- il y a 2 semaines
Navigating the Next Frontier in AI
Catégorie
🤖
TechnologieTranscription
00:00Thank you. I'm Melissa Haikula, AI correspondent at the Financial Times,
00:03and I'm joined by the wonderful Rohit Prasad, who leads Amazon's AGI Lab.
00:09Now, you are leading one of the most exciting fields in AI research
00:14at one of the biggest companies in the world at a time when AI is this incredible flux.
00:19How would you describe this moment?
00:21Yeah. First, great to be here. It's my first time at Weaver Tech.
00:25Thank you for coming. I think we are in a profound inflection point in AI and technology as a whole,
00:33where for the first time, we have with us a thinking partner,
00:41which is not just taking, responding to your commands, but it's actually understanding your intent
00:49and then accomplishing with you some tasks, either solving some problem or helping you create.
00:57And the way I think about it is that we have gone through eras of computing.
01:02Those of you who are old enough will remember the days of mainframe.
01:06Then it came to the personal computer, then the cloud computing, personal devices in your pocket.
01:12I think now we are living in the era of collaborative computing because this AI is now your thinking partner.
01:23And what this does, I think, is that this is the beginning of a new AI revolution
01:29where industrial revolution changed everything we know about human strength.
01:36It multiplied human strength.
01:37And now the AI revolution will be a force multiplier for human intelligence.
01:43Fantastic. Now, you research AGI, artificial general intelligence.
01:48Now, that's a bit of a buzzword. How do you define it?
01:52Yeah. In AGI, of course, everyone has a different interpretation.
01:57We have a very pragmatic definition inside Amazon.
02:00The way we think about AGI is that you have an AI that can do what you and I can
02:07do through a computer.
02:09What does that mean?
02:10That I can actually have an AI that can write software for me.
02:14I can have an AI that deploys an app for me.
02:18I can have an AI that goes and shop for me on Amazon.
02:22I can have an AI that can handle my meetings.
02:25I can have AI, which I actually used yesterday when I was traveling to make sure I can plan my
02:33trip to Paris.
02:34So that is what we are dealing with here, which is an AGI that is essentially enabling what you and
02:41I can do through a computer.
02:42And what does this mean?
02:44That you essentially have a new way of taking your ideas into production much faster.
02:53And the way to think about it is I think of that this AI is making every dreamer here into
02:59a builder.
03:00Each one of us in this room have had an idea.
03:04And usually when you had an idea, you said, oh, it's going to take too long to get it to
03:09be building it.
03:09I need to hire these many people.
03:11I'm going to go and build this app.
03:13And we have all abandoned our ideas and passion because it was too hard.
03:17And now suddenly with this AI, the way the AGI will essentially enable more of is that you can very
03:25quickly go to market with your AI and then with your idea and then actually generate a lot of value.
03:31And that's the utopian world I think AGI will deliver.
03:35That sounds amazing.
03:36But how far away are we from that?
03:38When are we going to get this amazing technology?
03:41I think with any prediction, especially around AGI, I can tell you any prediction will be wrong.
03:48So I do think, though, it's not really one misconception that we have is that AGI is this absolute bar.
03:56I don't think it's going to be an absolute bar.
03:59It's going to be incremental progress.
04:01Some days the progress will be super fast.
04:03Some error, the progress will be slow.
04:06And that's okay.
04:06But I think it's really the journey that's important where we are on the path to have the most trusted
04:13AI that can have maximum utility and essentially be an amplifier of human intelligence.
04:21And how do we get there?
04:22What do you think are the most promising ways to do that?
04:26I mean, some people in Silicon Valley say language models.
04:28Some people say absolutely not language models.
04:32Where do you stand?
04:34I think, again, the whole AI history has been proof of that there's always a paradigm change.
04:42Every five years or so, there's a massive change in the techniques.
04:46And then you see some slowdown, but then we find another way to make breakthroughs.
04:51So the current paradigm of large models, especially that are self-supervised trained and then reinforcement learning to make them
04:58into a reasoning, thinking, planning models, is a quite powerful one.
05:02I still think we have a lot of runway for that.
05:05But I also believe that there will be a lot more emphasis on learning the axioms of the world or
05:10what is called a world model to make these models be more reliable.
05:14But I think we're on the right trajectory here.
05:16You mentioned world models.
05:18Can you tell us a bit about what you mean by that and how they will change things?
05:22The world model is essentially about understanding, thinking about the physics of the world as well.
05:27So if you think about when I said one pragmatic definition for AGI is that it can help you do
05:34things with the computer, what you and I can do.
05:36Well, it turns out that you and I can control a robot through a computer.
05:42Now, does that mean that the AI needs to understand the world to navigate a physical robot?
05:50I think the answer is yes.
05:52And for that learning that one, you do need the physics of the world to be understood like here's what
05:58if I go into this path, then I'll get obstructed and so forth.
06:01So these are the kind of things that you have to learn the world pragmatics and learn them efficiently versus
06:07just relying on parametric knowledge.
06:10And concretely, if we do get AGI, how will that change our lives?
06:16You know, can you, you know, say I'm moving house and I need to sort out my life and logistics?
06:24How can this technology help me?
06:26Yeah.
06:27Let's take you're a journalist.
06:29My daughter is an aspiring journalist.
06:32Oh, excellent.
06:33Great choice.
06:34Learn French as well.
06:35And I think and I've asked her this question in terms of like and I think right now already this
06:41is true.
06:41Every morning we wake up in my household and for her, she actually has her Alexa set to French because
06:48she loves French that much, even though I live in the U.S.
06:51And in the in that setup, I think I really believe that your day begins with Alexa and the new
07:00reinvented Alexa, which we call Alexa Plus, where you get a morning briefing.
07:04Like here's how the weather is going to be so that you can decide what to wear.
07:08Here's how if you take your life, like you have to go and meet some person.
07:13So it tells you what your schedule is for the day.
07:17Now, if you were doing researching, let's say you were talking to me, you want to know everything about what
07:22I've done.
07:23And it takes a lot of time to do that research.
07:25But now with AI, you can go and do a deep research on essentially everything that has been happening in
07:30the field and what my viewpoints are.
07:32So you're better prepped as a journalist as well.
07:34Now, let's go to your other aspect, like you move to a new town.
07:38Then you, of course, want to know what are the service providers around you in terms of your Internet capabilities,
07:44your utility providers.
07:45Imagine how much work that is.
07:47And I really believe that the AI will take that complexity away from your thinking so that you can essentially
07:54spend more time on creativity.
07:56So if you wanted to, it can also be your interior designer, give you ideas based on putting, if you
08:03click at our Amazon Nova family of models, you have these creative models that can actually decide, can help you
08:10with images and videos.
08:11So you can actually have and you can go to our booth and think about try-ons, right?
08:16So you can put a particular furniture on an equipment in a particular part of your home.
08:21So these are the kind of things that will happen.
08:24And then, of course, I want to maximize time as AI with humans we love as well.
08:30So I really believe the AI should be there where you need it and recede in the background when you
08:34know it.
08:34And when you're in the evening, the AI is suggesting what to watch with your family and so forth.
08:39So I think the AI, again, as a digital partner for you, it's not about automation.
08:45It's about unleashing more creativity in you.
08:48What main challenges to get there?
08:51Like, what do you still need to crack?
08:53Yeah, before we get into the challenges, it's important to think about what's the next frontier in terms of how
09:00the world we live in will change.
09:02So I believe that the next era, which has already started, is where we are going to see millions of
09:12agents.
09:13This is how we assure the collaborative computing into this environment.
09:18Millions of agents that are built by a lot of different enterprises, different service providers, startups.
09:24And then you will have your personal agent as well.
09:28Alexa, for me, is my personal agent.
09:30And when you live in that world, you have to start thinking about what is the digital highway to be
09:38built.
09:38Just like we build the autonomous cars and we are thinking about how they operate, you have to now think
09:44about what does the digital version of the Internet of Agents looks like, which means you have to have massive
09:51infrastructure built.
09:52There was, you know, in the days when I worked for my former company, which was part of building the
09:57Internet, was essentially a lot of protocols came in being.
10:01But this AI, essentially, this AI future requires us building all the key components for agents to interoperate.
10:09And the challenges will come down to the following.
10:12First, every individual agent has to be more reliable because agents, by definition, do tasks on your behalf or another
10:22system's behalf.
10:23And they automate end-to-end workflows, which means the reliability should not be in the 70% to 80%.
10:29It has to be a high 90s.
10:31So the key challenge for there to make the reliability be that high, which is not just about accuracy, but
10:38also predictability and consistency, is going to require a lot of invention on reducing hallucinations in AI.
10:46And we already do work on that by working on retrieval augmented generation, where the AI retrieves some current information
10:53or some proprietary data you have and then gives an answer, or that it can know how to call your
11:01tools that you use on a daily basis in your environment, your CRM tools, your ERP and all.
11:06So all that integration, the model has to be more knowledgeable about getting there.
11:12So I think working on reliability is going to be the first key requirement to make millions of agents operate
11:19in the world.
11:20Yeah, I was hoping you could expand on that.
11:22Because language models today, they make stuff up, right?
11:25We all know that.
11:26And they're really easy to hack, and they have a tendency to persuade or manipulate people.
11:31How are you thinking about these kinds of harms, you know, especially when you roll this out and master millions
11:37of people?
11:38As we look at the agentic future, where we have more autonomous agents or agents that behave like us, I
11:46think we have to hold the AI to the same bar we attribute humans to.
11:50In terms of that, I think if I trust you as a human, I want you to be competent at
11:57what you do.
11:57I want you to be reliable.
11:59I want you to be discreet.
12:01If I share something with you and tell you, please don't share with someone else, you should respect that.
12:07I should, transparency on your decision making is important.
12:10If I make a decision, I should explain to you why I'm coming to that decision.
12:15That's the hallmark of great leaders, trusted leaders in that sense.
12:18Which also brings to one something very important, which is very important in this era of generative AI, is that
12:25if I don't know any something, and you ask me the question, as a trusted human, I would expect the
12:31human to be saying, I don't know the answer to that, but let's figure it out.
12:35So, same thing with generative AI and the future of AI, as we look at it in terms of the
12:40challenge, is to make sure we are working hard on saying, I don't know, but let's figure out the solution.
12:48So, how do you then learn that this concept is alien to me, and then work together with the human
12:55and other agents to actually have that particular agent improve over time?
13:00So, in my mind, this element of trust comes down to knowing what you don't know, and then being able
13:08to learn and reinforce that.
13:11And how do you think this technology is going to change our lives, but our sort of relationship to the
13:16technology, right?
13:17Maybe some of you in the audience saw that OpenAI had to pull back a model because it was way
13:23too sycophantic or complimentary, to the point where people found it really annoying.
13:28And people are often saying, you know, they externalize their thinking to chatbots.
13:33They're so reliant on chatbots that they grow addicted, which are really real risks in the age of super intelligent
13:40AI systems.
13:42Unpack that for me.
13:43How are you thinking about it and avoiding these kind of scenarios?
13:47Yeah, again, the AI, especially with this kind of a competent AI that we want as our digital partner, the
13:54bar, again, rises.
13:55And the bar rises in the sense that it needs to be aligned with human values.
14:01It should be able to learn and say, here's, just take this example of when you have to repeat some
14:09information to, and let's say you're just giving your name to a speech recognition-based system, and it misrecognizes you.
14:17So then you want, as a human, what you would say, did you mean a P or a B, right?
14:23So this kind of an intelligent way of interacting is important because you don't want to repeat the same information.
14:30You don't want to reinforce your errors.
14:32So this is where we have to work a lot on making the AI be better at reinforcement learning as
14:38well, in terms of how it reinforces the right behaviors and does not repeat the same mistake.
14:44And this requires a more massive alignment strategy, which is what we are working on.
14:51And what does that strategy entail?
14:53How do you make sure it aligns with human values?
14:56The way that is, again, you have to have pillars of trust, which I talked about.
15:01And, of course, reliability is key.
15:04So you start with making sure that the models are very good at orchestrating a lot of tools reliably.
15:11In terms of the personality, it's very important to exhibit the attributes of trust.
15:15And then you have to watch out for that it's not being trained by just a few humans versus more
15:22diversity of opinions.
15:23And yet it's able to then take the diversity of opinions and synthesize an answer which is factually correct and
15:30exhibits the personality that you want.
15:33And Amazon is a global company, right?
15:37But I think the problem with current AI is that it's often very American.
15:42How do you make sure that this technology is inclusive and takes into account, say, French or Chinese or all
15:49the other languages of the world?
15:51Yeah, I think in terms of the world we live in, let's just look at this room.
15:55It's very multicultural, multilingual.
15:58As Amazon, it's paramount for us to have our technology be working across the globe.
16:03And with the Amazon Nova Foundation model family, we took extreme care in terms of making sure it knows about
16:11a lot of different languages a priori.
16:14And then also, I strongly believe that it needs to be customized in different environments because we can't be an
16:22expert in every environment.
16:24So if you are in the Middle East, then you have a different cultural context.
16:27If you're in India, you have a different cultural context.
16:29If you're in France, it's a very different context than in the U.S.
16:34And I think giving the right set of tools with our partners and with our customers is key because you
16:40want the intensity of voice in every locale you're in.
16:45And is the plan for you guys to build these models yourself?
16:48I mean, you have a massive deal with Anthropic, right?
16:50One of the leading model developers in the world.
16:53How do you work with them?
16:54And is this goal still to build your own models or maybe rely on them a bit?
17:00Or how does that work?
17:01Yeah.
17:02First, we are thankful for the partnership we have with Anthropic.
17:05We built Amazon Bedrock, which offers the widest selection of foundation models in the world.
17:13And because it's built, Amazon and Amazon Bedrock specifically, is built with the premise of selection matters.
17:21Selection matters, why?
17:22Because there is no one model that does everything the way you talk, the way our customers need it.
17:28We just talked about languages and cultural context.
17:32Not every model exhibits the same things that you want in a different cultural context as well.
17:37Same thing we were noticing inside Amazon when we were looking at different applications being built because generative AI is
17:44virtually being used in every part of our businesses.
17:48And when we looked at, like, for instance, our advertising business had a very different requirement.
17:53They wanted video generation models that could actually be used for advertising products, which means your own product from our
18:03advertisers should be embedded in a short video.
18:05So there was no model that did that really well.
18:08So we built Nova Reel, which is our video generation model for essentially serving our internal customers.
18:14And when we built that one, we realized the same capability is being required by other external customers as well.
18:21And that's why we have Nova as an important selection choice for our customers, so that they can use it
18:28on the way they want different models to be.
18:31Because this is where we are living in the multi-model environment, and we want our customers to be able
18:36to plug and play easily, whichever model works great for their environment.
18:41So that's the selection part of it.
18:43The second part, for those of you who are looking to scale AI in your applications, at scale, the costs
18:51of inference become prohibitive, which means price performance becomes really important.
18:57So to drive costs down, we felt that with our models, we can be incredibly different for our customers, where
19:07we can give the differentiated value that we can make the models less expensive for them to deploy at scale.
19:13And Nova models, when we launched them, our first generation of the models, they were 75% less expensive than
19:18any other model in Bedrock in the same intelligence class.
19:21And that, customers noticed, we have thousands of enterprise customers now deploying Nova, and one of the reasons that resonates
19:29very well is exactly that price performance, that you can actually deploy these models at much cheaper cost.
19:35So it sounds like the future is not one big AGI model that's very expensive and cumbersome to run, but
19:42lots of different small ones that maybe work together then, or what do you say?
19:48Yes, that's a very, it's a very good question, because one thing about AGI, to me, is definitely we will
19:54have a more general purpose model.
19:57And I think general purpose is also open to interpretation.
20:00They are really multipurpose, they are out of the box, good at multiple things.
20:04However, it's very much like us going to universities and learning many different skills.
20:10And then we go specialize in every domain as a student, and then essentially you became an excellent journalist, someone
20:17else has become an excellent scientist, someone has become the world-class mathematician.
20:21I do think that we'll be living in this era where there will be a more general purpose or multipurpose
20:28model, but that's spawning many specialist models or specialist agents, which are very skilled at their domain.
20:36They can run very efficiently in that domain, they can also spawn a personal agent for myself.
20:43I want my knower, right?
20:45So that's the way we can think of, like, in terms of the future, is really going to be many
20:51millions and millions of agents.
20:52Some are just your own, some are working at an enterprise scale for your own enterprise, and some are these
20:58massive consumer agents like Alexa, which can help every customer in the world.
21:04So that's the world I think we're going to live in, and there is a place for very specialized agents.
21:09That sounds really exciting, but what do the next six months look like?
21:13This sounds like a long-term plan.
21:16What excites you in the next six months?
21:18Yes, six months means I have to go and work again from the fun conversation we are having here.
21:23The next six months, I think we're going to see, again, sustainable improvements in model capabilities, in their deployments, agents
21:34that are heading to more reliability.
21:35We have a research preview of NOAA Act, which is a software development kit we provided developers with, which is
21:44essentially trying to make web agents, as in agents that can go and interact with websites and take actions, be
21:51more reliable.
21:51So I think while we are looking at this massive number of capabilities coming in the model, the basic atomic
21:58capabilities of can you search, can you click a button reliably, I know it sounds a bit boring, but that's
22:04important to increase the reliability of it.
22:06So what I think is going to happen is that the model capabilities will keep getting better, and the reliability
22:11will also keep getting better as we go, ultimately to make that future of agentic AI really happen.
22:18What I really appreciate about this conversation is how you've really grounded this technology in real-life, concrete, practical applications,
22:25which is not what a lot of people in tech do.
22:27And AGI often has this, it's often called excessive hype or a buzzword.
22:34What do you say to people who claim this term is science fiction or nonsense?
22:39Yeah, I would say for this room, I would focus on essentially don't get sucked into the hype and try
22:47to do everything overnight.
22:49That's not going to be the approach here.
22:51I think go and look at, again, what matters to your customers, and that starts with use cases.
22:58Bring many use cases, start prototyping them very quickly, but be patient.
23:04I think it's very important to start with AI.
23:07I think one thing is going to be true, which has been true with every technological evolution, that the enterprises,
23:15the businesses, the startups that thrive are the ones who adopt the technology.
23:21So AI is the tool that I think, whether or not we get into AGI or not, on what timeline,
23:27that's not important.
23:27It's important that you have, for the first time, an AI tool that is so powerful in being your thinking
23:34partner, that start adopting it, because those are the businesses that will thrive.
23:40Fantastic.
23:41Now we have one minute left.
23:42What's your advice for people?
23:43How do we take advantage of this or prepare for this in the best possible way?
23:49Yeah, I think there's multiple levels.
23:51First, if you're a business or a startup or an enterprise of any kind, I would say start with your
23:57data and decisions.
23:58Make sure you have a good logging of them so that then you can plug in your whichever AI into
24:05it so that you can now start making deploying agents in your workforce.
24:09So that is very important.
24:11The second thing on the societal side, I think it's very important for us to make the future generation of
24:18workforce grow with AI.
24:20Now, there we have a lot of tailwinds because my kids grew up with AI.
24:24They are very well equipped in using the tools.
24:26And so they, in the universities, in the schools, I think that will happen organically.
24:32However, in enterprises where we already have a large workforce, it's very important to invest in upskilling that.
24:39And I think there, again, the emphasis should be adopt AI, start experimenting with it.
24:44And just like if you had laptops and computers and you weren't using it, you won't be relevant.
24:50That's why you want to go and start using AI because that's how we'll all stay relevant.
24:55That's exactly what evolution looks like.
24:57Fantastic.
24:58Thank you so much, Reid.
24:59And thank you so much for joining us today.
25:01Thank you.
25:02Thank you.
25:03Thank you.
25:04Thank you.
Commentaires