so to speak, is going to grow exponentially.
Well, that is very welcome to hear. I think you hit on something that is important because I remember when I first started doing this in the job, the conversations were all around these scaling laws and would they continue. And you mentioned the skepticism. Not only have they continued and continued to deliver this exponential growth, it feels very much like the last few months have seen an acceleration. It's what you touched on. It's what Dario and Dennis have also said to me and publicly, driven by this recursive self-improvement that the models themselves are now able to help us develop the new models quicker.
So we're at this quite exciting but also a moment of some anxiety because of that accelerating development timescale. I wanted to just then talk about trust. I'll give you a story. I have two young girls, and as they're using their chatbots and things, they're always like, please, you know, so-and-so, and thank you when they finish writing their request or they get the answer back. Lots of pleases and thank-yous everywhere.
And I thought that was interesting. But I said to them, look, you don't need to do that. There's not a person. It doesn't matter. And by the way, it costs a fortune when we have to do the compute for that. And they said to me, you know what, Daddy, they said, if AI takes over the world, we want to have been kind to the AI. And I thought that was a good insurance policy, but it touches on this anxiety.
Yeah. In that maybe across the world we're seeing different attitudes towards AI. In countries like India, where we are, there's enormous optimism and trust. And then actually in Western countries, we're seeing that anxiety is still the dominant feeling towards AI. Now I think closing that confidence gap is as much a policy task as it is a technical one, but I'd love to just get your thoughts.
Obviously you are at a B2C company dealing with billions of consumers every day. How do you think about this trust question? How do we get trust up in places where it is not, for something that you believe, I believe, and everyone in this room probably believes can do untold good?
Yeah, this is one of the most important questions, and actually perhaps even more so than the technology and the development of the technology, it may govern how successfully the technology diffuses more so than anything. And we think deeply about this because our vision for the technology is that we can deploy personal agents that know you better than most people know you. Know about your health, know about your goals, know about your relationships, know about your family and your friends.
And can really help you live your life more effectively. Can help you be healthier, can help you accomplish projects that you never would have had time for, can help you build the business you always wanted, can help you explore your interests. And to properly build that requires an immense amount of trust from consumers, from governments, from effectively everyone, from all organizations.
And so this is one of these interesting moments where I think because the form of the technology that we want it to take is so intimate and is so deeply tied with what we think is really the future of the consumer life, and the consumer is at the core of their relationship with technology, figuring out ways to develop the technology in partnership with governing bodies and in partnership with nonprofits and other organizations to garner that trust is very important.
This is something we've seen at Meta actually with WhatsApp. We were talking about this actually backstage, but WhatsApp has become incredibly ubiquitous and a real utility for much of the world, in large part due to its commitment to privacy and the trust that we've built around WhatsApp as an organization, oftentimes not necessarily for the optimization of our business objectives but for the optimization of trust in WhatsApp as a platform.
I think a similar journey will be necessary for the AI products of the modern age. And we've already seen various kerfuffles or sort of jabs back and forth between the existing AI products and labs around whether their actions are truly building these levels of trust.
And then I think the topic that you really kicked off at Bletchley in a global way around safety, I think is more important than ever. I think we're already seeing that the models are achieving levels of capability which cause some pause around some of the risks that we monitor for these models as it comes to various national security risks or various cybersecurity risks.
And so the importance of properly testing and then building the right systems to monitor and deploy these models safely is of paramount importance.
And you and I have been talking about this same topic for some years, and that was key to us at Bletchley, was establishing the AI Security Institute in the UK to do exactly that work. And the thought was there was a gap where the technical capability didn't exist amongst governments to evaluate some of these risks. So how could we in the UK hopefully take a lead but in a sense provide a global public good where we work in partnership with companies like yours and the other frontier labs to do pre-deployment testing on these models.
And my thought was always that by doing that, and that's not regulation, but it gives citizens confidence that the technology is safe because ultimately people don't adopt a technology that they don't trust. And my sense hearing from you is, look, there's a role for the companies. I thought the privacy example with WhatsApp is a good one, is how do they design AI products that will engender trust. And then there's a role, it sounds like you agree, for governments to do some of that technical evaluation.
And is your sense that we need more cooperation between governments on that, more interaction with the labs? Is that working as well as it should at the moment?
I think the interaction is working well, but I think this is one area where we always can be more thoughtful. And especially as we enter this phase of dramatically increased AI progress, I think we constantly need to be thinking about how we can do better.
I think one of the other areas where this debate on trust will be won or lost is in the public sector. I think as citizens, if you experience better healthcare, more efficient government services, a state that is quicker at dealing with you, I think this debate goes from abstract to you have a real way to demonstrate benefit.
Do you have a sense from where you sit about where the most obvious use cases or deployment cases are in the public for AI that we should be excited about and that governments should be thinking about?
