Rise of AI Conference 2025 · Berlin

The Future of AI Agents

Pascale Fung · Director at Meta FAIR

At Rise of AI 2025 in Berlin, Pascale Fung, Director at Meta FAIR, argues that AI is entering a new era defined by autonomous, embodied agents rather than simple assistants. She traces Meta FAIR's history, from PyTorch and the first Llama model, developed in Paris, to the No Language Left Behind project covering more than 300 languages, and explains why FAIR insists on open-sourcing its models: only open science lets the community independently verify claims such as which model performs best. Fung calls 2025 the "year zero" of AI agents, distinguishing them from earlier assistants by their ability to carry out long, multi-step tasks autonomously. She then focuses on embodied agents, in wearable devices, avatars and robots, that must model both the physical world and a user's mental state, and warns that hallucination is far more dangerous once agents act physically rather than only answering text. FAIR is exploring JEPA-based world-modeling architectures to build agents that are objective-driven rather than merely trained to predict the next token. She closes with a video scenario of a future dinner party run with wearable and robotic assistance, illustrating how embodied agents could remove physical and mental chores from daily life.

Key takeaways

  • Fung calls 2025 the "year zero" of AI agents because they can complete multi-step tasks that assistants could not.
  • Meta open-sources FAIR's research, including Llama and PyTorch, so scientific claims can be independently verified.
  • Embodied agents in wearables must model both the physical world and the user's mental state to act safely.
  • Fung warns hallucination is far more dangerous in embodied or robotic agents than in text-based generative models.
  • FAIR is exploring JEPA-style world-modeling architectures to build objective-driven AI beyond next-token prediction.

Detailed notes

Open science at Meta FAIR

Fung recounts FAIR's ten-year history: co-founded in 2013 with Yann LeCun, it created PyTorch in 2016, and the first Llama model was developed by European researchers in FAIR Paris before Llama was spun off into the GenAI team. She argues FAIR must keep releasing open models because AI is a multistakeholder, cross-disciplinary technology built on contributions from many scientific fields, and because open source is the only way the community can verify performance claims; she cites her own 2023 benchmarking paper that tested ChatGPT on 23 standard NLP tasks. She also highlights No Language Left Behind, a FAIR project covering more than 300 languages, developed with UNESCO and used by developers in the Global South, now being extended toward 1,000 languages with very-few-sample translation methods.

From AI assistants to autonomous agents

Fung distinguishes today's AI agents from earlier call-center and smartphone assistants such as Siri, which performed single, simple actions like voice search. She calls 2025 the "year zero" of AI agents because they can now carry out tasks requiring far more steps than a human would need, such as scheduling a meeting by researching purpose, timing and contacts. She notes startups can build purpose-built agents by fine-tuning open Llama models, and mentions GAIA as a benchmark used to compare web or digital agents' tool-use capabilities.

Embodied agents and the limits of generative models

Fung argues generative models are optimized to predict the next token or pixel, not to reason or plan, which is why they often fail to follow instructions precisely and why enterprises need fine-tuning. FAIR is researching alternative, objective-driven architectures, including JEPA-based world models, to move beyond this limitation. She describes embodied agents in 2D/3D avatars and wearable devices such as Meta's Orion glasses, which see and hear what the user does and must model both the physical world (laws of physics, what is dangerous) and the user's mental state, for example a math-tutor agent that hints rather than gives answers. She stresses that hallucination, tolerable in a text answer, becomes dangerous once an agent controls a robot's arm or tells a wearer to jump, and says FAIR is building safety and guardrails into these architectures from the start rather than filtering output afterward.

Robotic agents and a future dinner party

Fung describes the merging of robotics and AI into robotic agents that combine a physical body with an AI "brain," and says FAIR has published first results on robotic action planning using JEPA-based architectures. She illustrates the vision with a video scenario: planning a dinner party with a conversational agent that suggests a guest list and menu, an Orion-glasses agent that helps with grocery shopping and hints during cooking, and a robot such as Spot serving drinks and clearing the table afterward, freeing the host to spend time with guests instead of doing chores.

Q&A highlights

Asked when she expects to host a dinner party with a robotic dog, Fung says two to three years, maybe five, depending on what such a robot can do. Asked about her meeting with Pope Francis, she says he gave a speech on using technology to benefit humankind, particularly in healthcare, and said it would be selfish to want the benefits of medical AI trained on others' data while refusing to share one's own data for research.

Quotes

2025, I actually call it the year zero of AI agents.

Pascale Fung

AI agents are more autonomous than assistants.

Pascale Fung

Technology is not there to prevent humanity from progressing. Technology is here to enable us to progress further.

Pascale Fung

Questions answered in this talk

What does Pascale Fung mean by the "year zero of AI agents"?

According to Pascale Fung at Rise of AI 2025, 2025 is the "year zero" of AI agents because, unlike earlier assistants that performed single simple actions, these agents can autonomously complete tasks requiring far more steps, such as researching and scheduling a meeting.

What are embodied AI agents?

According to Pascale Fung at Rise of AI 2025, embodied AI agents are agents built into avatars, wearable devices or robots that see and hear what the user experiences and must model both the physical world and the user's mental state to act helpfully and safely.

Why does Meta FAIR open-source its AI models?

According to Pascale Fung at Rise of AI 2025, Meta FAIR open-sources models like Llama and PyTorch so the scientific community can independently verify performance claims and so AI, as a multistakeholder technology, can be co-developed with society.

Mentioned

a"": ""According to Pascale Fung at Rise of AI 2025

Keep exploring

More from Rise of AI Conference 2025

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 4,933 words

Show the full transcript

So dear ladies and gentlemen um the future of AI agent professor Pascal Funk is the senior director of AI research at Metafare an expert in the world economic forum future council member expert of UN advisory boards on the AI and was invited by governments and um in Europe but also in Saudi Arabia, Japan, Singapore and really important she met with the late Pope conferences in 2023 to discuss AI ethics. So, children of God, please with a warm welcome, welcome, please, Professor Pascal. Thank you, Daniela. Thank you, ladies and gentlemen. I'm very happy to be here. Um I um did all my studies in the US and uh spent 20 years as professor in Asia. Last year uh I moved to France Paris fair in France and I strongly agree with Peter Yonas earlier uh narrative on we need European valued uh centered human value centered AI development and research.

I also strongly agree that um here we have opportunity to provide uh the next frontier of AI research. So fair is the research arm of meta and I'm here to tell you the good news from fair in meta. Can I Yes. So we all agree and the previous panelists have talked about the need for AI economic development the future of precision medicine to make us healthier and to educate our uh next generation better and uh to have uh a to use AI for economic development. That's not the only thing I believe AI is creating is enabling us to have another time in history, another ren renaissance in humanities, in the arts and in science. AI will be used, it's already being used by artists to create new forms of art, by uh writers and uh creative uh creatives to create new forms of humanistic studies.

And there are people working on uh ethics of AI. And when we talk about ethics of AI, we're actually debating the ethics of different cultures, human societies. And AI of course is being used, starting to be used in discovering new scientific theories. And so I believe AI will bring about another renaissance in the history of humanity. And of course, Europe was the bad of renaissance the last time around. And I think we can be again the um you know uh a very important critical part of this um the second renaissance enabled by AI. So Meta has long been an industrial leader. used to be called Facebook and uh we started doing AI back in 20ou in the 2000s already using AI to enable recommendation engines and ads and so on.

And then fair was created about 10 years ago uh by a European um French American researcher named Yan Lung. Actually last time I came here on this stage was with him at another event in 2017. So in fair and in meta we believe in open science and open research. Uh what's why do we believe in that? Because unlike many other technological innovations in in our human history AI really is a full stack uh it's a multistakeholder co-development technology. It came out of the fruit of scientists in different disciplines. uh my first published paper was here in Europe in in in the 90s. So AI came from the field of seno processing, machine learning, uh mathematics, statics, statistics, pattern recognition but also cognitive science also computational linguistics also from uh uh philosophers.

So AI has always been a uh multistakeholder development. So we need to have a uh open science approach to continue to have this open science approach so that we can co-develop AI with the entire society. Every one of you here has a role to play and you are already play significant roles in the development in shaping the future development of AI. So to enable that we need to open source what we do our our scientific models and so on. Another reason to enable uh to to for AI development is I can stand up here and tell you llama is the best this our other model is the best in that but only through open source can you actually verify that claim only through open source can the scientific communities collectively benchmark what we do.

Right? So people back in 2023 people were saying how can what can chip beet do and not do and I actually uh decided to with my research team do a benchmarking paper that benchmark chibbt on 23 standard natural language processing uh uh open source natural language processing tasks and benchmarks. So only through open source can you can we collectively verify the scientific claims that's made by various uh uh players. So that's a game of science open science open source. Now this is uh about uh about I already talked about the history of AI a little bit. So we founded fair uh young lung co-launched the fair labs in 2013. FIA stands for fundamental AI research and uh the role of fair in meta is we are the upstream innovation center right the um you see the products in meta we we have today we have fair and then back in uh 2016 PyTorch was created and I think many of you are familiar with this it's the open-source platform for development and research of AI

today open AI uses it Google uses So that's the downstream from our fundamental AI research and 2020 2022 around the time we had our first llama model and did you know that the first llama was developed in Paris by European researchers in fair right so fair is a international research lab even meta is international uh innovation center because we have many researchers from all around the world they can be sitting in Melo Park or New York and a lot of us are in Paris and some in uh other parts of Europe in in in in Switzerland in UK uh you can debate whether that's Europe I think it's Europe um so Llama came out of Paris fair Paris and Llama then went uh was uh spin off into organization called Genai which today is the llama team so they develop lama so what do we do in fair we continue to do fundamental AI research.

We look at the technology of tomorrow. We are already looking at beyond lama uh beyond genai in the future. So there are researchers uh I will go to the next slide. So we have different research teams in fair working on the building blocks of AGI what we called ammy. The mission of fair is to uh is to build the uh what we call advanced machine intelligence. So it's AGI with a meta take. So different army capabilities and you can see some of this here. And I want to mention something uh I was very happy to hear the mention of low resource language development because I personally work on low resource language development since the 1990s first in speech then in languages.

We have six to 7,000 living languages in the world today. and the major language models indeed perform much better in the major uh languages uh and not so well in the low resource languages. So in fair we uh in a few uh previous years have wrote out uh a product called NNLB no language left behind which worked on more than 300 languages in the world and they were benchmarked on 300 languages in the world they're open source you can evaluate its capability right some of this is in llama some of it not yet but we collaborate for example with UNESCO to open source this kind of a uh language translation for uh for more than 300 languages in the world.

And I was very touched to hear at the Paris action summit uh many um developers from the global south in in Africa in South America talking about how they use these tools to develop applications in their local language and today in fair the same group is working on towards 1,000 language full machine translation for 1,000 languages in the world. So they are working on the next technology for low resource language translation with very few samples. So today's generative models are not good at um learning things from very few samples. But for low resource language translation it's a very good use uh it's a very good um benchmark of the future architecture that we are developing exploring in fair.

For example, you must have heard about the world modeling uh and the Japa models that Jan has been talking about. So we are exploring world modeling in fair collectively with different hypothesis. Some pursuing the generative hypothesis, some the JAPA hypothesis, some pursuing the reinforcement learning hypothesis. We would like models to be able to understand the physical world around us and also the mental world of humans of users. So this research is just the beginning. I'm very excited uh being in fair at the forefront of exploring this line of research. And another uh topic I'm going to talk about I'm going to focus on embodiment. So that's one of the building blocks of army or AGI which is um embodied AI agents and I'm going to talk to you about this and other you know we have a very big group effort on alignment.

What is alignment? It's about responsible AI how to align our AI models to human interests and human preferences and different cultural and philosophical contexts. So this is a rigorous line of research, right? Uh we don't just look at um uh uh regulations and guidelines. We must translate that into actual codes. How do we operationalize that into actual codes into machine learning? So there are group of people working on alignment. We also have people working on um I mentioned uh already uh communication language people of course working on reasoning and planning the four mentioned world modeling that's the focus of reasoning and planning how can AI models reason and plan better if you just use generative models today they are optimized to predict the next word and the next pixel in an image they're not optimized to actually reason and actually plan so we are exploring ing the architecture that will allow us to create what we call objectivedriven AI.

All right, generative AI is very powerful and the fact that they are only trained to predict the next token, next pixel and yet to yet able to manifest all this kind of capabilities and cap capabilities and task completion is a is is a great discovery. fair actually worked on the predecessor of tragic uh such as we had this blenderbot we had you know we had um uh a system called galactica we were already doing uh collectively I don't I don't mean just fair but also of course open AI at the time Google was doing the same line of research transformer came out of Google which was based on pietorch and open did a chat GBT based on transformer and fair you know it's it's it's always been this collaborative collaborative work.

So generative models has its great capabilities but it has it it has its limitations right because they are not optimized for reasoning or for planning. So today if you have if you've been using this kind of gen AI models you must have the feeling that it doesn't always do what you want to do even when you say something and it says yes I will compress the data that like you said but it's it's saying yes it's doing something else why because it's trying to optimize for optimize for multiple objectives the objective of answering user requests the objective of speaking fluently that's what chatb and all these genai models love to do. They want to sound professional.

They want to sound expert-like and the objective of preventing harm and the objective of doing many tasks at the same time. So, Gen AI today that's why you need fine-tuning for specific languages. You need fine-tuning for enterprise applications. So, please do fine-tuning because llama is there free for everybody to use as a foundational model and to fine-tune for all these different purposes right now. But we are looking at different architecture that will be more efficient, that will be smaller and they will be objective driven. The the next part of the talk I'm going to focus on embodiment. Um people are you're probably familiar with all these AI assistants. I've been working on conversational AI since the 90s, right?

At the beginning, we have AI assistants that handle call center requests on the telephone. you call and they say, you know, AT&T for example at the time will have a service that's how may I help you and then you some of you old enough you remember those AI assistants on in call centers and then in the 2000s we saw series and the other assistants come out on smartphones and uh you talk to it and it it does different things. These assistants were kind of errorprone but also it was very simple. It does one thing at a time. very simple things like five me restaurant give you a whole bunch of search right it was voice search but today we have evolved from AI assistance to agents 2025 I actually call it the year zero of AI agents right because these AI agents are kind of learning so what's the difference from between AI agents AI assistants AI agents are more autonomous than assistants they can carry out tasks at a higher level right it can

carry out tasks that will require more than 1000 steps of humans to do. Right? Search is just one step. Replace search by assistant, you just replace it by voice command. But if I want an AI agent to schedule my meeting, I just want to say please schedule the next meeting between me and Mark Zuckerberg. I don't do that but very often. But you know there are a lot of complexities. Imagine if Pascal wants to have a meeting with Mark Zuckerberg. lot of capacities it needs to look into what's the purpose the a means the scheduling and what is the good contacts and so on right agents should be able to do that AI assistants cannot so we're working towards this kind of agents and today the agents are you see demos um and on the other hand you can build your own AI agents I always tell startups you can build your own AI agents by fine-tuning instruction tuning on llama models that's advantage of open source foundational models.

You can take them and you build your own agents, purpose-built AI agents. Right now, what I want to talk to you about is beyond that, beyond what I call web agents or digital agents, Lama agents or what you see out of, you know, claw agents, right? What you what they can do today is you use tools on the internet to complete tasks for you. And we also have a benchmark for these agents called GAA by the way. And people can you know compare different uh web agents in terms of their capabilities and two use. I also want to talk to you about embody AI agents. They are embodied in avatars uh 2D 3D avatars in meta.

We have other organizations reality labs developing these 2D 3D avatars. So AI agents embodying them. AI agents embody the in wearable devices. Why do I call them embody? They actually embody in the human user because on the AI agents on the wearable devices, they will see what I see, hear what I hear, right? So this is uh the second the this is what uh embodied AI agents are coming in the future and you must have seen the meta device called Orion under development. So in a few years you imagine that you won't be using your smartphone anymore. You will be wearing this holographic uh this kind of uh wearables with holographic displays. You will see your computer everything just like this.

You can be sitting here pretending to listen to me and working on your computer with your you know with your fingers like doing this and nobody would know. You can also uh listen to me and check whether I'm saying something is true immediately in your devices. Right? It's all holographic. Well, the idea is to break down the barrier, right, between human to human because the smartphone is I see people in restaurants looking at their smartphones, right? You want people to actually connect with each other more with the added benefit of technology. So they will be in wearable glasses where you will get the information you need without putting a barrier between you and your interlocutor, your friend, your colleague, your client, your family, you'll be continually interacting with them and while being augmented by technology truly technology augmenting human and not putting a barrier between us and then we go to robotic agents.

So today I'm very excited. Since I was seven, I wanted to build robots. I went into AI because I wanted to build robots because robotic development was too slow. You know the mechanical part of robots development has been very slow. So I went into AI to build the brain for robots. But today we finally see the merge of the two field. The merge of robotics with AI into something we call robotic AI agents. Right? When robots are equipped with brain of AI brain, they become much more intelligent. They can do all the things that I imagined that we wanted to do when I was seven in the science fictions. Right? So what are the capability we built?

Sorry. Um so wearable and virtual agents. What do these agents need? The capabilities they need number one physical modeling. when they are in they are agents with my egocentric point of view they need to know what I see what I'm experiencing the laws of physics generative AI cannot do that today very badly they need to know the laws of physics what's happening around you what is about to happen if you cook know if you put something too too hot you might burn yourself it needs to know that at the same time as you're doing this kind of human actions in the physical It needs to know the human mental world. I want my daughter to ask me, I want to have a wearable agent who is my math tutor.

It just looks at me when I'm doing my math homework, but it does not tell me the next step solution. It knows what I know and try to hint as a good math tutor should do. Today, we don't have that. We have uh uh we have uh generative models trying to solve math problems, but not tutoring. So coaching tutoring requires these agents to understand the physical world and the human mental world going into robotics even more so they need to understand the physical world right the laws of physics they need to control the actions of robots today one of the biggest problem in generative models which I've been working on a uh many years is hallucinations all right hallucination is AI giving you uh untruthful answers is okay if it's a text answer you don't like the answer you go for search maybe you can solve the problem imagine robot hallucinating you know fllaying its arm around your kitchen you don't want that right you don't want the wearable agent tell you to jump when

you shouldn't jump right so you don't want that so hallucination cannot we cannot have models hallucinating that's another reason we work on architectures such as ja because we want architecture that's fundamentally objective driven fundamentally built to have guardrails and not as like a filter at the end. Generative models cannot do that. So in robotic agents we we want to build uh we are actually coming out already publishing some first results on robotic agents robotic action planning using jetpack. You can Google Google the paper. It's all open source, open science. You can verify that. You can take that home and try it yourself and build your robotic agent based on that, right? And don't forget, of course, there are a lot of ethical issues about agents and then about the AI agents.

Agents are more autonomous than LLMs. Embodied agents are autonomous AI agents in the physical world. So the implications and the danger um so the risk can be um uh detrimental more detrimental than if it's just a web output right so we have a large team of researchers myself included uh we're working on models that are fundamentally taking AI safety into account right I don't want to give up geni uh generative models I use generative models and improve them and then use them as building blocks towards objective ive objective and safe AI. So that is that is a one part or important part of all AI research today and with that kind of research we can ensure that embodied AI agents will unlock human potential.

Imagine the applications we can have with this kind of agents. Nobody's touching it. Nobody's nobody's even talking about it except me and some of us. This is my first public talk on this topic. Before I've only talked internally, right, about since a year ago and now we have some initial results and I'm very excited to see because as a scientist I want to have wearable agents that can tell me whether how to set up my experiments, right? Chemist who can have wearable agents that tell them precisely what portion uh of uh of uh of chemistry solutions to use and so on. So imagine the human potential that can be unlocked by this kind of AI future AI models and imagine the renaissance we can have that will elevate humanity.

Technology is not there to prevent humanity from progressing. Technology is here to enable us to progress further. Not just technologically, not just economically but also humanistically. Right? uh elevates our thinking, elevate our civilization and so on. Now to end this talk, I have two minutes. Um perfect. I'm gonna show you a video illustration of what I see as family of embodied AI agents working do together. So I'm going to use I'm from France, so I'm using a very bourgeoa example of a dinner party, which I love to have in my house. When I move to Paris, I pick apartment where I can have a huge space for dinner parties. And so for dinner party, how do I plan a dinner party?

Before I plan it, I I you know, imagine this is the future. Not today, right? Next year maybe, two years later, I have my Orion glasses and I have my web agents. I say, "Okay, I want to have a theme for my dinner party. Okay, what's this theme?" Okay. And uh I start to interact with the online agent and say, "Okay, what what could be the uh invite invitation list? What kind of people I should get, you know, get together at this dinner?" some philosophers, some engineers and some writers and maybe some filmmakers which I happen to know in Paris and uh let's have this theme and suggest some menu right what are these menu suggestions okay thank you I take this menu and I go grocery shopping I love to going grocery shopping in Paris because I love touching the product produce so can you help me figure out what to make for dinner tonight hopefully something where I can use some of these fresh veggies and stuff I look at different produce And

I have my Metag I have my Metag Glass Orion telling me here these are the produce you need for your menu uh for your for your dinner menu. But then I'm like you know these these radishes are not so good-looking. I want to change the menu. Okay, it will change for me. I say okay you can replace this by this while staying in the theme of my dinner party. So I go shopping and I get all this and I have them delivered through e-ervice. maybe Amazon delivery, maybe others. And uh to my home and I go home, I prepare and I start cooking and I will be cooking with my uh is it I will be cooking wearing my Orion glasses and I already have the menu and I'm actually a pretty good cook.

I'm not like zero, you know. I start cooking but then it will tell me oh but this is a new dish I want to try instead of you know now when I do that I look at the menu I do it I'm not so sure right so the glasses the agents will follow me it knows the plan and as it watches me cook it will it will hint to me when necessary only if not it's it doesn't say anything what's the next step right um so no back. I cannot go back. Um anyway, so it will hint and uh I might, you know, suddenly I put things in the oven and I have a phone call.

So I went to get a phone call talking to my friends telling them how to get to my home and then the glasses will say, "Hey, don't forget it's five minutes. You need to turn off the oven." Right? So this is uh how wearable agent um wearable agent helps me. Then my guest come for dinner. My dream is to have my robot spot. serve drinks to my friends when they arrive. Go to the kitchen, prepare the drinks and serve them. So I can just talk to my friends instead of going between the kitchen to my friends and then my friends are all sitting in my in my dining room and they say, "Pascal, come come join us."

I said, "No, I got to make drinks. I got to cook, right? I want a robot to help me doing that." That's the whole idea of embody AI agents. They need to take away the chores, physical chores and mental chores that we don't want to do to allow us to spend more time with our friends and families. So, okay, then we have our dinner. You know what I do not like to do in the dinner party is after the dinner party when I say goodbye to my friends? Cleaning up. All right. I want a robotic agent to clean up. Today, I have a robot vacuum cleaner that vacuums the whole floor. Great. But who's cleaning the table?

Who's bringing the dishes to the dishwasher, right? who is uh putting the things from the dishwasher back to the you know the racks and so on. I want the robot that would do all that and that's what fair robotic team is focusing on domestic robots. So I want cleaning robots. I want ironing robots right and cooking assistant or sue chef let's call it. So this is the vision, the future vision of a family of AI agents working together to improve my life and your life and there will be many other beneficial applications come out of this and we're working towards AMI in fair uh in in fair meta. Thank you very much. Thank you so much Pascal.

Thank you. If before you leave, just one question. Uh when do you expect do you have your next dinner party with your robotic dog? Robotic dog. It's uh um maybe two, three years. Okay. Or five. Yeah. Okay. Depending what a dog can do. You all invited by the way because no cleaning is necessary then. And um we can have the party here. Yeah. Yeah. Perhaps that that will that would be an idea. And if if you may if I may ask um could you tell us perhaps just in a in a few words your meeting with the late Pope Francis? How would how does it an interesting experience? Was it more a ceremony or it was actually he gave a very important speech about the benefit of uh how can we make technology to benefit humankind in particular in healthcare and medical care.

In particular he talked about this you can look it on the internet. Um his speech contains the line that for medical research, for AI, precision medicine research, we will be selfish if we would do not want to share our data and yet want to benefit from the treatment of the latest technology. Meaning that I don't want to share my patient data, but I want to use that technology to treat my illness, the technology that has been trained on other people's data. So he said that would be very selfish. That's what he said. Okay. And we talked about the uh various aspects of beneficial AI and how we can enable, you know, use AI to enable a healthier and more prosperous humanity in the future.

Ladies and gentlemen, Hask. Thank you. Thank you. Thank you.

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.