OpenAI is Making a MASSIVE Mistake

In this episode of The Delphi Podcast, Tommy sits down with Travis Good, co-founder of Ambient, to discuss why open-source AI may ultimately beat the closed labs and what that shift means for developers, businesses, and the broader AI economy.

Travis explains how Ambient is building a decentralized marketplace for AI inference, matching demand with underutilized GPU capacity while verifying that users receive the exact model quality they paid for. They also explore the risks of building on closed AI infrastructure, China’s growing advantage in open-source models, and why OpenAI and Anthropic may be creating long-term distrust among their own customers.

The conversation also goes deeper into what would happen if OpenAI achieved AGI first, why prompt injection remains one of AI’s most important unsolved problems, and how the current AI infrastructure spending boom could eventually trigger a broader funding shock or AI winter.

Timestamps

00:00 Intro
03:20 What Ambient Is and How It Works
18:40 Verified AI Inference
40:20 China and the Open-Source AI Race
49:00 Why Closed AI Is Making a Mistake
1:03:00 What Happens If OpenAI Reaches AGI?
1:17:10 The AI Bubble and a Possible AI Winter

Tommy: https://x.com/Shaughnessy119
Travis: https://x.com/IridiumEagle


🔗 Connect with Delphi

🌐 Portal: https://delphidigital.io/
🐦 Twitter: https://twitter.com/delphi_digital
💼 LinkedIn: https://www.linkedin.com/company/delphi-digital


📺 Watch on Youtube: https://thedelphipodcast.buzzsprout.com

DISCLAIMER
This podcast is strictly informational and educational and is not investment advice or a solicitation to buy or sell any tokens or securities or to make any financial decisions. Do not trade or invest in any project, tokens, or securities based upon this podcast episode. The host and members at Delphi Ventures may personally own tokens or art that are mentioned on the podcast. Our current show features paid sponsorships which may be featured at the start, middle, and/or the end of the episode. These sponsorships are for informational purposes only and are not a solicitation to use any product, service or token.

Transcript

SPEAKER_00 0:00

Open source is going to win, and you still have time to be the hero. Like, you don't want to be in the crosshairs of an opinionated infrastructure provider. Your distribution doesn't matter because your customers hate you. If you're Sam Altman or Dario, like you need to have a little bit of rational self-interest and say, like, maybe we should do some nice things. If we're taking your data, if we're perhaps disintermediating some of your business relationships, that creates a fear response. And say that they legitimately achieve a hard takeoff. And they've got a model which is just incomprehensibly smart. You know, what happens then is probably like a two-year AI winter.

SPEAKER_01 0:43

Hey everyone, it's Tommy from Delphi Ventures, and welcome back to the podcast. Today I'm joined by Travis Good, who's the co-founder of Ambient. Travis is one of the rare people who can go deep across open source AI, crypto economics, token design, and real distributed systems. I've known him for years. Many of us in the Delphi Ventures team has. We've invested into his project, and we're very happy that he accepted us in. And I'm excited to uh to chat with him today. So, Travis, how are you doing? I'm very well, Tommy. Uh, thanks for having me on.

SPEAKER_00 1:17

I appreciate it.

SPEAKER_01 1:18

Yeah, for sure. I I mean, listeners of the podcast know you've been on before, so it's excited to have you back.

SPEAKER_00 1:23

And it's uh it's a real moment uh for uh open weights uh open source AI. Uh a lot has gone on.

SPEAKER_01 1:31

Did you see the uh the Alex Carp video recently where some people argued he'd crashed out on open source, some people said he didn't. What was your take?

SPEAKER_00 1:40

Uh, you know, I think he makes uh a really good point, which is that we've all experienced a pain that I don't think the people at large had experienced before, uh, which is what happens when your trusted provider turns against you. Uh so all of a sudden, you know, Anthropic revealed that they'd been silently sabotaging ML research. Or I should say they didn't really emphasize this, but you know, somewhere in their 300-page release on Fable, they mentioned, oops, you know, we are gonna sabotage uh ML research queries. Of course, they walked that back and they said, now we're going to just uh fall back to a less capable model, which is also sabotage. Feel rugged on that one. Right. And so, I mean, but what this was very clearly pointing out, which I think Alex Carp was pointing out, is that if you're competing with anthropic, you are in their crosshairs. And that is not a good position for a business or for anyone. Like, you don't want to be in the crosshairs of an opinionated infrastructure provider. Like you just want to do good work. You want to be economically productive. And I think that that conflict of interest is something that uh, you know, crypto and credibly neutral infrastructure have a real uh possibility to resolve. And there's huge opportunity uh created by the conflicted infra providers.

SPEAKER_01 3:09

So, Travis, one of the reasons, like just to start off the podcast, we have a lot of topics to cover, but one of the reasons why I'm just so excited for Ambient is I mean, selfishly, it just plays into so many things that I'm excited about, which is open source AI, which is and the explosion of open source models and fine-tuned models and being able to get your costs down and being able to use AI more and the ability to not be um you know nerfed by an AI lab on my usage and things like that. Maybe just give us your overview on like what Ambient is, because it plays into this all really well, and I'm just excited to learn more about it.

SPEAKER_00 3:47

Sure. So I'm gonna give kind of the business version of it, and maybe we can also dive into the technical underpinnings of it. Uh so on the business side, you could kind of think of Ambient as uh Uber for inference. Uh we don't own the cars, uh, we own the network that matches every request uh to the best available supply and guarantees the quality. Uh, we focus our delivery on uh two canonical models, the most popular, highest intelligence large model, uh, and the same on the small model side. Uh and you know, Uber for inference is a good analogy, but if you're gonna uh go down a layer deeper, I'd say I'd actually anchor you on something like Costco. So you know Costco doesn't own its inventory. Uh suppliers fund Costco's inventory, uh, but Costco turns it fast. Uh, it keeps prices low and it owns the customer. And that is us. Uh the neo clouds, the independent uh GPU operators in the world are our uh suppliers. Uh and about 25% of the world's uh enterprise GPU capacity uh is currently sitting with people who are earning very poor margins on it, unfortunately. There are a lot of smaller inference operators who have 50 to 100 GPU nodes who can really only rent these GPUs out on marketplaces. They don't have the technical skills or capabilities to get higher margins on those. So to give you a perspective on this, uh rental margins for these operators are typically like 10%. Uh, but if you're providing inference services, you can get 40 to 50% margins. So there is a real uplift uh for uh these providers. And uh Ambient is able to help them capture that. Uh, we can be a trusted brand uh for them uh that guarantees the quality of delivery uh to the consumer and uh helps all our suppliers uh make money, improve their utilization, uh, and provide the credibly neutral service uh that people need and want.

SPEAKER_01 5:55

So, Travis, maybe just to recap a little bit uh to drive this home. If I'm an end user, if I'm an agent, if I want access to AI inference, a lot of people will go to Together AI, they'll go to OpenRouter, they'll go to Venice, maybe they'll use their ChatGPT or Claude subscription for their you know news Hermes agent or something like that. You're saying that they can come instead to Ambient. So if they come to Ambient, why do they get a better service? Like why is the cost lower? Why is the latency lower? Why is the model better? Like drive that home for us a little bit and we could get into it more.

SPEAKER_00 6:33

Absolutely. Uh so you know you could think of Ambient as kind of uh underneath the hood, high frequency trading uh for inference. Uh, we have a very sophisticated uh routing engine uh that is running all the time uh against an auction of capacity uh on our network. Uh and it can be profitable for uh all sorts of providers to dip in to Ambient to provide uh capacity. Uh and so uh as a supply aggregator uh with the ability to increase utilization of our suppliers, uh we have favorable scalability, availability, and reliability uh characteristics. Uh so one by one, uh, you know, if you use together as a consumer, you may notice that they're largely booked up. Uh they're pretty severe rate limits on what you can do. Uh and with Ambient, uh, you can come in and uh there's an option that's going on. People have spare capacity, arbitrary GPUs can join on. And if there's the equivalent of Uber surge pricing, you can still run all the requests that you need to run uh to get your business done. Uh, there's no arbitrary rate limit on you. And the network economics kind of guarantee uh that you have the capacity. Uh the network economics are also helping with the availability and reliability story. Uh so uh, you know, of course, we had this unfortunate event where there was a data center that was bombed uh in Dubai, uh AWS data center, and Anthropic uh went down uh for the southern part of the United States, which you know was kind of like a surprising outcome for people.

SPEAKER_01 8:18

Uh it's kind of crazy that that could even happen in today's day and age. You just don't think that that's a thing.

SPEAKER_00 8:24

Yeah. Um, but what it shows you is that you know, providers like this, and same with the open weights providers, by the way, uh providers like this are operating at the edge of their capacity envelope. So if any data center is disrupted, uh then a lot of service gets disrupted because there's nowhere for the traffic to go. And so a network that can address the existing neo clouds and get improved utilization on those, but they can also access 25% of the world supply that isn't being addressed right now is huge for your ability to do business. Uh and uh the pricing is also favorable because uh people are competing continuously to deliver service to you. Uh, and uh they are focused on delivering services on the models that you care about. So you know, we deliberately pick top models, and then the world's resources are focused on optimizing the delivery of those models. Uh, and that's so powerful. You look at what happened uh with Bitcoin, uh, you know, the early uh proof-of-work network, ambient is also uh a proof of work, uh and we went from doing CPU mining uh to GPU mining to ASIC uh delivery. And so, you know, I think the the question that I would the way I would put it is uh if you are using one of the canonical ambient models, be that the large or small model uh in the future, like why would you go to any other provider? Because you're going to have a global network of people entirely focused on delivering these uh in the most optimal way. And it's the same type of dynamic as you get with uh open source AI versus closed source AI. Like the diffusion of knowledge and the global expertise and resource arbitraging that goes on uh is much, is much greater in the end than what any given uh neo cloud or hyperscaler can do. Uh and there's kind of a collective wealth that you're tapping into.

SPEAKER_01 10:31

Travis, it's exciting to hear your thoughts on this. And I'm trying to think through like what this looks like in a couple of months and a couple of years, probably a couple of months given how fast this is all moving. But uh the Bitcoin analogy you laid out is is a really interesting one because uh the neo clouds that we see today, a lot of them don't own their own hardware. They're middlemen and they're taking a take rate. They're connecting supply of GPUs with those using it. So they don't own the GPUs, but they're also serving a huge quantity of models, like dozens, if not more, right um on their platform. It's a little confusing as to what what to use and and what not to use. I think the pros do. But looking at ambient, there's a lot of contrarian decisions you made versus that model to deliver or what you're delivering with your network. You're uh using people are using real GPUs on your network, and you're hyperserving or a few select models to uh do that, and people are competing within those models in insane depth to serve them the best way possible, whether it be uh electricity, whether it be software optimizations, whether it be changes. So there's this crazy competition within the model, which I think is unique. There's a lot here to go into, but uh I'd love to just talk to you about uh why you've made such uh polar opposite decisions for the network versus what you've seen out there.

SPEAKER_00 11:57

Uh this is going to sound a little bit strange, uh, but I think that in the past crypto has lacked a certain kind of ambition. Uh and the specific ambition I'm talking about is the one where you go head-to-head uh with established Web 2 business models and you lean fully into the advantages uh that crypto provides in terms of decentralized economic uh coordinations and you disrupt those businesses. And so Ambient's specific objective is to create uh high-quality supply that is uh highly available uh and uh become a supply aggregator in a way that is impossible uh for a traditional uh provider to do. Because, as you mentioned, uh the traditional providers are heavyweight. Uh they are into these uh leases, they're into these business uh uh uh relationships, which are great for providing raw capacity, but which also make them biased inherently. Like if I'm renting a bunch of GPUs, I need to serve traffic to those first so I can pay my bills. I can't uh credibly arbitrage supply across a wide variety of uh providers because I need to feed my book uh first. And so uh, you know, we want to, Ambient wants to lean into uh, first of all, a credible neutrality posture of that. Um, but secondly, uh the idea that once you can make something trustless, uh you can compose that and scale that in a way that is impossible before. Uh so uh right now, uh these providers can't bring on arbitrary supply uh because there's no way of knowing what people are doing. Uh and uh you know, you could have uh people who are serving models very badly, you know, the uh data could be corrupted along the way. There's just no way of uh bringing that on uh you know scalably. Uh ambient solves that problem algorithmically. Uh so we solve the supply problem uh with verified inference and we guarantee a level of quality uh immediately. Uh we also solve the delivery problem algorithmically uh through uh enhanced routing. Uh and so you know if you talk about us versus like one of these uh larger uh sorts of providers, uh, we're very asset-light, we're nimble, uh, and we're anti-fragile. And we are those things because we build uh on the strengths of crypto. Uh to talk about the contrarian aspect of this, uh, you know, I think that what we've observed uh, you know, with something like Bitcoin is that if you focus uh extreme amounts of economic attention on a limited number of problems, uh then uh the network solves those problems for you because they're economically motivated actors. And you want like 10,000 people working frantically to use whatever hacks they possibly can uh to make money uh to be working all the time on improving the nodes on your network. Uh and that's the strength that Ambient leans into uh for crypto. And it's why we choose to focus on the delivery of a small number of models as a core function. Uh so but let me let me expand this because I think that this also kind of touches on uh you know where the where the future goes. Uh so I use a shopping mall analogy. Uh, you know, anchor stores in shopping malls uh drive a lot of foot traffic. Uh, you know, you have uh Dillard's, uh you have uh a Nordstrom's, and people like to go to those spaces. Uh and uh, you know, this is true uh for uh suppliers as well as customers. Like uh customers like to be uh in uh interesting places and then they go other places uh in the mall. And uh the stores, the suppliers, uh like to be around anchor stores because then the foot traffic uh is available to them. Uh and you know, this is kind of how Ambient is set up. We want to be exceptionally good as an anchor store, uh, delivering the most popular models uh that at the best possible way, uh, to the point where ideally uh no one would want to try and compete with us on those particular models, right? Like because the delivery just was so efficient that like you wouldn't want to uh do a margin game uh with us. Uh and we can talk more about the economics uh of that. Uh and but if you're accomplishing that, uh then you know uh the suppliers uh who are already there uh might want to do higher margin things as well. Uh and we can create an economy around that. And the customers might want to take advantage of those opportunities too.

SPEAKER_01 17:17

I I really like the analogy, and and I want to just maybe reiterate a couple of things to drive home Ambient. Like I think the demand side is is pretty straightforward, right? If you're a user, if you're an agent, if you need AI inference, you come to Ambient, you know you're gonna get a hyper-serve model, you know, the cost will be low, you know, the latency will be low. That's the goal. But the supply side that you dove into is really interesting. Like those that are serving these models, those that are competing. And you brought up a point about being both asset-like, and then also just having a totally different model versus the centralized competitors I talked about, together, base 10, et cetera. And one of the things that you mentioned was uh verification of the models. And I know you and I have spoken about this a lot offline, but when you're a big company, you need to talk to these suppliers, do KYC, get contracts, have salespeople, has have biz dev because they need to verify this party is serving this model and uh this will be the uptime, yada yada. But on ambient, the barrier to serve a model to me seems uh much lower. Like if you have capacity and you can serve it, you go on the network, you provide these AI inference requests, and the network itself is verifying that you're serving the model you say you are with all these other things. I think that verification is really important. Can we talk a little bit about how that works? Uh, because I think it's like highly technical, but I also think it's like really important for the story because it verifies all the demand or the supply for the network.

SPEAKER_00 18:55

Yeah, absolutely. So uh, you know, getting what you paid for has become a real issue. And and you see this uh when people are complaining about what I'm just gonna call intelligence compression uh for the closed weight models. It's like, wow, you know, Claude was dumber this week, OpenAI was dumber this week. Uh the result of that uh for the consumer is like a bad experience, right? Like you uh are not getting a capable assistant. You're having to correct uh the code, you're having to uh edit the briefs that it writes for you, uh, and it's just bad. And it doesn't uh, you know, and that happens with open weights too, by the way. Like serving patterns are are quite variable. Uh and uh what what that means if you don't verify is that you don't have a scalable solution. Uh people are not going to keep uh coming to your business if the solution quality is highly variable. And so you know, Ambient treats this incredibly seriously. It's the core of the customer experience that everything else is built around. Uh and you know, over time, uh our algorithm for doing this has gotten um more and more efficient. Uh so you know, if we were to have this conversation a year ago, I would say, uh, well, uh, you know, the algorithm is three different uh levels. Uh you know, we uh operate on logits, we operate on internal model telemetry, we create uh some complex mathematical relationships, and uh, you know, we derive a result. Um, but today I can tell you that uh we are you know 100% logits. Uh we do check all the logits, and what the logits uh represent uh is uh the state of the model's thinking, uh, which is very characteristic and can be expressed in terms of uh the choice that the model has made, uh, the spread of choices that the model could have made but did not make, and the associated probabilities. Uh, you know, all the things that uh you would consider kind of represents you in some way. It represents your personality, it represents your knowledge. Uh, and Ambient has got an extremely efficient way of getting that fingerprint uh that is 100% secure uh that verifies that all the words in a particular text, for example, were produced by the model that you expect. Uh and we've gotten this technology to the point where we can turn it on and off on a per request basis. So you can run whatever inference engine that you want. Uh we do uh VLLM and SGLang ourselves. Uh, we also are going to support uh Llama CPP. We have a beta version of that, and we can talk more about that. But you can, as a miner on this network, like you can essentially turn verification. On and off. Verification is basically costless. Like you can't see a difference in the speed for verified versus unverified inference. But what it buys you is the ability to just hop on to the ambient network and deliver high quality supply at the drop of a hat and deliver increased capacity for serving inference, which is the critical business goal. And it's that something which is scalable. It's a scalable, high-quality uh product for consumers and enterprises that all of a sudden you can participate in.

SPEAKER_01 22:42

I love that. It's just so it's so simple. I I guess like what the attack vector is someone trying to join the network and serving a worse model because it costs them less and they get rewards. Like that would be the like the simplistic attack vector.

SPEAKER_00 22:58

Ambient has dealt with different aggregators. Uh, you know, I'll give you an example uh uh of how this goes on one of the aggregators that we deal with. I don't want to name names, um, but uh you know, we deal with uh aggregator that does tests on a weekly basis, uh, you know, intelligence tests uh to help verify the supply. Uh and uh what we've observed uh is that the serving patterns differ greatly on the day that the tests are going to be administered uh versus on other days. Like, you know, everyone slows down and it's like, wait a minute, like how come everyone is operating at like 50% of the speed uh that they normally would be? And I think that the simple answer is because uh every other day of the week, they're running a compressed version of the model, which goes a lot faster. But when they want to be tested, uh, you know, they go back to serving the full fat uh thing uh just to be on the safe side. Uh and uh you know it's it's slow. Uh and so, but that you know, that's also like the core of the problem. Like as a consumer, you want consistent delivery. I mean, pick your quantization, but you just want the experience to be consistent. And if things are just up and down because someone's trying to game a metric or uh because they're dealing with high uh demand and they can't cope with it otherwise, like that's that's a real problem for you.

SPEAKER_01 24:28

No, I I totally agree. Um people want what they pay for. And I think my higher level view was I always just thought of that as hey, if I want GLM 5.1, I want that model to be served to me. But I think what you're describing is one layer deeper, which is that, hey, you might be getting that model in title, but you're not getting the full brain of that model. You're getting a compressed version or part of the version or something like that. Is that fair?

SPEAKER_00 24:59

Yeah, and I I think you know the counter-argument you'll sometimes hear uh from people is well, it's like it's good enough. Like uh, you know, people are are still are still using that. Um, but I like to think of this more as a function of desperation. Uh the question I always ask people is uh if you could get verified and unverified inference for the same price, which is what ambient offers, why would you ever choose unverified inference?

SPEAKER_01 25:25

I I don't know why you would. Yeah, it'd be actively choosing for less intelligence is a weird choice.

SPEAKER_00 25:31

Right. Um I don't want guarantees about what I'm getting. Like, oh, okay. Like you know, it's just because nobody nobody offers that right now.

SPEAKER_01 25:40

No, I'm it it makes sense. And maybe Travis, just to push back a little bit to to learn more from you on the network, like part of the sell of the supply side is you want uh people around the world to compete to serve these models in the best way possible, right? And I'm just wondering a couple of things. Like the first one is like why will they serve like put their intellectual capacity into serving these models on the Ampient network, right? Like w how do we compete with the open source labs at like GLM headquarters uh abroad? Like, how are we gonna serve models better than say them or better than a Minimax or an Alibaba, like the model creators? Like that's always a question for me because like they're extremely technical, deep. They've built these models, they're serving them. How do we get them served better on Ambient and why? I guess my main question.

SPEAKER_00 26:38

Right. And you know, I have I have a cheeky answer and a real answer to that question. Fuller function. Um, so I mean the cheeky answer is that we're uh I mean we're doing better than Moonshot at serving Kimmy on open router today. Like uh, you know, Ambient is right now. Like, yeah, we could we could pull it up, but uh you know, we're uh very typically uh top three. Uh we're usually number one uh on uh open router for serving uh Kimmy 2.7. Uh and you know it's because of some of the algorithms uh that we've built uh to improve utilization, uh to improve routing. And you know, I think that people discount the amount of work that is just associated with delivery optimization. And it is a little bit different than uh what is required for model training uh or uh you know just uh other types of uh patterns uh that you would typically uh serve. Like inference is like a very particular beast. Uh so you know, if you're talking about uh like what motivates people, uh there's an economic answer uh which is pretty straightforward, which is a lot of the folks that we're addressing uh don't have economic opportunities, right? Like uh they can't uh serve traffic on open router, they can't be recognized as an independent brand. And as a result, they're stuck in low margin world. Uh there's no one going and getting contracts uh for the demand side on their behalf. Uh they don't have the capacity. Uh you know, they might have loans that are 17% and they're making 10%. Who is this business development arm who's going to be fighting on their behalf to bring demand onto their network? And so the simple answer is like going from 10% margins to going to 40% margins is transformational uh for them. And that is a reason to fight, uh, to participate, uh, to uh improve things because the more share on this network uh they have, uh, the more rewards uh they can get. And the the share part is very important because you know, on Ambient, you don't just earn a transactional-based reward. Uh you know, this is a difference between us and like a purely uh web 2 operation. Like in web 2, like you get the transaction and that's it. You know? Uh and uh you better be profitable on that transaction or you're gonna be losing a boatload of money. Uh you know, with Ambient, you're getting the equivalent of a transaction as well as some ambient stock, you know, based on your contribution uh to ambient. Uh the inflation-based rewards is like uh almost like a stock-based uh compensation uh that you're getting. And you can choose to cash that in, uh, or you can choose uh to hang on to it uh and see the value appreciate. Uh, but you're actually getting uh double the rewards because with a useful proof of work network, uh, we are paying the security budget as rewards for something that is also a useful computation, in this case, inference. And so, you know, you've got this opportunity to uh underbid people. Like if you think that your stock-based compensation is going to be worth a lot of money in the future, you can come in on a per transaction basis, you know, like maybe 10 or 15% under, uh, and still realize a huge percentage uh on your overall gains. Uh and the more share that you do uh in a given epoch, which is you know, tokens in and tokens out, uh, the more of those stock-based compensation type gains you get. And so you want to participate uh in the ambient network because uh, first of all, you didn't have the opportunity uh to participate in an economy like this with rewards like this uh before. But also, the more you do, really the more ownership stake uh that you have uh in Ambient. And other people aren't offering you that same deal. I'm not getting uh shares in together because you know I I do a lot of uh inference for them uh or or open router, like you know, they're just taking their their cut, right?

SPEAKER_01 31:20

Um I think the uh I I think the idea that somebody serving uh models for ambient or or those running GPUs can think through, hey, I really am potentially bullish on ambient, and I'm gonna earn these tokens, and I'm gonna figure out my own view on what those are worth today. I'm gonna discount them. And maybe if I'm serving a dollar of inference, maybe I can serve it for 95 cents or 90 cents or something like that. And that undercuts our competitors, it lets them compete because they're all thinking that or viewing Ambient as valuable in the future. And the comp you bring in together, like yeah, nobody's getting together stock for serving inference because it's not a network. Right. Yeah. Yeah. I Travis, one of the I I know you like the Costco example. I really like the together AI without a company example on the AI side where people can just come and hypercompete. And I really like the Bitcoin example on the crypto side because folks are hyper competing to mine and serve requests. One potentially interesting question I have for you is like, how much design space do you think there is to hypercompete to optimize these models, right? Is it one like tip and trick? Is it like let's deploy 20 ML guys here and let's figure this out? Like, what is the design space to better serve these models on Ambia?

SPEAKER_00 32:46

Yeah, you know, uh, I think that there is a lot. And, you know, at a high level, we've gone through three uh iterations of our own uh design. And I think there's still uh mileage to be had. Uh, you know, very simply, we started out with a pure auction system uh that was doing 10 million tokens a day on open router, uh, which is really nothing. Um we were sad about that. Uh we worked really hard and uh we uh produced uh like a cache-based uh routing system uh that combined with the auction. So we would look uh and uh if you were uh qualified on your cache, in other words, if you cache this work so it was going to be easier for you to render the next tokens, uh then we would favor you a bit uh in auction terms. Uh and that uh saw us get some improvements. Uh and uh, you know, then we're getting like 100 million tokens a day of uh inference, which is still, you know, it's 10x, but it's still pretty bad. Uh and so you know, then we're unhappy about that. Uh and I'll spare you a lot of the technical gyrations, but where we ended up uh is with an incredibly uh complex, high-speed, predictive HFT style uh algorithm uh that looks at what someone can actually serve in a moment uh based on historical traffic patterns. Uh and we have it's uncheatable because it just relies on external metrics. Uh and we got very, very good at routing exactly the traffic that any given miner could take at a very particular point in time without tipping over. And that improved everybody's utilization, and that got us to the point where you know we're serving 10 to 15 billion uh tokens per day uh open router, like a vast uh improvement. I still think there is gas in this tank.

SPEAKER_01 34:51

I like that there's more gas in the tank. I like that you guys are already live on open router serving these models. This isn't just in theory. Um I uh yeah, I don't know if it's worth asking. It's a very specific question uh based off what you said, but like you know, one of the the my favorite parts about humanity is like the ability for anybody with a really good idea to rise up, make something for themselves, or like crush it, like the whole capitalist market economy idea. Um but like if I'm a really smart researcher and I have this crazy optimization for a model that's being served on ambient, and I can make it 10 times more or 10 or 10% more efficient or cheaper, or there's there's an optimization. But I don't have the capital to buy the GPUs to serve it on Ambient. Is there something that we can do or you can do to like bring that person's innovation to the network despite their inability to buy the GPU to serve those models?

SPEAKER_00 35:46

Absolutely. Uh so one of the things that we want to explore uh is uh pooled support for miners. Uh so uh, you know, I want to be clear, like I think someone could absolutely uh deploy like a small model uh minor type uh advantage, and we could talk about that in the future. But for you know, these really large models, uh, you know, I agree with you. Like if someone has a brilliant idea, uh Ambient the Foundation would like to create a means by which that they could um deploy that uh in the world. And so one of the things that we're exploring uh is a way for uh people to crowdfund that essentially, uh, you know, using uh ambient tokens. And then uh you know, we can go out and uh rent uh the supply uh for them uh through our big book of uh providers uh and just sort of transparently provide them access to uh work on it. Like I yeah, I think that that is uh absolutely uh something that helps spark uh innovation in the network and is the exactly the type of ecosystem uh that we want to uh fund.

SPEAKER_01 37:00

In my personal view, something I would love to see is um people think of Ambient not only as inference, but as the beneficiary of like AI optimization moving forward. Um and I think you've opened the door for that.

SPEAKER_00 37:15

Yeah, you know, the the thing I'd pile on here with is, and I I haven't talked about this a lot, but in a strange way, Ambient is a programmable machine learning economy. Uh so what I mean by that is this inflation-based reward is currently being put on inference, right? The more inference you do, uh the the better uh you do as far as the inflation that you capture. Um, but that doesn't necessarily need to be true in the future. Like that inflation is a budget uh that we can spend programmatically uh to spin up new lines of business. Uh so uh you know, you could imagine uh that there is uh like uh a crypto AI uh training concern uh you know who really needs GPUs. Uh and you know they come to us and they're like, well, you know, we'd like to make some sort of economic arrangement. Uh what Ambient has the ability to do is, you know, implement these proposals, flip the inflation, uh, so that some portion of that uh favors miners who are doing, say, training-related jobs uh for a period of time uh such that those jobs uh get completed kind of transparently uh with the network's incentives uh in real time. And then the profit is shared back uh you know with the network for those things. And uh what that what that means is that like we can essentially be, if you think of the shopping mall, a host uh for the boutiques. Uh, you know, there are these high-end jewelers, there are these high-end boutiques in the shopping mall. Uh, they sell to a specialist audience. Uh, we can go make deals with them and stack a bunch of them up uh in our shopping mall and transparently use uh the supply which is present uh delivering the bread and butter of the network. Uh those GPUs, uh, they're getting fed by delivering commodity models at high speed for good margins. Like that's how they're getting fed day to day. But if they want to go up the economic ladder, they can then volunteer uh for these other types of activities. So I think that in some way, ambient becomes the foundation uh for a bigger economy. And you know, that's that's what excites me is I think that a lot of the dreams uh that we have, a lot of the aspirations uh that we have uh in crypto uh start to become possible if you create this type of uh programmable uh economy. I I think that we have a very uh very good model for that.

SPEAKER_01 40:00

No, it is it's a good, it's a good viewpoint. Helps drive home kind of what you're building and and why it's exciting. Travis, I'd love to get your thoughts on some potentially harder questions on the the network side and and where where the world's going with AI. Um I mean we have a a closed source America, um, which has been insane, and we have an open source China, where we're getting incredible models of you know, GLM, Cloud Mini Max. Um it kind of begs a question like where things are going, right? Like it seems like uh the US is not slowing down on closed source deployments. Uh if anything, our access is diminishing as they restrict our access to models and give us nerfed versions of fable and things like that. And then the Chinese open source model side, like you know, Zipu and and others are trading at you know, 1400 1500 times sales the last time I checked, like they're not really making any money because the neo clouds in the US serve their open source models. People don't want to send their data to China, but they want to access them. I think one thing that could potentially hurt Ambient is if China just says one day, hey, we're going closed source. Like we don't we want to serve our models, we want to make money out with them, or they go to neo clouds like together and they say, Hey, even you guys could serve them, but they're private and we want a revenue cut or something like that. Like, how do you think Ambient fares in a world where in a world if China goes closed source?

SPEAKER_00 41:28

The thing that has made it possible uh for China to uh be open source is and to operate at this level is that the barrier to entry for training, like a really good open source model, has come down radically. So, you know, if we would have had this conversation like a year ago, it would have been inconceivable that GLM 5.2 would be nipping on the heels of the latest closed source models. Uh, it would be a radical heresy that they could train that thing for $25 million.

SPEAKER_01 42:04

GLM 5 is really good. I yeah.

SPEAKER_00 42:08

Yeah. And so I think that you need to look at this from a macroeconomic perspective. Uh, you know, let's say that next year it costs you $5 million to train a really great open source model. And uh, you know, you think that uh you could follow the same progression that you're talking about uh with Zeepu and uh uh uh GLM5, uh where uh you know you get on the map uh by putting out your open weights model that's really good. Uh you get some sort of high valuation, investor interest, public interest uh based on that, and then you slowly uh close it off. Uh well, I think a lot of people are just gonna want to do that, right? Like, you know, the barrier, you know, if the barrier is like 5 million, I feel like there's a lot of investor capital in the world that is going to uh fund teams who are gonna release model that it's like, oh, by the way, ours is open weights and we're trashing uh GLM 5.6 uh in the ratings, right? Um I mean, didn't we see like just like a Chinese DoorDash competitor release some ridiculous OWL model? Crazy. Yeah, I don't see that coming.

SPEAKER_01 43:18

Yeah.

SPEAKER_00 43:18

And so I, you know, I think that my my judo move, if you want to call it that, Tommy, is I can agree with your premise. Like, you know, it might be that the future for these particular companies uh is to shut the doors. Uh, but I would say that the economic incentives on the other side to like compete with those guys and undercut them and undermine them in the perceptual uh marketplace is so high that actually this dynamic is going to keep going uh for a long time. And if it keeps going for like another three years, we get to the point where this pre training capability is so commoditized and the knowledge about how to To train these models well is so commoditized that it really becomes all about model delivery and supply aggregation, which is where Ambient sits. And so I agree with you, it's a risk, but I think the macroeconomics of it mitigate the risk and actually, you know, ironically perhaps create opportunity for supply aggregators like Ambient, because then everyone is going to be competing on the delivery side.

SPEAKER_01 44:28

Yeah, that is that is really interesting. So just to feed it back to you, what you're saying is that China could go closed source, but if they did, it wouldn't be expensive to release the models they're doing, and somebody would pop up and do it. And so it's in their best interest to keep doing I guess Travis, one of the debates we like we see everywhere is like is China actually innovating or are they just distilling? And I don't read all the AI papers, there's too many, but uh I did read the Deep Seek ones and I see what's going on, and I remember them really innovating. They're using tier two hardware, they did a mixture of experts. Like, do you view the China landscape as uh innovative, or do you view it as distilling and stealing the US stuff? Just curious what you're saying.

SPEAKER_00 45:15

Yeah, uh, I think that they are largely uh innovative, and I think that this is gonna sound strange, but I think that distillation is almost incidental at this point. And I'm gonna make a bigger point with this. Uh so you know, we've been aware a lot in the Twitter sphere or the X-ray, whatever, whatever people are calling this these days, um, that uh everyone's saying I X'd.

SPEAKER_01 45:41

I just like saying I tweeted. Yeah, it's annoying.

SPEAKER_00 45:43

Yeah, I I X sounds weird. Uh I agree. Um uh like X2, like sounds punitive. Yeah. Um so you know, I think that everyone is a 10x engineer. Like we are seeing like ridiculous open source contributions all the time. You know, I made one of them. Like uh Ambient Desktop is our agentic harness, uh, desktop.ambient.xyz. It's over a million lines of code, right? Like this was mostly created by me and uh like a small percentage of a couple of our uh teammates uh like in our off hours. Uh we have a regular business to run. Um and it works pretty well, and it's MIT licensed, and and people can use that. And you know, the time that we spent on that model was focused on fixing the bugs and making it an end-to-end functional product and putting in lots of tests uh on that thing. And so that things like that are being introduced to the data sphere at an exponentially increasing rate. These are LM products that have been perfected by humans. And if you think about it, these are like perfect training artifacts. Like uh, you know, people always used to say, oh, like Anthropic has this unattainable edge because they've got all the interaction patterns. It doesn't matter because we end up with perfected outputs. And you actually often see the history of how the outputs were perfected. Like you can train on all the pull requests that were created on these repos. You don't even need like the anthropic user data. Like you can just look at how Claude revised like each of these things. And so I think in a world where everyone is ingesting all of this data all of the time, you have effective distillation. Like we're all distilling each other, and we're all essentially doing human feedback on all of the outputs of the LLMs implicitly. And that's going into training. And so, yeah, of course, like there, there's probably some you know, hackery involved. There's probably some level of like direct uh distillation. Um, but the reality is like this is just a much different and easier game on the high quality uh data sets.

SPEAKER_01 48:11

Um It is an interesting take. I mean, if you if you had to think through like uh where we're going in AI, and I I hate to use like a multi-year outlook because it's just too much happens or too fast, but uh maybe if you had to take like a three to a six month view given uh how fast all this happens, like it's it's hard to imagine we just keep playing tit for tat, right? Like, uh Claude releases a god tier model or China like quickly has one that catches up. Like it seems like eventually we get to a point where things really change. Like, where do you think uh the closed source side is going? Where do you think the open source side is going? Like I'm curious like how you view these end states because you think about this all day and you're looking at it much deeper than than I am.

SPEAKER_00 48:58

Yeah, so I have uh maybe a direct message for the closed source side, uh, and then I have a view on where I think these things are probably headed. Uh so my direct message uh to the closed source side is open source is going to win, and you still have time to be the hero. Uh America has always been great culturally. We have achieved the greatest soft power in history by sharing our cultural artifacts, by sharing our research, by sharing our perspective widely. This has always worked for us, and it has created like a durable world peace that I think is you know unmatched in some ways. Like there's a world stability that exists that you know you couldn't say existed like 200 years ago. Um, and it's because of like a shared uh cultural space. Now, I think that that we're in a time where this is fraying, uh, right? Like, uh, and I think it is a huge mistake to completely change our approach to this. Uh and so like specifically, I would say to the closed labs, like, would it kill you to release some research on a six-month time delay? Would that would that kill you? Would that kill your projects? Would that kill your edge? Uh, I tell you, it would create a lot of goodwill. Uh, it would create a lot of architectural innovations around the world that would bend your way. Uh, it would probably create, if you release open weights models more often, a huge mind share uh for American models. And I think people tend to prefer American models. Like, you know, GEMA is like really, really popular uh and it's you know getting a little bit long in the tooth, they periodically refresh it. But I think Google is the best one about refreshing this. Um imagine if Google released a slightly bigger, more capable model. Imagine if OpenAI released another open weights model. You know, I think that that uh there's still time uh to turn the narrative around a little bit because uh right now I think that the world feels like they're being beaten up by America's closed AI.

SPEAKER_01 51:28

Travis, I mean, I I agree with you. And it is crazy. Like I walked around near IPS last year, and there's no researcher from Anthropic or um OpenAI, it's all closed, right? There's a ton of research from Google there and and others, thousands of really smart people. But I guess if you had to take it a step further, not altruistically, but what is like uh the reason dollars and cents that uh a Sam Alton or a Dario would wake up and say, Hey, we're gonna open source this, because it clearly takes a huge hit because as we both know, they're making insane margins on the API. So just curious.

SPEAKER_00 52:07

Yeah, so I think that um open AI and Anthropic have distribution. This is what everyone fights for. Like you want to have distribution, you want to have direct connections to all these uh businesses. And Anthropic and OpenAI have the world's ultimate Rolodex. I mean, the the world's VCs have dumped more into these companies uh than probably anything else combined in history. And so as a result, they have the world's best Rolodex. Like, that is an amazing advantage. And so if you start out with that advantage, the question is like, what can you give up to build goodwill uh while you just spend your time capitalizing on distribution? Uh and I think that there's probably a decent amount uh that you could give away, and your edge starting this would be uh still sufficient uh to take you forward. For I think the risk uh that is created uh by uh trying to use a crypto term uh max extract, uh, which is what we were talking about uh with uh you know the Alex Carp anecdote, is that your distribution doesn't matter because your customers hate you. Um and and so like you if you're Sam Altman or Dario, like you need to have a little bit of rational self-interest and say, yeah, like maybe we should do some nice things, maybe we should have some gestures uh towards the open source community, maybe we should give away some stuff. Um, because if we're taking your data, uh if we're perhaps disintermediating some of your business relationships, uh if we're replacing some of your core tools in your organization, like that creates a fear response. And if people are afraid of you and you're also acting in a very aggressive manner, they're unlikely to continue to do business with you. Uh and so I would say that is the self-interested dollars and cents response for why they should change uh postures.

SPEAKER_01 54:14

Maybe up that line of thinking, Travis, it the Palantir interview with Alex Carp was really interesting, right? Because it's it was the first, I think, major company that came out and said, you should own your model weights, you should own your data, like you should own what you're building. And that is a polar or opposite viewpoint of what the AI labs are selling us today. They're saying use our API, build everything with us, and we'll have your data. You don't need fine-tuned models, we could eventually compete with you. The Palantir side, and I'd call it the open source world says not just the model, but everything around the model should be yours. The orchestration, the memory, the fine-tuned model weights, the harness, the company data, the permissions, the approvals, the users, all this proprietary business knowledge you should maintain and you should use as a business because that's your resource, right? And that's very different from what the labs sell us. So I'm curious about your view. Like I know you probably tend to agree. So I guess my more pointed question for you is are we actually gonna get that world or not? Because right now it doesn't seem like it, given how fast anthropic and open I've grown. So I'm just curious, like I because I know you agree with that, but I'm just curious, like, what is the realistic outcome that we get to with with that viewpoint?

SPEAKER_00 55:36

Yeah, so uh, you know, I think that I want to just highlight this and then properly answer your question. Um, you know, networks like Ambient can offer you like complete privacy as an enterprise. You know, we support running end-to-end encrypted uh in trusted execution environments. Um, you know, if you're a consumer, uh, we can do that. Uh, we can also onion route uh your requests uh so that you're not identified. Uh we can remove personally identifying information up front. And so uh there is a way that uh your privacy and your data integrity uh can be preserved while you're getting a guaranteed quality of service. And you know, you can access that service however you want to. You just get the API and like go to town. Uh and that can be true for enterprise contracts as well as for just consumer interaction. So I think it's really important to highlight that I think that's a better service inherently uh than going to like an enterprise sales call with Anthropic and having them twist your arm and tell you, you're not really a big customer of ours. We're not gonna give you favorable rates, and we need a minimum number of tokens per month. Uh, and uh, you know, we're not maybe gonna do ZDR for our latest models because of safety. Um, you know, it's just I think it's just like a better experience uh that uh we can offer uh in in Web3. And you know, it does you don't need to take our word for it. Like we can we can have attestations, we can have uh uh verification, uh proofs, hashes, like all this. And so I think that's like a better experience. But I want to answer your question and say, like, you know, in terms of uh where this is headed, I actually believe that uh you know open source is going to win, uh, but it's going to be like a very painful, tumultuous period uh for closed source, like if it continues on its current course and speed. And so I think that that actually operates on like almost three different levels, and I'll try and explain the levels associated with that. Uh so you know the first thing that I want to highlight is something that uh you know I highlighted maybe at the beginning, uh, which is that everyone in this space uh is almost entirely asset heavy. Right? Like, and what asset heavy means is invested in the current paradigm. Uh so the current paradigm is NVIDIA GPU clusters, maybe with Vera Rubin, they're liquid-cooled. They exist in specially constructed data centers, which have plentiful access to water and power and you know have a specific uh performance uh profile, uh, and they're very, very expensive. Uh and so on a data center level, asset heavy means that people are taking a hugely long bet that this paradigm uh remains and continues and is the dominant paradigm. And I think that that's an incredibly risky bet because ASICs are a thing. Like, you know, we saw this progression in crypto, right? Like an ASIC is 14 nanometers. Like you can run it with uh SRAM, like you don't need you know high speed um uh HPM, right? Like you don't need DDR5 or 6 or whatever they're gonna come out with, right? Like you can do a 14 nanometer process that anyone can do, and uh anyone in the world can do 14 nanometer. You don't need special expertise. Uh and you can just win, and I think this is what China is doing. Uh, you can just win on sheer scale of that, because you can make so many of those boards uh that you can you know darken the sky with them. And it, you know, if you spend the same amount of money that you would on an NVIDIA installation, then you might get actually vastly better capacity. Uh and uh you know it's very cheap uh for you to just keep churning on those things. Uh whereas you know, NVIDIA you have these long cycle times. So if you look at that first layer, that data center layer, like potentially very vulnerable uh to disruption. And uh, you know, people are sort of following each other, sort of diving into these things. Uh and you know, ambience perspective on this layer is like we're asset light because we want to be anti-fragile to shocks uh that could happen here. Uh, you know, if someone is stuck uh in huge data center investments, like what are they gonna do if an ASIC player uh comes out? Uh you know, maybe they have to dump all their GPUs on the market. It's like great, great for ambient, you know? Um because they can serve to the network.

SPEAKER_01 1:00:37

Yeah.

SPEAKER_00 1:00:37

Yeah, they need to find a home, right? Uh but then let's let's go like a layer up. So the next layer is uh like the models themselves. And like the big bet, uh the all-in bet that people are making is that we're going to get self-improvels. So we're going to have models that are capable of designing themselves, and it's going to be a flywheel. And uh the speed that you get out of that uh for delivery is going to be unmatched and it will achieve escape velocity, and then no one can compete. Like, because if you were to bet on the other thing, if you were to bet like, oh, like uh they're just gonna be generally faster uh at delivery than other people, like three to six months ahead, like you wouldn't put a trillion dollars on them, right? Like you wouldn't value that at a trillion dollars. So you're you're betting that at some point their edge is just gonna be completely insurmountable, they're gonna crush everybody else, and they're gonna be the only game in town. That's the bet. Again, I think this is a very, very risky bet. Uh, and the reason for that um is because everyone is a machine learning engineer now. So we we talk a lot about like make uh make AI into a machine learning engineer. But the reality is that AI itself has commoditized this skill so much that if you put a moderately skilled human in the loop, like they can do pretty great like ML research. Um, you know, you don't actually need the self-improving machine, you might just need like a bunch of moderately skilled humans uh who are making constant improvements that are overseen by you know average intelligence LLMs. Uh and like that might be just good enough.

SPEAKER_01 1:02:32

I I I no, I guess, Trous, maybe like one pointed question there though. Like the bet is risky for sure to get self-improving AGI. But you know, if you if you're gambling and and there's a 1% chance you win and the payoff is a hundred trillion dollars, you you take it, right? Because the EV is one one hundredth of the trillion. Like so you you would do that. Like, is there an argument to do that and to pursue it? Because it seems like burning the boats and spending everything you can to achieve it honestly is the highest EV outcome for those players, right?

SPEAKER_00 1:03:08

Yeah, actually, like let's play that one to its conclusion because I think it's a fun one. I mean, I don't think people often talk about like how this how this ends up. So let's run the scenario.

SPEAKER_01 1:03:20

Hell yeah.

SPEAKER_00 1:03:21

Let's say that uh let's pick on open AI because I like them better than Anthropic. I like individual people at Anthropic, I just do not like them as a company at this point. Um interesting.

SPEAKER_01 1:03:33

I've swapped back and forth a lot over the years, but yeah, yeah, keep going.

SPEAKER_00 1:03:38

But let's say I I okay, yeah, let's say let's put our chips on open AI and say that they legitimately achieve a hard takeoff, and they've got a model which is just incomprehensibly smart as a result. What happens next? I would say that the first thing that happens is that nobody in the US except the government gets to access it. Yeah, this is the default response. Um, I mean, we've seen it, it's reflexive. It's like uh, you know, if you can hack the NSA, then you are getting banned uh immediately. You know, it doesn't matter like if the world is using your model or like a bunch of companies are depending on it, like you're getting banned.

SPEAKER_01 1:04:24

That's why I'm scared about Pliny the Liberator, because he keeps hacking Claude, and I feel like the NSA is gonna just go find him.

SPEAKER_00 1:04:30

Yeah, Pliny, uh like slow down, man. Like just take a vacation.

SPEAKER_01 1:04:35

I hope he has good OPSEC.

SPEAKER_00 1:04:37

Yeah, seriously. Um so okay, well, what does that mean? Like the US government has it, and none of the companies have it. Well, like if none of the companies, if none of the US companies have it, then they can't like really do anything with it. You know, like they can't achieve economic productivity ends with it. And so then the question you asked is like, well, can you uh you know deliver like a limited form of this uh to uh people? And I think that the technical state of the industry at the moment is no. Like you just fall back to Opus 4.8 all the time, and you haven't achieved anything. And then, okay, like but the US government has got it. So we haven't improved like economic productivity because it's too dangerous to widely release the model. And then it's like, what can the US government uh do? Well, uh, they can start a war with every nation on earth, which doesn't yeah, which doesn't seem like a great idea to me, actually. Like, you know, like I I mean I love the US, I'm a patriot, but I just don't think that taking on everyone in the world or even making a really hostile uh targeted cyber attack towards one of our enemies is going to inspire the kind of relationships uh that we want in the world. In fact, what I I think that would lead to is immediately everyone making an alliance against us and working frantically to develop their own alternative uh AGI.

SPEAKER_01 1:06:10

I I I totally agree with you. And and the interesting part too is like I think it's very simple for the government to say, you know, hey, here hey robotics company, you can use AGI because the end result is a new material or a new robot design or something physical we can see and touch that's nice. Or or a drug company can build you know solutions to cancer, like very straightforward. Like we can test it, we can use it, things like that. But to your point, it only drives the flywheel forward if there's a economically productive value that's created. And uh the other point is how do you pay for it? And I don't think that if 20 companies in the US have access to fable six through ten, they're gonna pay $25 per million input. Like I feel like they need to own a high percentage of the creations they're creating to pay for all this CapEx, right? It's just it's a weird flagwheel on the econ.

SPEAKER_00 1:07:09

So I I'm gonna make uh like the strongest counter argument I can think of to the points I've been making, just so we can steel man this thing. So uh, you know, maybe you know, you could be like, Travis, like, look, like actually uh, you know, we're not gonna get direct access to this model ever. But like the fruits of the model, the golden apples produced by the model uh can be distributed uh into the economy. Um and it's like okay, uh, but now like I've I've got like a different problem, right? Uh yeah, because uh I don't think the US has ever been like a fully centrally planned economy. Um but you know, you kind of uh I don't think that's a good paradigm, by the way. Um and and so it's like if the government is the only one who has access and they're like creating like specific products, then you know that puts them in the business of uh picking winners. Um and you know, I don't trust become the China we've wanted to avoid. Right. Like I don't trust them to do that. Uh I don't think anyone should have that power. I don't like that. Uh I don't think most of Americans would be cool with that. Like uh, you know, you know, you see, oh, the government like picked this business over the one that was in my hometown, and now my hometown business is destroyed. Uh like I just think that um, you know, that breaks everything. And so like maybe the government doesn't even know what to do with this thing, right? Like they they don't they can't really distribute the disruptive innovations from it, they can't like really actually start a war with somebody because it would cause everyone to gang up against them. Like, we're at really great at cyber defense at that point, so we we know we can't get hacked. Uh but it's unclear to me um like what benefit we have in a case where the technology is not diffused. So like the the biggest counter argument to this is like if you look at a non-diffused world, uh, which is kind of what we've been describing, and you look at a diffused world where capabilities are balanced, people are at rough parity, like the diffused world is a lot more functional. Like we still have vibrant competition, uh, we still have a balance of powers, uh, you know, we still have interests that can compete uh to achieve different outcomes for different countries. Um, that looks a lot healthier to me.

SPEAKER_01 1:09:42

Yeah, I really don't want the centrally planned economy. That the problem is it's just so hard to fit capitalism in with AGI tomorrow. It's so difficult in so many ways because the AGI companies obviously need to make money, um, but we can't trust everyone with it. I have you ever seen the movie Automata or Automata? I f I don't know how to pronounce it, but have you seen it? I don't think so. It it's an older movie, and it's a cop investigating a robot death who who goes sentient. And the crazy part of the story is it's a world with robots, and the only way that they've figured out how to secure the robots is to build the 10x AGI version or have it build the security mechanisms for the lowly normal robots, rather than just straight up kill the AGI robot because they say, hey, it could never be hacked. So I don't know if maybe there's a situation where a God-tier model is private and can consistently secures a lower-tier model on some recurring flywheel that's public, but that also feels a little dicey as well.

SPEAKER_00 1:10:45

Oh, you know, I think this is a really interesting one. I want to just pull on this thread because I think there's been like a real misallocation of resources in AI. And my simple um proof of this is that prompt injection is still a thing. You know, uh God knows that we've had AI safety researchers who've told us about existential threats, about their P doom level. They've given us these cyber war scenarios, they've like stirred the pot a lot. Um and yet we have prompt injection. Like all these people care about is uh you know addressing intellectual threats. They're not solving like the most basic thing uh that you know we're struggling with right now. And I think it's like a trillion dollar business opportunity, right? If you solve prompt injection tomorrow, uh you enable everyone to use full fat fable, right? You can't jailbreak anything uh anymore. Uh it's because of prompt injection, essentially, uh, that you can jailbreak models. Uh and so I guess you know where I'm going with this is that uh these models like we aren't caught up with the state of the art where we can really manage these models. Uh like we haven't, and because we haven't caught up uh with the safety side of this, uh we have like all of these problems. And there's like this continuous hurdle uh that we have in model deployment where we try and create like patches uh to address like a really fundamental uh research issue. And so, you know, I you know, Leopold Ashenbrenner uh has this graph uh that he goes on, and I think it ignores a lot of things. We've talked about this. Uh, I think it ignores social friction. Um, but uh I think it also ignores this problem, which is that the uneven development of uh model safety research with fundamental model research uh could actually uh handicap the whole thing. Uh like you know, let's say we don't go full AGI, uh, you know, maybe we go halfway to AGI. We still not might not be able to release these models uh because of prompt injection.

SPEAKER_01 1:13:07

Well that well, that's an interesting point. So you're basically what you're saying is the models can be jailbroken through prompt injection. And because this is an unsolved issue, we'll never be able to release the models because that's how these are fundamentally uh designed. And I I don't know. I mean that that is kind of scary because if the US or maybe it's not scary, but if the US models keep getting restricted and nerfed and taken offline because of that issue, what we've seen with Fable recently, doesn't that allow China and competing countries to just release better models that basically are just better than what we could release publicly? I guess that's not bad because they're open source, but America loses its advantage.

SPEAKER_00 1:13:50

Yeah, I and I think that that's really what could happen. And you kind of get into this, I'm gonna call it like a regulatory prison dilemma prisoners dilemma. Uh where the other guy could release a model which uh someone could corrupt, which could hack all your infrastructure. And because we don't, we've been safety conscious and we haven't released the models that could defend against those attacks, uh, you know, we get hacked six ways to Sunday, even though we have superior model capabilities.

SPEAKER_01 1:14:21

It is crazy that Anthropic has been the poster child since its creation on safety, but they can't solve prompt injection. It's odd.

SPEAKER_00 1:14:28

Yeah, it's uh so you know we talked about like this this model layer, um, but you know, I think that there is like one more layer that could go any number of ways. And you know, depending on how you think of it, it's like fragile or anti-fragile. Uh, and that is, I'm gonna say, harnesses and specific applications, right? Uh so uh, you know, uh my experience has been that uh harnesses offer radically different experiences right now. Um like clawed code is pretty bad uh as a harness, like codex is pretty good, Hermes is way better than open claw. Um and like harnesses make a big difference in terms of your uh capabilities. Uh and so uh you know the question is like you know, is this a durable advantage? Do we think that um you know the models are going to sort of implicitly become harnesses themselves because of the way that we train them with harnesses? You know, it's kind of like chain of thought was originally a prompting technique. And then eventually, like we just train the things with chain of thought and they you know they become um and uh you know I don't know the answer to that. I I suspect that the delivery mechanism is always gonna have a role.

SPEAKER_01 1:15:51

It is interesting though, because you and I spoke about like global versus fine-tuned models years ago, and like we always thought there, you know, two camps, like global model and then fine-tuned model that's really smart and specific for a specific reason. But it kind of feels like we're getting the fine-tuned model experience through the harness because if all your company data and memory and proprietary data and all the specialized stuff that you've built out uses an open source model, you're not training the model to change the weights and to fine-tune it, but your de facto end result is very similar to having a fine-tuned experience.

SPEAKER_00 1:16:29

That's I think that's exactly right. Um, and I think that that is going to be uh true unless we solve continual learning. Uh and I also have a belief that solving continual learning is gonna make prompt injection like a thousand times worse.

SPEAKER_01 1:16:49

It'll continually learn how to prompt inject.

SPEAKER_00 1:16:53

You don't know like what the training data is at that point. And if you've ingested like a prompt injection somewhere along the way, uh then you know, maybe your model goes rogue on you and it's like because like 10 million tokens ago it saw this attack from a uh a Fisher, right? Like it's ridiculous.

SPEAKER_01 1:17:12

Travis, could I I want to ask your take, like changing gears a little bit, just on like the the AI build-out, right? Uh-huh. Um so I I released a a thread like a month ago that that got a bunch of views on the AI CapEx cycle. Um it's it's been nuts. I mean, we're we're seeing basically you jump kind of bottleneck to bottleneck. Recently, it's been memory and and there's been others, and the stocks go crazy. But my my key like lower IQ take is that these businesses are just spending way too much money. Um you're seeing businesses cancel their AI subscriptions because the labs are too expensive, and there's a perfect substitute, right? You get 90, 95 percent of the intelligence for a hundredth or a tenth the cost, and or you get 10x the usage by using open source, by using ambient, by using other players. And my thought is that long term, a lot of these businesses will shift to open source, which benefits Ambient and others. And obviously a small percent will stay with frontier intelligence. They want the God tier models at any cost, they're making new medicines, they're making new materials, or whatever. The the question I have for you though is how does that ripple through like the funding cycle? Because if that the if the growth switches off the labs, like and they don't have a way to fund it, like we talked about earlier, owning part of the creation, they don't have the money to build data centers, get talent, yada yada. And that could invert the flywheel a little bit and hurt our ability to get real AGI or something like that. So I'm just curious if you agree or disagree or kind of how you're seeing the CapEx cycle here.

SPEAKER_00 1:18:44

So I, you know, I fully agree that we are in the moment of open weights inference. Uh, you know, if you look at Together AI, fireworks, modal, base 10, uh, you know, we're talking about like five billion dollars of value per player added in like the last five months. Um, you know, investors are like, we see it, we see the wave, we're gonna surf the wave. Uh Ambience is also surfing this wave. So, I mean, I'm I'm selling my book here a little bit, but like the reality is it's a really big book. And I think that you're absolutely right. Like, this story is going to uh continue. Um, but I think what you're pointing out is also that there is this peril to the whole interconnected um ecosystem, right? Because a lot of the data center CapEx, uh like Core Weave and OpenAI, for example, are inextricably connected. Uh Core Weave funds other data center players, Nvidia funds Core Weave, like there's this ecosystem. It's very circular. Um, and I think that you're you're right. Like there could be a huge uh shock to the system. Uh and uh you know what happens then is probably like a two-year AI winter. Uh, you know, because uh and and this affects, and this is why Ambient is constructed the way it is. It's why it's constructed to be asset-light, uh, is because it goes like this. Uh OpenAI isn't able to go out at a trillion dollar evaluation. Um, they come in at like a $500 billion evaluation. There's a downround, like something, something happens. Uh and all of a sudden, they aren't able, as a result, to fulfill their data center commitments uh with Oracle. Uh Oracle has like bet the company on OpenAI if you look at their uh financials. Um so Oracle like I don't want to say goes to zero because it seems like they would be rescued. Like I feel like the US government is dependent on Oracle enough that like maybe they're too AGI to fail. Let's see. Yeah, like um maybe Oracle gets marked down like 80%. Um now Oracle can't fund data centers, right? And you know, together, fireworks, all these guys have been drafting off of data center capacity that is essentially being created for uh OpenAI and Anthropic. Uh so now they have maybe more demand, uh, but no like real ability to fill that demand because the data centers aren't being built. Uh and so uh I don't think their businesses are gonna fail. I think they're very good businesses, uh, but they're sort of locked at this capacity ceiling and they have to start turning people away. Uh, this, by the way, is like the perfect storm for ambient success. Like, you know, people are gonna, I need capacity, give me capacity, you know. Um, but uh like I think that uh that hurts American businesses um because they can't access the breadth of services that they want. It slows down the overall economy. Uh it hurts everyday Americans because like big pension funds, um, big uh private credit is involved in data center build-outs. And like these, you experience uh like a shock to this, and like people's 401ks are gonna get hit. You know, so I think it creates a wave perhaps of unfortunately anti-AI sentiment. People are like, you're taking our jobs, you've hurt our retirements, uh, you know, you throw in some propaganda about water usage, even though data centers are super water efficient. Uh way more. Uh and yeah, it's like a golf course or avocados, like uh almonds, yeah, like uh almonds in California, research it. Um that's that's homework for those who are still skeptical of this. But like I think it creates like kind of an anti-AI backlash that creates real political problems uh for uh the country. Because if you get that, you can have lawmakers reflexively pass uh laws. We've kind of seen some of this happening, uh, you know, where there's already sort of a populist uh socialist movement uh that's kind of uh taking fire uh in New York. And I think that you it's a quagmire. Uh and so you know, that's why I believe that it's incumbent on the closed source players to actually change course a little bit uh and change their approach uh and sort of diversify uh in some sense and not just not just bet everything on uh you know AGI, but really be nice to their distribution uh and be and you know lobby the US government to let us be really nice to the world. And uh, you know, then I think that it makes us more resilient to this kind of shock. But it's a huge problem. I don't know if I have a great answer to uh you know what we what we do in this case, but uh that's how I think it plays out.

SPEAKER_01 1:24:13

I don't disagree with you. I mean, you've thought it through more from a American and political perspective than I have. I I just I view it as long term we will get AGI. I'm very bullish on that. But the concern I have is short term. It's businesses realizing they can shift off of those labs. And you know, the silicon LLM token index, the the price to use a million tokens, is down 20 or 25 percent in the last 30 days. Like you're seeing people use cheaper models. And my worry is I mean, in the best case scenario, we get AGI, we get all these new materials on Frontier Intelligence, and then on the open source on your side, everybody has access to intelligence as cheap as humanly possible to advance both sides of this barbell. Um but I'm just worried about the the funding cycle in the near term and that's slowing down and what it what it could cause. So hopefully we we don't get there. Um but but Travis, maybe shifting to final thoughts here, like you've been building ambient for years. Um you're the guy that wrote the response to situational awareness called situational blindness way back when like it was it was great. Like where do you think ambient goes? Like, what would make you really happy to see a year from now if we were recording again? Like, where do you want to see Ambient? Like people using it, the software, the GPUs, the network, like what would be the the success here?

SPEAKER_00 1:25:36

Yeah, for us, like uh you know, since we're supply aggregator, we want to have a huge book of supply. We want to bring all those mom and pop operators on. We want to bring the smaller neo clouds on. We want to give them real economic opportunities that are allowing them to serve inference uh to the world and make good margins on it. Uh and we want that flywheel to be spinning uh continuously and self-perpetuating. We want people to tell their friends, hey, like uh you don't need to rent your GPU. Like you can actually serve inference and you can make a ton of money. Like we want the word of mouth uh to be strong, and we want people to be scrappy and like actively finding ways to optimize delivery on our network. Like that is the supply side of it. And then the second part is like global distribution. We want to be a recognized brand. You know, we've started a little beach head here on OpenRouter where we're number one on the delivery of one model. We'd like to be number one on the delivery of 10 models. Uh, we want to be a name in everyone's lips. Uh, then we want to uh have adoption of our own uh subscription offering. Uh we want people to be using uh our API and also like our X402 uh offering if they're doing agentic uh commerce, uh and uh you know getting uh directly onto the platform uh to build the future of agentic commerce. And I think those two things are uh what create uh the flywheel uh for ambient. We've got a complete supply and demand uh loop uh that we want to build out. It's premised on the supply uh side because we think we can deliver the supply better and in a market that carries uh that cares a lot about uh quality and speed, like that's a key uh advantage. Uh but then you know we want to turn this into the global intelligence utility uh that we have been promised by all the closed source people, but who are too conflicted uh to be able to deliver that. Uh so you know, Dario is never going to be able to deliver intelligence as a utility. Uh his hands are tied by defense contracts. Uh, but Ambient can. Deliver a high intelligence network that's useful to everyday businesses and everyday people, that's credibly neutral, that's scalable, that's highly available and reliable. And we think life will be good.

SPEAKER_01 1:28:11

I like that. And Travis, how can people get involved? Like the demand side is pretty straightforward. They could access you via open router, they can download Ambient Desktop, they can access maybe an API, a subscription, but maybe on the supply side, if I've missed things or on the on the demand side, yeah, just curious how people get involved.

SPEAKER_00 1:28:30

Yeah, so you know, one thing that we didn't talk about at all is that Ambient really implicates two models. So one of those is Ambient, which is the big model, and uh the other one is uh Ambient Mini, uh, which is a small model that can kind of run on any GPU. And our intention with our network is actually to intelligently orchestrate these models in the future so people can get the most bang for their bug. And I'm happy to announce uh that in about the next week, uh we are going to open up first the closed beta uh for small model mining, uh, and then the open uh PvP season of uh ambient mining uh for the small model, uh, where anyone can actually uh come on our network, uh try out uh small model mining using the ambient desktop app, uh, which is cross-platform, uh Mac Linux Windows. Uh you can you know mine it on your laptop uh if you want uh and uh send requests essentially to other people uh on the network, uh give back responses. Uh we're gonna uh offer that on our own platform. I'm gonna try and offer it on open router as well as like a free um model. Um you know, they're they're a little bit touchy about uh data, so I probably have to create like a sub-brand of ambient uh in order to uh to do that successfully. We'll see how that goes. But um people can become uh directly involved in in mining ambient like pretty soon. Uh and you know, touching on something that you mentioned earlier, uh we'd like to create the opportunity for people who want to fund uh big nodes uh to do so uh collectively uh to mine the big model. Uh we will take uh no profit on those nodes. So 100% of the profit uh will go back uh to the mining pool. Uh we will do a 100% uh pass-through, uh just straight costs uh for those nodes uh to incentivize uh big model mining. Uh so yeah, I think that there are gonna be some opportunities to get uh directly involved in the near technical future. And then, you know, we're launching this thing this year. Like it's it's coming out, it's gonna be uh live, and we would like everyone to use it, subscribe, push hard on it, mine for it, and help build the kind of future that I think we all implicitly want.

SPEAKER_01 1:30:54

Travis, ambient's making the summer fun again. Um it it's super exciting. Um it just it's gonna be really fun to watch it play out. Uh and not not just fun, but but exciting, right? Like watching everybody compete to serve these models in a competitive manner, like watching the crypto crew get excited from the network perspective, watching the end users be able to access intelligence at cheapest possible cost with lowest latency. Um it's it's really exciting to see. So um yeah, we're obviously investors in you, and we we love that you've accepted us in and uh I really enjoyed this convo. So Travis, thank you so much for the time.

SPEAKER_00 1:31:32

Thank you, Tommy. Uh, we appreciate all you do. Um, if you are uh wanting a really active and helpful investor, I can uh you know happily say that uh Delphi is uh you know one of the best investors on our cap table. They've been tremendous uh for us and and you in particular, Tommy. Uh so thank you so much.

SPEAKER_01 1:31:51

Really appreciate that. We're excited to back you. So thank you everyone for listening and thank you, Travis.

SPEAKER_00 1:31:56

Thank you.