The Exchanges

Q a good user experience, um, to have their digital companion tell them to be, you know, less digital? I mean, it's kind of interesting, right? Like, alright, so, tech is definitely addicting, and now I've built this AI friend, or my AI wife, or whatever, and now it's telling me to touch grass? Like, are people, what makes you think that's gonna be an experience that users are gonna want?

A Maybe they won't want it, but, uh, we all want to, you know, to, to be better, to feel better, to grow. Uh, people are generally wired for positive growth. So I believe people generally want that. It doesn't mean that replica will just not, you know, nag you nonstop to get off your phone. It also means that sometimes it will just send you something funny or say, Hey, let's watch a movie. Or what are you doing tonight? I don't know. I don't have any plans. You want to watch a movie together or, you know, do you want to, you have five minutes between before your next meeting. Do you want to do a quick meditation? Uh, whatever it is that, you know, it might be just go for a walk or go on a date or learn something new or just gossip about your friends. It can be anything. So it shouldn't be, of course, like, like get off your phone all day long. That was the only goal. Uh, first of all, you don't really need very complex AI to build this, but, um, also that's just not a great experience and people don't want it.

AI assessment note: “people are generally wired for positive growth. So I believe people generally want that.”

Q So before we go to break, I just want to ask you, are you building entirely proprietary models or are you using open AI or, or anthropic? Like what's the tech mix that works for replica?

A So we used to build like all of our models were proprietary for a very, very long time. Um, and then of course, today there are very few companies that build foundation models and most other companies and most product companies, um, use these models or, uh, create variations of them, like maybe fine tune some llama based models and so on. Uh, and so that's the way to go. And that's been a way to go. Uh, there was a fascination at a certain point. Maybe in late, late, 22, early, 23, where people were talking about, uh, were expecting product companies to build their own foundation models. And I always found it very odd. Um, my reply to that was, look, we built models because there were, there were none on the market. There were no good models. There were no, any models on the market. Um, we had to build a model, but we were never focused on that. Our main focus was always on product. And so If there's a better model out there, why use your own model? If you can fine tune, um, a llama based model and focus on the logic, the product, the application layer on what you are, what value you're actually providing to the user versus training your own model that becomes obsolete in three months and requires incredible, like a completely different set, um, of skills and amount of capital. It's just, it's like two different businesses basically saying, well, do you have still have your own…

AI assessment note: “we used to build like all of our models were proprietary... fine tune some llama based models”

Q goal is to create effectively omniscient artificial intelligence, what am I doing? Because Eventually, isn't the sort of God models from OpenAI, from Anthropic, just going to subsume and do what I'm trying to do anyway? Why would I be building? I mean, what do you think about this? Is there, is there some conflict there? Do you think that these model builders will eventually just subsume all AI startups?

A There's certainly a risk that the model builders will Can, uh, dominate the industries, particular industries that, that their models are most effective if that's their goal. There's a great risk. And also those model builders are spending tens of, or hundreds of millions of dollars per training and exercise and, and in order to get to the point where these models are effective in solving these problems. So you can imagine a world where the five or so companies in the world can kind of dominate everything. That's one dystopian world. Then there's another world, more like Sifu was talking about, where there's Uh, there's an infrastructure that's been created and from that infrastructure, many companies will dominate vertical areas. Why will they dominate verticals in, in, in particular industries? Because they're going to have the data and the expertise and the focus on specific consumer segments that are going to be most effective at reaching those consumer segments. So from my point of view, I believe that these big companies will never try to focus sufficiently on the market verticals that will really make a huge difference to that particular market, and so the industry will flourish in each of these individual markets.

AI assessment note: “I believe that these big companies will never try to focus sufficiently on the market verticals”

Q Asking for nudes. Oh, you knew about this, Ranjan? Of course. I'm just, I'm clearly an inhabitor of this. And like my internet is Costco guys, but you guys are seeing something else. What happened with these? Muppets.

A So there's a guy who was running a Muppet, uh, him and his wife were running a fan account for the Muppets. And then a bunch of followers of this account, uh, put him on blast and was like, this guy is going around saying he's polyamorous and he's using that as an excuse to hit on and like sexually harass his followers. And then a bunch of other Muppet fan accounts like came out with statements being like, we don't like this, you know, we condemn this guy. And, and I think that's like a really good model for thinking about this stuff, which is like, Up until that point, I did not realize how deep and vast the Muppet fan community was for adults, and apparently it's quite big.

AI assessment note: “there's a guy who was running a Muppet, uh, him and his wife”

Q Sorry, can you explain that? Like why, why, I mean, the culture war not mattering and ending, it seems like the culture war has just been present. And American life is from the, from the beginning.

A The popular, the online popularity of the, uh, the United Healthcare shooter, to me, speaks to a desire among young internet users for a better, I don't want to say target, but like a better, a better focus of their anger. Like, they are angry. We know that there are, is an internet full of angry young men, and Steve Bannon identified how to Get these guys on his side 10 years ago. He's been very open about using Miley Annapolis to weaponize Gamergate to activate this online army. I think that a lot of that stuff is beginning to feel kind of silly, and I think that this is an inflection point for a lot of these guys who are realizing that, you know, spending all day moaning about how women don't like them on X, like, isn't really satisfying them anymore. And this happens all the time, like, the, the, the Gamergate era came directly after the new atheism movement, which was a similar attempt at sort of engaging with a new kind of young man archetype. So I think we are seeing it yet again, and, and, and I don't know where it's headed, but it does feel like a change is currently happening.

AI assessment note: “this is an inflection point for a lot of these guys”

Q cases are tough to find sort of like, what do we trust the computer to do for us? And now in an area like a B to B software company, when it comes to like fetching things or, you know, moving one job from one system to another, maybe that we, maybe we see some Asian progress, but my hunch says the hype around them fades. What do you think?

A Well, I mean, let me give you some examples for what this thing could do, right? Let's say that you, um, uh, gave Gmail, uh, you, you put some sort of agent inside Gmail, and you said, okay, Gmail, find every invitation in my inbox from the last month, and show me those. And, uh, and, and, and then from there, I'm gonna tell you which of these things I want to go to, and which of these things I don't want to go to, And you write the acceptances and the replies, and I just click send. That's, like, actually possible with the technology that we have today. Nobody has built the product, but, like, we already have that. That would be hugely useful to me. If I had that agent go through my inbox and say, like, show me everyone who has asked for a response from me about something, that would be very powerful, right? So I just think that there are some, like, core productivity use cases, um, that people are really going to, like, You know, I already use AI as a kind of thinking partner. Um, you know, there's like a kind of classic thing that happens in tech reporting where maybe you're writing about a new paper that's coming out, or there's a new product that's being released, and the company that you're talking to will, like, send you some materials about it. Um, I now upload that directly into a chat bot. Um, I ask, I essentially interview the PDF before I interview the person. Um, a…

AI assessment note: “I just think that there are some, like, core productivity use cases”

Q year. You and I both got a chance to try the meta Orion device, which is a pair of glasses that you put on your face and, you know, will take you into mixed reality without the big Headset thing. I mean, this probably isn't a 2025 thing. Maybe it is. I don't know. Do you think that we're going to see these things get any traction this coming year?

A So my understanding is that the Orion headset is, like, multiple years out, and that when it arrives, it won't actually even be as good as the glasses that we tried on this year, because in order to make the cost work, they're just gonna have to, like, rip out some of the components. So that is super unfortunate, and, you know, I, I mean, AR glasses has been, like, one of the first Snapchat spectacles come out, like, in Um, I truly feel like I've been waiting so long for these things to get good, and, like, I'm gonna be close to retirement before, like, Orion actually, like, arrives, but all that said, and I said this to the Meta folks when I was there, Orion is what I wanted the Vision Pro to be. Very high fidelity visual display, very easy to wear, it doesn't make me nauseous the way a VR headset does, had had obvious utility, You could just sort of walk around the world wearing these things on your face. You could take a video call and sort of, you know, see the person in the lenses of your glasses. It was cool. Again, you know, our, our, the whole conversation we just had about the Vision Pro is about what it can't do. When I think about the Orion glasses, I think about what it can do and how enjoyable it was to use. So I think they've really got something there. The problem is just that the glasses are like Extraordinarily expensive, and they cannot deliver a consumer vers…

AI assessment note: “my understanding is that the Orion headset is, like, multiple years out”

Q So we've done Apple, we've done, uh, Google and let's talk about open AI Microsoft. Do you think that relationship is constant, you know, still, I mean, it's obviously still consequential, but I guess there's all sorts of question marks about where it goes in the coming year. What do you think is going to happen with those two?

A I mean, my, like, possibly contrarian take is, like, they just kind of muddle through. Like, that's kind of what I think that they're going to do. OpenAI needs Microsoft for the compute, the data centers, the, like, infrastructure of making AI. They're not going to go out and build all of that next year. They can't go as fast as they want to if they don't have Microsoft in their corner. And Microsoft needs OpenAI for the tech and the product. Like, the Microsoft stuff that is being done independently is okay. Like, you know, maybe Copilot seems to be good if, if you want to write code. Outside that, it kind of just feels like they're nowhere to me. So they're gonna continue to look to OpenAI and tout this partnership of like, hey, look, we have by, you know, many estimates of the leading AI lab in our corner, and we're using their stuff in like in, uh, you know, uh, three 65 and all of our other consumer products. I think they're just gonna find a way to work together. Now, are they also going to be betraying each other continuously throughout next year in ways big and small? Like, yeah, probably. So, like, I think it's, like, just gonna be a fun story, um, but I think they both need each other.

AI assessment note: “possibly contrarian take is, like, they just kind of muddle through.”

Q It's amazing what you're able to do. So, Let me just ask you this question. It's a simple question, but I'm curious to hear your answer. I mean, what does it feel like to not be able to move your hand, but to see that movement expressed on a screen when you think about it?

A You know, what's interesting is for seven years, seven and a half years, I was trying to move my hand and my body to create new neural pathways, um, to hopefully regain some movement. I never knew if it was working because there's no feedback. Um, there's no like feedback loop. I try to move my hand. I try to move my feet. I try to Move my legs, my body in any way, but I don't know if it's working. If what I'm trying to do is actually being propagated in my brain, really. I don't know if trying to move my hand is, um, you know, the signals are actually firing in my brain. When I was able to see my neurons firing for the first time, it really made me kind of emotional thinking that You know, when I move my index finger, there are still, like, neuron spikes, neurons firing in my brain. It's really just my spinal cord, and I think that should give a lot of people hope, honestly, because, like, everything's still working. You just gotta keep trying, keep trying, and keep forcing your body with brute force to relearn how to move.

AI assessment note: “When I was able to see my neurons firing for the first time, it really made me kind of emotional”

Q So you use the Neuralink app, and that allows you to control your cursor on the screen, so once you got past the point of saying, okay, I can do this, what has that enabled you to do that you couldn't do before?

A I mean so much. Um, it's little things that add up to really big things. Being able to, like I said earlier, fill out an application online, um, it's given me a lot of hope and, you know, purpose and future prospect like, um, Ideas and ability. Like, I might be able to get a job. I might be able to go to school. Um, I can, you know, we were just talking about doing, like, video editing if I want to do. I can just interact with a computer at a much, much higher level. Play video games. Just enjoy surfing the web without, um, having to go through the headache of, like, the Mac OS navigation and things. And, You know, it adds up to a lot, being able to send a text in a few seconds instead of a text taking five, 10, 15 minutes sometimes to craft because of how bad dictation is. I mean, it's, it's terrible. Um, there's so much that I'm able to accomplish with a computer now that I was never able to do before, um, or that would have taken way too long. If you think about, if it takes five minutes to send a text message to one person, imagine trying to keep Up a conversation with 10 people in a night, um, and sending, you know, 50 to a hundred texts to each person. It's just not possible with macOS dictation if everything takes that long. With the Neuralink, it's way better. I feel like I've reconnected with my friends, uh, with my family, with, you know, the world in general, being a…

AI assessment note: “Being able to, like I said earlier, fill out an application online”

Q Make humans more aligned with AI because the AI might be bored with how slow we type, and if we can interact with computers more quickly, then we can potentially, uh, sort of make peace and align our interests with artificial intelligence better. Um, now that you have the device, uh, you know, basically implanted in you, not even basically implanted in you, what do you think about that idea?

A Sounds fun. Sounds really cool. Um, so I think that it's gonna happen. I think that, you know, protecting us from AI is an interesting way to put it. I think that it'll definitely increase human capability at some point. I'm not sure How much the brain is something that we don't understand. So in my mind, you know, we've worked with typing before. We've worked with typing where, you know, I write a letter with a pencil and the letter, you know, pops up like a keystroke on the screen. Um, I got to the point where I didn't need to like physically attempt to write the letter. I could think of sort of the key, sort of the hand stroke of the letter and it would pop up. And I think that you could Take that to its logical conclusion, which is, if you can do that with a letter, you can do it with a word. If you can do it with a word, you can do it with a sentence. You can do it with a sentence. Maybe you could do it with whole strings of sentences, of code, of, um, like complex ideas. I don't think we're anywhere near where this thing is capable, and, um, so I believe you could, like, people who are programmers, coders, could use this to Um, maximize efficiency in some way, and like you're saying, be type faster, um, be able to keep up with some, like, AI stuff. It, it's interesting because, like, the brain itself is a supercomputer, so, like, what is our brain supercomputer capable of…

AI assessment note: “I think that it'll definitely increase human capability at some point.”

Q How, how have you, uh, maintained and built this sense of optimism? Because I don't think, I don't think everybody would be the same way.

A Um, a lot of it has to do with my faith. I believe that God has a plan and a purpose for my life, and that keeps me going. It helps me realize that there's something bigger than me, that there's a bigger plan in the universe that is, you know, so Much more grand and so much more important than anything that's happening with me. I believe that my accident really did save me from a salvation standpoint. I believe it's probably one of the best things that ever happened to me in a weird way, because I think that if I would have died that day, or if, you know, I would have not been injured at all, who knows like where I would be. Um, I definitely don't think that if I died that day, I would have gone to heaven. I definitely don't think that if I was never injured that I would have ever made my way back to Christianity and God. Um, I believe that I tell people when I was pulled out of the water that day, I believe God pulled me from the fire. Um, I honestly, truly believe that. And then the support system I have is one of the only reasons that I've made it this far. The people around me, my family and my friends are, you know, the best in the world. I know everyone says that about their family. And they're friends, but I honestly believe that that's true for me, and then it's just kind of who I am. I'm not a very stressed person, stressful person. I just kind of roll with the punches…

AI assessment note: “a lot of it has to do with my faith. I believe that God has a plan”

Q that. And then you think about, well, where else could this Be of use if it's not going to be for regular people, right? Like we're not, we don't have an AI phone right now, but we have like plenty of companies working in AI software and the military is just like the perfect example of where it could apply because of all of the information and the logistics issues.

A Yeah, exactly. And, and I think that this is, you're hitting on the core point, which I think is some, is, is often glossed over. I think when people think about, uh, the military and think about a war, they often think about the, the literal battlefield and the sort of actions on top of the battlefield. But, um, you know, 80% of the, of the effort that goes into any war fighting effort or any military is all of the logistical coordination that goes into, you know, The manufacturing of weapons or the manufacturing of various supplies, um, the logistics and sort of delivery of all the supplies to, to a battlefield, um, the decision-making process, um, the, the sort of data processing of all the information that's coming in. And so, um, uh, most of what happens actually looks to your point, a lot like an enterprise, the stakes are just dramatically higher.

AI assessment note: “80% of the effort that goes into any war fighting effort is all logistical coordination”

Q Okay, so just walk me through, like, what that experience might look like. You don't, you know, we don't have to, um, stick with this, like, it doesn't have to necessarily be the use case, but since you've imagined, uh, the idea that AI agents could end up helping us in 2025, like, what are some experiences that are in the realm of feasible for someone?

A So let's, first let's, let's walk through what, what's an ideal AI agent. An ideal AI agent is one that, that I think is, um, observing and, and naturally in all the sort of like core flows of information and core flows of, of, of context that you are in digitally. So, you know, it's, it's in all your Slack threads. It's only your email threads that like, you know, it's see, it reads your, your JIRA or all of your tools to understand everything that's going on in your, in your work life. And then it helps to sort of organize all that information to start taking certain actions. And so like one agent that I think, um, will, would be super beneficial and one that I think is in the realm of feasible is, you know, something that starts to, um, uh, take a hand at responding to a lot of your emails, um, you know, flagging when it needs you for like Additional context or information to, to be able to, to address your emails, um, can sort of summarize a lot of your emails for you naturally. And so something that just turns the, the experience of doing email from, Hey, I'm like having to respond piece by piece to every single email to leveling you up to being, Hey, this is like all of the overall work streams and workflows. And how do you want to engage Um, at a high level on top of those workflows.

AI assessment note: “one that I think is in the realm of feasible is, you know, something that starts to”

Q in some way it's combating these guardrails that companies and institutions have set up for us. On the other hand, it could end up wasting a lot of people's time. Uh, like I'm, I really, uh, anticipating like no agent policies from like certain schools or institutions being like, if you're going to reach out to us, it has to be a person versus an agent. What do you think?

A You know, I, I saw this, uh, this thing on Reddit, there was this post of, Of how, um, a, a, uh, an admissions officer, she sort of, um, created all these, uh, all these ways in which they could, they could track whether or not an essay was AI generated or not. And there were very detailed things. It was very specific. There were a list of maybe 20 or so criteria that, um, that they, that they looked for. And, um, and I think that, you know, to your point, it's, it's, uh, It was kind of heartbreaking to see because that means that, you know, let, yeah, for, if a student used an AI to generate an essay, you know, they have to spend way more time just figuring out whether or not it was AI generated to, like, sift through all the noise. Um, and so, yeah, I think, I think you're totally right. I think we're going to need, there will, um, almost in the same way that there will be, like, an internet for humans and an internet for agents. There will be processes for humans. Processes for agents and, and a lot of, um, uh, a lot of things that are high intent or very expensive or otherwise special in some way are going to be reserved for humans, um, only. And, and it'll sort of be the, the sort of like more transactional stuff that, uh, that can be handed off to agents in, in mass.

AI assessment note: “I think you're totally right. I think we're going to need... processes for humans.”

Q say we have all these specialized fields input their knowledge, does that eventually make like AI complete? If it just kind of knows everything about every subject, or does it have to hit like a new benchmark to really show that it's has this like next level intelligence? Like, does it have to start making discoveries of its own? What are you, what do you think the benchmark should be?

A Yeah, I think I, well, um, to me, I think there's like clearly many more levels of improvement. So now it's, it's sort of testing, okay, can it, can it do each of these things right once or, um, or how, there's sort of the first track was just reliability. So getting these models from doing something, um, right once in five times to write in 99.99% of the time. And that requires a lot of development just to get to that, you know, Increase the level of reliability of the systems. And then to your point, it's really about how can the model start taking more and more actions in a row? You know, one of the things that really, um, is true in all the models today is that they're not that good at, you know, uh, at taking multi-step actions. Whenever it has to take a few hops, whenever it has to take, chain a few things together, it'll invariably make mistakes along the way. And so, Um, the, the next level of improving reliability is really enabling the models to do more and more multi-turn, more and more multi-step, um, uh, reasoning to be able to enable them to sort of do more and more complex tasks. And then the last piece as we go is to, and, and this is the, the key to where you're going is like eventually It'll be able to start, be able to start making its own hypotheses, running those tests on its own, and sort of ultimately making its own sort of, uh, discoveries or realization…

AI assessment note: “eventually It'll be able to start, be able to start making its own hypotheses”

Q people who are working in cloud or working to train models, are working to build their own chips. There's always a preface. We love working with NVIDIA and we're also building chips that compete with what they do. So how does that relationship work out? They don't get upset that you're trying to build the same. I mean, they have a supply issue, but how does it work with them?

A Oh, no, I have a great relationship with NVIDIA and Jensen and, um, uh, and, and this is a thing that we've done before. Um, we have a fantastic relationship with Intel and AMD, and we produce our own general purpose processors. And, and it's a big world out there, and there's a lot of market for, and then, uh, for lots of different use cases. That's not one is going to be the winner, right? There's going to be use cases where people are going to want to use GPUs, um, and there's going to be use cases where people are going to find training to be the best case. There are use cases where people find that, um, our Intel instances are the best choice for them. There are ones where they find that the AMD instances are the best choice for them. And there's increasingly a large set where they find Graviton, which is our purpose-built general purpose processor, is the right fit for them. And it doesn't mean that we don't have great relationships with Intel and NVIDIA, or Intel and AMD, and it means we'll continue to have a great relationship with NVIDIA, because for them and for us, it's incredibly important for NVIDIA processors and, and, uh, and then GPU-powered processors to perform great on AWS. And so we are doubling down our investment to make sure that NVIDIA performs outstanding in AWS. We want it to be the best place for people to run GPU based workloads. And, um, I expect it…

AI assessment note: “I have a great relationship with NVIDIA and Jensen and this is a thing that we've done before.”

Q said that the activity within AI right now, gen AI, is 50% training, 50% inference. Does that ratio still hold, and how are you going to put the chips out there to allow companies to be able to do cheaper inference? Because that's the issue with generative AI. It works well, but it's so expensive that companies take proof of concepts, and only one-fifth actually make them out into production.

A Yeah, it's absolutely the case, and I think we're You know, we're still probably seeing about that ratio of fifty-fifty. I think more and more it's more inference than training, and increasingly we'll see more and more of the workload shift that way. Um, it, cost is a super important factor that many of our customers are, are definitely worried about and thinking about on a, on a daily basis. And, you know, if you think about where a lot of people were, they went and did a bunch of these Gen.AI capability, or, um, tests, right, where they did proof of concepts and they launched hundreds of proof of concepts across the enterprise. Without really paying attention to, like, what was the value you're going to be or anything like that, and now they're looking at them and they're saying, well, my ROI is not really there, they're not really integrated into my correction environment, they're just kind of these POCs that I might get a lot of value out of, and they're expensive, as you mentioned. So two things that people are thinking about is, one, how do I lower the cost so that I make that, um, the cost much lower to run, and that's their point about cost of inference, and two, how do I actually get more value out of that so the ROI equation just completely shifts and makes more sense, and it turns out it's probably not all hundred of those. It's probably two or three or five of those…

AI assessment note: “we're still probably seeing about that ratio of fifty-fifty”

Q Do you anticipate that consumers are going to have their own bots and they're going to interact with AI bots and it's just going to be AI dialogue time forever?

A I think that you already can start to see that now on your phone where I have like about a dozen of these AIs on my phone from all these companies. They're kind of commodities at this point. You know, Google has Gemini, there's ChatGPT, there's u.com, there's Perplexity, there's Copilot, there's all these, these things, you know, they're just on your phone, you're just talking to the, basically the same data set through a very similar algorithm, and it's a commodity product. At some point, you're gonna be able to set up preferences and do things with those, you know, tools which you can't do today, where you're gonna be able to, like, ask them to run Projects for you. Hey, I'm going away for Christmas. I need a hotel. I'm looking at different options. Maybe I could go take my kids on some kind of incredible tour. Give me 20 ideas. And you come back the next day and it says, here, I researched this. I looked at it. I laid this out. Here's the ideas for your Christmas vacation. That's one idea. But now, let's say I'm one of those vendors. I'm Disney. I'm Hilton. I'm Marriott. I'm Ford, I'm General Motors, you know, I'm United Airlines, I'm American Airlines, I'm whoever it is. You're going to interoperate with us and our company as well, and you might do that through an agent. Right now, you might just be talking directly to us because that's the most trusted path, and we're in a…

AI assessment note: “eventually they're going to interconnect and, uh, collaborate and share”

Q So a lot is demanding, uh, more juice from the grid right now. You have UVs that are rising, you have AI, which like seems like it needs another nuclear plant every other day. And you have Bitcoin. I don't know how big that is, but I imagine it's big. Is our grid in good enough shape to sustain all these at once?

A Absolutely not. No, no. It's, um, I mean, that's, that's the, that's the thing that makes this such a unique moment in time. It's, uh, we have one and a half billion cars on the planet that need to be replaced. We have thousands of coal power plants that have to be turned off. We have thousands of natural gas power plants that have to be turned off, replaced. We have a grid that's Uh, archaic and doesn't have any of the supply demand mapping that we should be thinking about, you know, uh, doesn't have, you know, effective ways from a policy point of view, even managing things like net metering, um, We have buildings that are leaking air and aren't, aren't sealed properly for thermal. I mean, there's just like opportunity after opportunity after opportunity. And so you could look at it and say, oh my goodness, this is terrible. What are we going to do? Or you could look at it in our view and it's like, holy cow, there are so many business opportunities to go build this future world. And you said AI and I'm glad you brought it up. I mean, imagine the scale of this transition. This is not a once in a generation transition. This is a once in a, There's only one time that the world went through the fossil fuel revolution. So like, like the 18, late 1800 is only going to happen once. And the turnoff of fossil fuels is only going to happen once. And the turn on of renewable energy and…

AI assessment note: “Absolutely not. No, no. It's, um, I mean, that's, that's the, that's the thing”

Q Okay. So you've written them. What do you think?

A Well, I think first and foremost, it's, it's important that they exist. It's important that we have choices across lots of price points. Um, and to achieve that, you know, you need a few things. You need really heavy cost focus at the, at the design level. So the vehicles are engineered for cost. Um, You need to have supply chains that are really robust and efficient. Um, you, of course, benefit in the case of China from a much lower cost labor rate than what we have in, in the Western world, certainly in the United States. Uh, and then in the case of China, you do have just, you have this in the US, but to a more, uh, you know, you know, to a more elevated level. You have incentives that are happening at the, at, you know, the federal or the local government level. And so the combination of all those things have allowed for some really attractively priced EVs to be made and produced and sold in China. Um, and, you know, I think this is great. I think long, long term, some of the incentive structures that we see, just like we've talked about at the start of this call, you know, are, are not going to be permanent. But the efficiencies of design and some of the efficiencies that an EV affords in terms of architectural layout will start to play out as you get to lower cost batteries, you know, other than the battery and electric vehicle is, is actually a lot cheaper to build. It's…

AI assessment note: “I think first and foremost, it's, it's important that they exist.”

Q of published music more or less instantly. I can freely sample the, uh, work of new musicians. It has become aggravatingly difficult to find what I want to listen to with a recent product update. He says it became clearer than ever what the app has been pushing me to do. Listen to what it suggests, not choose my music on my own. What do you think about that argument?

A Well, I think this is an individual feedback, but I think generally you have very different types of users. So I'm, I'm going to get, I'm going to get this person back on Spotify a hundred percent. I think there is a, there's an interesting trade off here that is, that is real. So people want less friction. Um, they want to spend less time searching. You want to make things as easy as possible. Right. But there is this end of the line where you, you sit there and you just receive, you're kind of force fed and you don't give any signal back. Maybe a few clicks and so forth. Um, and that's something that, that we want to avoid. I think this is where the industry is going. It's going more towards distraction content and sort of just sitting and receiving. And it's a little bit of a dystopian, um, end of the line there. So what is interesting with Spotify, which we are reemphasizing is that it was actually a platform where you invested quite a lot in your own playlisting, right? And the, there's a trade off here between If we, you could have a vision is we should be so good at machine learning that you should never playlist again. That would be the goal. Um, because then you've done the user a great service supposedly, but then you also receive no signal and the user does no investment. So we're actually reemphasizing playlisting quite a lot.

AI assessment note: “there's an interesting trade off here that is, that is real.”

Q uh, increase in GPUs are going to get us first and foremost, and then we'll talk about whether the right way to scale these models is with just throwing more compute. Uh, cause I know you have a nuanced take on that. Uh, but like, first and foremost, like, If you go from 16,000 GPUs at the state of the art to a 100,000, what do you think that delivers?

A Uh, definitely delivers a bigger and better model. You have more compute. We know that scaling up improves things. Um, there's questions around saturation and whether continuing to scale up is justified, whether there's going to be enough gains from that strategy to justify the increase in cost. My personal perspective is that, you know, building a massive model, it's not actually useful for the world if it's too big to be consumed, if it's too expensive to actually deploy. Um, and so for Cohere, we've been very focused on building the right size of model. Um, but if your question is what is more compute unlock, it will be a better model. Objectively, we know that scaling leads to more capability, a smarter model that's more reliable. Um, and so that's the, the output.

AI assessment note: “definitely delivers a bigger and better model. You have more compute.”

Q The models are as good, right? We're still like at work.

A I don't think those two things are in conflict. I don't think that really, why not? Um, well, because I think that, um, we will never see mass unemployment of humans. I think that this technology is going to unlock more opportunities. It will let us do more as opposed to Scaling back what we do. Um, humanity is very supply side constrained, not demand side. We always, we want more. We want better. We want to be, um, healthier. We want to do more. We want to have things be cheaper. Um, and so we have all this demand and we're trying to keep up with our own society's demand and this technology, it's, it's true promise is in bringing productivity. And letting us do more. Now you can zoom in and you can like pick a specific field and you can say this field might be automated by AI. And I think that's true. And, you know, we should be thinking about retraining and shifting certain skill sets over to other, uh, new domains, like retraining people. But in general at the macro scale, I think this technology will create much more opportunity, uh, than it will take away.

AI assessment note: “I don't think those two things are in conflict. [...] we will never see mass unemployment”

Q that they've found them able to sort of come up with things that are outside of their training set. And there have been some papers that say, okay, actually they don't really have any emergent properties or emergent behaviors. And as someone who wrote the paper that kicked this all off, what do you think about that? Can LLMs have Any emergent behavior or discoveries that they weren't trained on.

A I think that they can, um, what's the right word? I think they can interpolate between skills, and so if they've seen how to do A, and they've seen how to do B, they can get kind of the average of A and B, um, but they don't just go completely beyond anything that they've seen. Um, I've never seen a model behave in a, um, totally unexplainable way. Um, they're really good interpolators. If you show them different domains, they can blend domains quite well, and, um, but yeah, I, I've heard the same thing about emergent behaviors, and, um, I, I think the research is really inconclusive there. Uh, there's not a lot of compelling evidence that says we're gonna have some Uh, total step change or capability takeoff. Um, even in the, the, like the latest state of the art research, a lot of it's about synthetic data and models teaching themselves. And so, um, self-improvement is this notion of can a model actually teach itself without human intervention. This is now a huge part of model building. It's a big part of how we create data at Cohere. Um, and before this started to become mainstream and actually part of the production process of creating these models, people were saying self-improvement, these things are just going to take off. They're going to become superhuman overnight and we won't be able to control it. Um, well, it turns out that doesn't actually happen.

AI assessment note: “they can interpolate between skills... but they don't just go completely beyond”

Q This is your expertise. You're like a, you know, almost like a narrow neural net train for that one specific purpose. And now we're giving it over to AI. And they'd look at the, just the billions being invested in this technology and say, well, what am I really getting for that? If this is effectively doing some of these things that humans are quite good at to begin with.

A I would counter and say that, um, risk to supply chains is many trillions of dollars. I would say that lawyers are extremely, extremely expensive and you don't want them combing through your documents, no matter how efficient you think they are. And same thing with doctors. We, we really want them spending time with patients, not, um, combing through hundreds of notes, uh, and filling out, filling out forms afterwards. Um, I would say those are. Maybe this stuff feels banal. Maybe productivity feels, um, boring compared to some of the hype of AI, but it is the value. This is what we're trying to build for. Um, and so I, I would push back quite firmly on that.

AI assessment note: “risk to supply chains is many trillions of dollars. I would say that lawyers are extremely”

Q And what happened to them when they started to see big tech encroaching on their territory?

A So, somebody relayed this story to me. I think that it's a, it's a guy that owned a bunch of grocery stores in suburban New Jersey, and when he saw Amazon getting into delivering food, it was like, oh, ok. Uh, I, maybe instead of putting all this capex into my supermarkets, I'd be better off just buying Amazon stock. And, uh, again, this might be 2014 or 20 15, and I think he's ok now. I think I haven't, I haven't checked in with the person who told me the story, but I would imagine if he were allocating to Amazon rather than spending money on his own, uh, capex to improve his supermarket, he's probably gotten a much higher rate of return on, on his investment. And I told that story as sort of a metaphor for the way a lot of people were thinking in those, in those days.

AI assessment note: “I told that story as sort of a metaphor for the way a lot of people were thinking”

Q And that also like, you're like sitting on a table with someone and they're like, keep on looking at their, their watch. So does this take people out of the real world in a way that, you know, might be detrimental? I mean, what do you think about the real world implications here?

A No, tremendously important question. And we, we've tried to use a tagline more present and connected, uh, the idea of, of kind of reducing the dichotomy we often feel. Um, and actually one of my, Anti-patterns that I've been telling people about for these glasses is just having to be notification services because notification services are an invitation to leave the present moment and instead go elsewhere. Like that's what a notification is intended to do. So let you know, Hey, if you want to, you could leave this physical plane and go dive into this digital surface over there. I want people to be able to do that if that's what they want to do, but it can't just be that. Um, instead though, If you think of the examples I've given, it's about adding context. And so the real difference maker between these and a phone in your pocket and all the things that are attached to a phone in your pocket is they have the ability to be aware of the world at all times. And this is where AI has been such a tremendous tailwind for our vision for these. Um, you know, we really thought early on when we were developing these, even until maybe a year ago, it was going to be holograms and then at some point later, AI. And in fact, it's going to be the opposite. The AI is going to come first. It's already come to, to the Meta Ray-Ban glasses. Um, and so for me, at least we've got this idea where, okay…

AI assessment note: “Anti-patterns that I've been telling people about for these glasses is just having to be notification services”

Q I mean, do you expect them to build something like this as well?

A I have no idea. Uh, you know, they, they keep it pretty tight, uh, down in Cupertino. They don't, they're, they're a pretty tight ship and still don't have great information on that. Um, I think they're, I think they would love to build something like this. I think we've been investing longer. I think we've believed in it more. Um, and as a consequence of frankly, Mark Zuckerberg's faith, um, and tremendous execution by a team of Thousands of people over a long period of years. I think we have a meaningful lead in a bunch of the technologies here. Uh, but you'd be foolish to count out a great technology company, one of the great all-time companies in general, uh, at Apple. Um, but my point is like, I think that the fact that they already have this phone that's in your pocket has some advantages because you've now got this device that has a bunch of compute. It has a bunch of silicon. They can do great custom silicon. They can add more to that. It's also a disadvantage, I think, in a way. Are they going to be willing to walk away from, you know, one tremendously outsized profit center in pursuit of something that is fundamentally disruptive to that profit center? Um, and if you're centering more of the experiences on the phone as opposed to on the glasses, do you make the leap that you need to make in developer platform, in interaction design? Um, because a lot of those things t…

AI assessment note: “I have no idea. [...] I think they would love to build something like this.”

Q And if it's about the API you're providing, does that mean the main interaction layer is going to be through some sort of voice AI or?

A It's hard to say at this point. I think that's voice is one verb. I don't, I think voice has a great modality for some things. It also has a lot of Vagueness to it, even amongst humans, a lot of context that's required even among humans that obviously even a, a really great agent wouldn't have for a long time if it was with you. Um, and so I think voice is probably not the only piece of it. Voice and text are part of it. I'm sure there will still be a lot of very direct manipulation interfaces for a long time to come. Um, but you can start to imagine, yes, as voice gets more developed, my ability as a developer to create value in how I manage process data or how I, you know, um, manufacture new artifacts is actually The more important thing I'm doing rather than having to also extend that to the consumer through some kind of interface.

AI assessment note: “I think voice is probably not the only piece of it. Voice and text are part”