Tommy Shaughnessy: [00:00:00] Hey everyone, it’s Tommy from Delphi Ventures. Welcome back to the show Today, I’m extremely excited to host Travis from Ambient yet again. Uh, it’s Travis’, third time on the show. Travis, how are you?

Travis Good: I am great. Uh, Tommy. Thank you. Uh, just, uh, out here in sunny San Francisco, uh, doing the, uh, crypto startup accelerator, uh, with, uh, a 16 z, uh, so intense and a lot of, a lot of fun.

Tommy Shaughnessy: Travis, I’m so glad that we could finally chat now that ambient is outta stealth. Um, we’re happy to, to have co-led your, your pre-seed. And at the time we couldn’t talk about it. Um, I need the one-liner on ambient. It is so good. Tell me what, tell everyone what ambient is.

Travis Good: Uh, well, uh, ambient is a useful proof of work, uh, L one, uh, where the proof of work is verified.

Inference, fine tuning and pre-training. On a single large language model that runs on every node, uh, of its network. It is [00:01:00] a fork of Solana that converts Solana, uh, from proof of stake to proof of work and operates at full solana speeds.

Tommy Shaughnessy: It’s insane. We’re gonna get into all of this, but before, let’s, let’s walk backward.

What is the end goal of ambient? Are you building an, are you building an L one to build an L one? Are you building an AI model to build an AI model? What, what’s the end result?

Travis Good: Yeah. Uh, well I think this, uh, this gets into the nature of, uh, crypto and currencies. Uh, so, uh, you know, if you look at, uh, at Bitcoin, you know, uh, Bitcoin is a magnificent project, uh, that achieved one of its goals really well and did not achieve another one of its goals.

Uh, so Bitcoin’s original goal, uh, was to be a currency. You know, it’s something that we wanted to be able to. Pay for pizza, for pay, you know, pay for pizza with, and, uh, transact with. And, uh, you know, if we had achieved that goal, I think it would’ve [00:02:00] reordered the world in a really healthy way. Uh, it would’ve decentralized, uh, a lot of the power structures, uh, that I think are, uh, grinding against each other, uh, today.

Uh, and, uh, it would’ve been wonderful. Uh, but, uh, you know, in the face of its success as a store of value, um, Bitcoin pivoted away, uh, from being a currency. And today it is denominated in dollars, uh, which I, which I think is really sad actually. Uh, you know, it’s an abstraction on top of an abstraction. Uh, and, uh, you know, ambient, we want to go back, uh, to the beginning.

We think that we have an opportunity to do that because the world economy is changing so profoundly. We’re moving towards a world where there’s going to be more AI agents doing work, [00:03:00] uh, than people. And, uh, you know, those, uh, AI agents are powered, um, by inference, um, by machine learning. Uh, and that’s their cost basis.

And so in that world, uh, it makes sense, uh, to have a currency, uh, that represents a unit of machine intelligence. And that is what ambient is seeking, uh, to provide, uh, to the world. We want to be this unit of intelligence that can, uh, be usefully applied in any sector of the economy, in any nation of the world, uh, because every nation has adopted.

AI and AI agents to do useful work. And we think that, uh, you know, if, uh, AI is the work that we’re gonna do in the future economy, uh, it makes sense to use AI to secure [00:04:00] blockchain. If we’re devoting all of the world’s resources towards running more and more models, uh, faster, uh, you know, and they’re, they’re not that many models, but, uh, you know, the, the big ones, uh, are, are gonna do a lot of the real work.

Uh, then, uh, it makes sense to take advantage of that supply of hardware, uh, to secure a blockchain. Uh, and, and that’s the, the goal of ambient. It’s to provide this currency, but also a store of value that reflects what’s going on in the world, uh, which is that the AI dominated economy, the AI portion of the economy is growing and growing and growing.

And, and that’s why, you know, something like ambient becomes more valuable, uh, over time, uh, because, uh, you know, that, uh, that supply, uh, is becoming, uh, more and more value valuable and more and more prevalent. [00:05:00]

Tommy Shaughnessy: Travis, that’s, it’s awesome to hear. It is sad when you bring up that we denominate Bitcoin the way we do, uh, after all these years.

But focusing a bit on ambient, what exactly does that mean, right? Like, I hold your token, what, you know, with Bitcoin I have, there’s many different ways to splice it. The easiest is, you know, I have $82,500, uh, this horrible night. ’cause Trump likes to be unhinged. But, um, what does it mean to hold ambiance, token?

I have access to a model or a blockchain or inference. Like, like what do I have access to? That’s a great

Travis Good: question and I, I, I wanna take a step back and. You know, talk about something that’s been happening in the world of models. And I’m gonna relate it to your question. Uh, and what’s been happening in the world of models is a convergence of capabilities.

Uh, so we’ve seen closed source models and open weights models, uh, start to achieve a very similar outcomes in most meaningful problem solving domains. [00:06:00] And our thesis is that that standardization and convergence is going to continue, uh, because the economic incentives to do that are simply too strong.

Uh, like all these players, uh, need to try and balance each other out, or they’re gonna be left behind. Uh, and I think China has discovered maybe to the US economy’s detriment, uh, that they can achieve a sort of geopolitical balancing of the scales, uh, by releasing open weights models. And I think other nations are gonna do that.

Um, but what that means is when you have these, uh, convergent capabilities. Uh, you can pick one large model and get most of the benefit of machine intelligence. Uh, and, uh, you know, in ambiance case, going back to your question, uh, what we do is we say, uh, we are offering, uh, this large model to you. Uh, it’s going to be the highest scale delivery of this, uh, such that, uh, [00:07:00] you’re getting the cheapest possible cost, uh, for verified inference at the fastest possible speed.

Uh, and that value proposition is represented by a token of ambient. It is the world’s commonly defined capability for machine intelligence. Uh, that’s manifested in a token.

Tommy Shaughnessy: It’s, it’s crazy to hear out loud because if I own OpenAI stock, I have no access to the models of the inference. If I own, if I own Ambient, it’s totally different.

That’s right.

Travis Good: And I think that it has the power to be a, a hugely democratizing force. Um, because, you know, our belief is that, um, people should have ownership, uh, in this phenomenon. And it shouldn’t be through, uh, proxies. Like you should be able to own it directly. Uh, you know, uh, open AI [00:08:00] and all of these other, uh, closed source providers are very dangerous choke points, uh, for a lot of, uh, you know, phenomenon, uh, authoritarianism, uh, censorship, um, uh, propaganda.

Uh, and, uh, we think that, uh, no one should have that kind of power over the world. Uh, that it’s, uh, it’s like handing, you know, no matter what you think of Sam Altman, uh, it’s like, uh, handing him, uh, the, the one ring, uh, giving him that kind of power. What are the for? Yeah. It’s just, it’s just corrupting.

Tommy Shaughnessy: It’s not even him. Right. It’s, it, I don’t know of a person that you can give that to that it wouldn’t corrupt. Right. That’s right. It could be me. You, it

Travis Good: could be anyone. And so, you know, with That’s exactly right. And so with Ambient, [00:09:00] uh, you know, we want to leverage, uh, the design, uh, the decentralized design of crypto to create a natural balance of powers, uh, that keep bad things from happening, uh, and keep abuse, uh, from happening.

Uh, and that’s, you know, that’s the beauty of crypto. Like, you can, you can do that. That was the wonderful technical innovation of bitcoin. Uh, that, uh, you could motivate, uh, the creation of strong emergent communities just with the proper, like decentralized, incentivize incentivization mechanisms. Uh, and, uh, we want to be a counterbalance, uh, to these other sort of, I guess you could call ’em, oligarchical forces in the world, um, that I think are slowing humanity down.

Tommy Shaughnessy: So, Travis, I know we’re gonna jump around a lot from mining to inference to the models. I wanna spend a little bit more time on the, the mining side. [00:10:00] So, uh, John Girl, one of my colleagues who, who you know very well, he had a, had a tweet about your announcement and one of his bullets was. Ambient will incentivize miners from all over the world to power a single model showcasing the power of crypto and coordinating the much needed compute that AI demands.

There’s a lot to go into there, but maybe one of my more specific questions is, if you are successful in convincing GPUs around the world to mine and secure ambient, it’s, it’s become so nuanced how deep seek versus OpenAI versus others train these models and make them better, right? Mm-hmm. Like, why will getting more GPUs make your end product this model, uh, better?

Travis Good: Yes. Uh, so there’s a lot to unpack there. Um,

Tommy Shaughnessy: uh, but I may have asked the question wrong too, so feel free to take

Travis Good: No, no, it, it is totally fine. So let’s start there. And then there are a number of other related things to, to talk about. And so I, I think you make a great point, uh, which is that, uh, you know, ambient is starting with something that has been [00:11:00] provided in the world.

Uh, you know, we might start with deep seek R one. Or it might be R two, or it might be llama, it’s whatever the state-of-the-art open weights model is. At the point that we do our test net and our main net, that’s gonna be the model. Uh, and we think that model is gonna be, you know, 600 billion parameters plus, uh, because that’s what’s providing, uh, the really high intelligence and value, uh, for people.

And we could talk about other things, but you know, the question is like, uh, what, what’s going on? Um, you know, with, with ambient in the meantime? Like what’s, uh, and what value is ambient itself adding? Uh, and I, I would start with, uh, you know, something simple, uh, which is that a distinction between, you know, just an open weights model release and a model that’s really kept up to date.

Uh, so, you know, as a power AI user, you’ve experienced, like, it’s really nice when [00:12:00] GPT-4 oh comes out with a new release. The knowledge is fresh. Uh, like it doesn’t give you 3-year-old Python library references. Um, you know, uh, like probably questions you’ve asked on the internet have made their way into the training set, and all of a sudden it has answers for them.

Uh, you know, uh, there’s a combination, uh, memorization and genuine innovation that happens, uh, that you miss actually, uh, with a lot of the open weights models because they don’t have a steady, uh, release cadence. Uh, and so one of the things that Ambient as a network is designed to do is to take that base model and steadily improve it in a few different ways.

Uh, so one way is to continuously fine tune it, uh, with up-to-date knowledge and provide those fine tunes on the network, uh, so that, hey, like if you just wanna do logic problems on the base model, that’s totally fine. But it’s pretty easy to flip a [00:13:00] switch. Uh, have a hundred megabyte LoRa update that gets you into a fine tune with all the financial news, uh, on the market, uh, for example.

And because this is a proof of work blockchain, like we can build into the program of the system, uh, activities that keep the systems’ knowledge up to date. Uh, so that’s one thing. Uh, but then another like really exciting thing is, uh, related to the deep seek paradigm. Uh, so, you know, for those who who don’t know, uh, deep seek, uh, pioneer, this virtuous, uh, circle I like to call it, where, uh, you use reinforcement learning on an already good base, uh, model, and then it produces a bunch of synthetic data, uh, that is really useful for training the next base model.

And so, uh, ambient because it is a proof of work network and is continuously churning, uh, [00:14:00] can devote a portion of the network capacity to generating huge synthetic data sets, uh, with reasoning traces, uh, that can be used to increase the intelligence of the model. Uh, and, uh, so that is some of the value, uh, that Ambient provides.

It’s really changing the paradigm from having this increasingly stale snapshot of an open weights, uh, model release, uh, to something, uh, that is, uh, cultivated carefully, uh, by the network, uh, to maximize utility, uh, to the communities that it serves.

Tommy Shaughnessy: That is, uh, it’s, I read, uh, Sam Lehman’s Post, say, the world’s RL gym.

I, I think I saw you mentioned as a yes as a contributor, and it really didn’t click for me how important this. Synthetic data reasoning. Trace flywheel is for models, right? I always just assume this outdated mental model of, you know, [00:15:00] data trains model output. Go back to train the model. Let’s double click on that for a second.

Sure. So what is a reasoning trace? Like, walk us through a reasoning trace, how the network produces that and, and why it makes the network smarter. Yes. And if you could, Travis, maybe just parlay that with how it’s so different for just updating the zeros and ones of a, of a model.

Travis Good: Yeah, absolutely. So, uh, when you’re talking about reasoning traces, there are a few different ways of going about it, but I, I’m probably gonna focus on two.

Uh, so, uh, one is where you have very well-defined, uh, problem domains, uh, that you’re training, uh, a large language model, uh, to deal with based on kind of automated feedback. Uh, so imagine, uh, you’re trying to get a model to write a Python program, uh, and you give it access to a Python debugger. And so, uh, what you do is you have the model output, a bunch of reasoning traces, and the ones that don’t, uh, trigger Python debugger errors are [00:16:00] the ones that get rewarded.

Uh, and, uh, you know, if you’re looking at harder problems, uh, the ones where, uh, there were all the problem solving approaches, uh, kind of shared similar problem solving steps, we reward the similar problem solving steps because we, we have this idea, uh, that like statistically we’re kind of honing in on the correct problem solving approach.

And there’s sort of a majority of, uh, you know, voters in that, uh, approach who reflect a certain viewpoint and we sort of reward them, uh, for agreeing with each other. Uh, so that, that’s sort so Reasoning trace to answer your question, actually is a series of problem solving steps. Uh, that gets rewarded in different ways.

Uh, and of course we’re trying to reward correctness, uh, with those, uh, reasoning steps. Uh, and so, uh, was your next question, like what do we do with those, or, or [00:17:00] which

Tommy Shaughnessy: Yeah, maybe, maybe both. Right? Like how do we take those reasoning traces and my mental model is just sort of like a lot of different pathways the model takes until it’s successful, but Yes.

Like how do we take those and make ambiance model better? Like what’s the, how do we go back?

Travis Good: Yeah. Um, so we, we take those reasoning traces, uh, and we do more pre-training on the model. Uh, so that’s the other part, uh, of what, uh, ambient is designed to do, uh, is, uh, you know, when the time comes, when we’ve built up sufficient reasoning traces on the network, uh, the whole network takes a portion of its compute power.

We can talk about how that works economically and is devoted to. Upgrading the capabilities of the model, um, by, uh, pre-training it with these reasoning traces

Tommy Shaughnessy: that is, I have to ask another stupid question, but so you’re, you’re taking the reasoning traces and you update the model. How is [00:18:00] it, why is it so much better for the model to update with reasoning traces than it is to update with just raw data, a new encyclopedia, a new manual, a new textbook, like Pure Raw Data versus the reasoning traces?

Why is it so much better to have the reasoning traces?

Travis Good: It’s a great question and it almost gets philosophical, but I’m gonna try and give you a pragmatic answer here. So, uh, high quality data is incredibly important, uh, to these models. You know, what we discover is that, uh, smaller curated data sets are better.

For pre-training, uh, for fine tuning, uh, then bigger noisy data sets. Uh, and the reason is that, uh, you know, you could think of these models as, uh, pattern recognition engines. And so the more clearly that you can spell out, uh, good patterns, uh, the better it is for the models learning rate. Uh, and the more it’s going to converge on something that, you know, humanity finds useful.

Uh, to give a little bit of a, [00:19:00] a silly example, uh, like if you throw a bunch of romance novels at these things and you mix it in with a bunch of logic problems, you know, you’re, you’re give some incoherent, uh, like value system here. Like not a good Hallmark movie. Yeah. Yes, exactly. Um, and so that’s why, you know, these reading reasoning traces, you could think of them as sort of the purest distillation of patterning.

They are logic patterns that in some strange way, as you introduce more and more of them to the model universalize, the model’s intelligence.

Tommy Shaughnessy: That is so cool. And I, I gotta keep going. You have the best answers and takes here. So when I read the world’s rl Jim, when I’ve, when I’ve talked to like the news research folks and all these others, there, there’s always this recurring theme of these, uh, you know, there’s always domain specific data, domain specific models.

And now there’s like these domain specific reinforcement learning, uh, gyms as Sam calls them and Right. And other areas where, [00:20:00] you know, if you’re a doctor, you want specific reasoning traces to train a specific model, right? Mm-hmm. But in your, in your situation, it’s one giant model. So That’s right. I, how do you view the trade-offs of like, having many ambient models with specific reasoning traces versus the one global?

Travis Good: It, it’s a, it’s a great question and, uh. The first thing that I would say is that the IQ of a given base model is kind of fixed like once you’ve, once you’ve baked it. Um, and that is, that can be a detriment if the IQ is low. Uh, and it can be, uh, uh, an asset if the IQ is high, because what happens then, you know, ambient supports a model and its fine tunes.

Uh, and so, uh, you know, on our network you’ll be able to go like train, pay to train a fine tune of ambient using your own gym. Uh, and the validators will check that that was a valid, uh, training [00:21:00] process. Uh, and, uh, you know, that can uh, be something that is like perfectly obscure to your area and you can just load that, uh, fine tune on the network and call it as much as you want, uh, and achieve good results.

And the reason that we’re confident you’re going to achieve good results is because. You’re building off of the base IQ of the model. And as long as we can pick something with a high enough base iq, uh, we’re very confident that it’s going to learn the right patterns. Uh, you know, I, I do, uh, like probably way too much fine tuning, uh, in my very limited, uh, spare time.

Uh, I, I occasionally, I occasionally, uh, tweet about it. Uh, but, um, you know, one of the things that is really funny to me is how much better, uh, bigger models do with smaller data sets. It’s kind of a weird counterintuitive thing. Um, but like, if you feed, uh, a really high [00:22:00] quality dataset, uh, to, uh, llama 8 billion, uh, it kind of, it kind of converges, like it kind of gets better at answering the problems and then it kind of plateaus.

Um, and if you feed the same dataset to like a 27 billion parameter model. The convergence happens really fast. Like it’s just smarter. And so it’s figuring it out faster and you have to, you have to stop a lot faster, uh, in some cases because like, for example, the, the Gemma models that were released, uh, recently, like they’re really good for fine tuning.

Um, and, you know, I’m ending my training jobs early, uh, because the model has already learned everything it could possibly learn, uh, from the dataset. Um, but that, you know, that gives us a strong indication that if you have a really big model and you’re fine tuning that, uh, you’re gonna get a lot of benefit, uh, from those fine tunes.

Tommy Shaughnessy: That is really cool. Yeah. I, I, I, I guess I go back and forth the smaller model, I would assume it’s like less people in the room to convince, but then more, it’s, it’s hard to go back and [00:23:00] forth on that one.

Travis Good: Yeah. The, um, you know, there’s some really interesting trade-offs that are going on in the space. I don’t wanna go on too much of a tangent, but I could talk to you briefly about my testing of QWQ.

So, yes, please. This is actually, you know, it’s a 30 billion parameter model, roughly, um, maybe 32, I don’t remember. And, uh, it’s really up there in terms of its benchmarks. Uh, like it’s right on par if you just look at the benchmarks, uh, with, uh, deep seek. Uh, and, um, when you use it, uh, you start to feel a great deal of disappointment.

Uh, and, uh, the disappointment comes in. Uh, you know, maybe it’s a form of the bitter lesson. I don’t know what it is, but the, uh, the disappointment comes in the fact that there is like a timespace trade off. Uh, so, uh, it turns out that in order to get an intelligent answer out of QWQ, you have to [00:24:00] output like 60,000 reasoning tokens.

Um, you know, I actually almost burned my Mac down the other day, uh, running this thing locally. Uh, you know, I gave it a very basic math problem. And my God, that thing was thinking so hard about that math problem, and it was asking itself rhetorical questions like, am I really sure about this? You know, and it was going, but wait, but wait, that’s the classic, uh, uh, phrase that these models give.

Wait, wait, uh, and then it goes back round. Um, and it spun for a solid 16 minutes. Uh, my God. And, uh, you know, deep seek, uh, which I was running on some H one hundreds, uh, just went, yeah, it was like 15 seconds. It had a very succinct reasoning trace. Uh, the answer was immediately clear. It got to the same answer.

Uh, it’s just that the cost of shrinking that model down. [00:25:00] So much, uh, was that it took an incredibly long time for it to emulate the level of intelligence of deep seek. And that’s not something they discussed in their benchmarks for obvious reasons. Uh, but really when I see these now, I think that they should just normalize them based on the tokens of output, uh, like intelligence per token of output.

Uh, which strangely enough is the same premise of ambiance. It’s like, well, it’s like we’re really measuring intelligence in a standard way, sort of per token of output. If we think that the intelligence of these models converges at a certain scale, uh, then that’s the value that you’re getting. Um, so

Tommy Shaughnessy: the, the output intelligence, does that include.

So like two plus two equals four. It’s 15 minutes on one model, 15 seconds on the other. I, I’m greatly simplifying the math problem. You probably did, but the intelligence per token, does that include all of those reasoning traces or is it just the answer?

Travis Good: It, it includes all those reasoning traces, uh, because that’s, you know, you can think of that as a proxy for the amount of [00:26:00] compute that it’s spending.

Tommy Shaughnessy: Okay. So if you both arrive at 15 and one uses 15,000 tokens, the other uses a hundred. The one model is much more intelligent. Yeah. ’cause it includes the time dynamic and the correct answer.

Travis Good: That’s right.

Tommy Shaughnessy: Okay. This brings us to another interesting question. Um, sure. How will Ambient be as fast on the inference side as the hyperscalers an open ai, clawed philanthropic?

How do I get my answers fast, if not faster than the centralized players?

Travis Good: Well, I think, uh, there are a few different answers to that. The first obvious answer is that you have to be operating on a fast blockchain substrate. And so, uh, you know, we’re very, uh, happy and, and proud to be building off of something that has proven to be a speed king, uh, which is Solana.

Uh, and we’re not doing anything to that blockchain in our fork of it, uh, that is impacting its speed. Uh, and it provides if you have that [00:27:00] fast finality, um, that, uh, allows you to present a similar user experience, uh, to what people are expecting, uh, which is that they pop the transaction in. Um, you know, models can even be streaming and our protocol and have streaming verified inference.

Um, but these transactions are happening very quickly. Uh, and, uh, you know, I, I think, uh, the future is bright, uh, for salon and we’ve set ourselves up to, uh, be downstream from the performance improvements, uh, that they make. Uh, but they are going to push the limits of TPS that we think is possible. Uh, you know, we love the engineering, uh, team, uh, team over there.

Uh, we, we know some of the members, uh, we’ve also, uh, met with Anza and we think they’re absolutely cracked in terms of their engineering talent. And, uh, you know, their contributions represent the, the first sort of pillar of, uh, ambient, uh, you know, which is absolutely like nailing the core blockchain, uh, tech.[00:28:00]

Um, but then, uh, you know, the second thing, uh, is, uh, really enabling efficiency on two sorts of, uh, use cases, uh, for, uh, LLM inference. Uh, so the first of those is what you’re talking about, uh, which is the. Bitcoin, uh, hyperscaler, you know, I just got into GPUs now. Um, uh, sort of, uh, profile, uh, and, you know, to do that, uh, what we enable is patching of, uh, open source, uh, highly competitive, uh, inference engines.

Uh, so things like A-V-L-L-M and SG lane. Uh, you know, these projects are spending, you know, really, uh, huge numbers of cycles on achieving bleeding edge performance. Uh, and we want to leverage that, uh, and also raise their profile as well because we think they’re great products. And so what we do, uh, for our verified [00:29:00] inference, uh, is we patch them, uh, to make them trustless, uh, without really impacting their performance much at all.

Uh, and, uh, so they support high scale batch inference. Uh, you can configure them to do routing around different machines. And, uh, it’s a really good experience. Uh, and you know, the benefit of that, you know, building on top of successful platforms, uh, like Solana, like VLLM, uh, is that you sort of for free inherit all the innovations that they implement.

Uh, and in the case of like VLLM, you know, you can be sure that they’re pouring over the deep seek, uh, code releases and are starting to include some of these routing, uh, optimizations and things, uh, in their engine. Uh, and, uh, they also are very configurable, uh, to different situations. And so if you’re provider means you have a lot of documentation and support, [00:30:00] like just out of the gate, uh, because, you know, we’re not some like completely custom thing.

It’s like, well, you just apply the patch, uh, to VLLM and it basically operates exactly the same as vanilla VLM, except it gives you verified inference instead of unverified. So sort of, these are two of the approaches. And the other one I kind of already mentioned, uh, which is that we support, uh, streaming, uh, verified inference.

Uh, you know, to some degree, uh, there are, there are limits, uh, you know, on how many tokens you can get down to before that starts to become a poor, uh, experience. But you could think like 10 to 20 tokens at a time. Like if you have a really, uh, long answer, you’re kind of approximating the performance of these reasoning models anyway.

Uh, because like oh three mini high, you know, it’s spinning. It’s saying, I’m thinking about this, I’m thinking about that. And, you know, we can very much do the same thing, uh, just out of the box.

Tommy Shaughnessy: That is, those are awesome answers. Um, and I guess the reason to [00:31:00] fork Solana would be you don’t want non ambient related transactions on your network.

You sort of want to keep the fees low for what you’re doing.

Travis Good: Well, uh, it’s actually, there, there are a few reasons. Um, so first of all. Like philosophically, uh, ambient is designed to, uh, serve up, uh, machine intelligence at scale to the world. And so in order to do that, you have to create a very specific economic system.

And what you specifically need to do is be the best place for miners. Uh, so, uh, because, you know, miners provide your quality of service on your network and, you know, let’s be honest, they deserve the rewards. Uh, you know, they’re, they’re doing the work. Uh, and so, uh, in order to have a network, uh, in crypto economic terms that rewards miners of inference, miners of inference actually need to secure the network.

Uh, because that’s [00:32:00] how crypto works. Uh, you know, those who secure the network get the rewards. Uh, and you know, so in solana’s economic design, it, it’s a different thing. It’s, you know, it’s optimized to be like a stock market. And it’s not optimized to be like a minor reward platform. Um, and those are, you know, these are different objectives.

Uh, they’re both very useful. We’d like stock markets, uh, you know, in, in crypto to be super functional, and we’d like to be able to run all sorts of complex financial engineering problems. Uh, you know, on a crypto network. It’s awesome use case. And we’d also like to have, you know, a high scale inference provider.

It’s just that the high scale inference provider needs to be, you know, catering, uh, to the compute that is really, uh, bringing it, uh, you know, bringing it the quality. Uh, and, and so, uh, you know, specifically what we change about, uh, Solana is that, uh, we, [00:33:00] we change the definition of fake. Uh, so, uh, stake goes from, uh, you know, rewards for locking your capital up.

Uh, two, uh, validated problem solving contribution to the am ambient network on two different timescales. Uh, sort of a short timescale, uh, and a longer timescale. Uh, so what that means in, uh, Solana terms is if you are a big miner on our network, uh, you are more likely to be selected as a leader, uh, or a validator, uh, in ambient terms.

Uh, and if you’re a big minor, you are more likely to be selected as a member of a committee. Uh, that verifies inference. Uh, so, you know, it puts minors in the driver’s seat. Uh, it gives them the rewards, uh, it gives them, uh, the decision making power. Uh, and it preserves that sort of tower byzantine, uh, consensus model, [00:34:00] uh, that Solana has and the power and speed of that, uh, but makes it into a proof of work network by just tweaking one little thing.

Tommy Shaughnessy: Travis, maybe a, a basic question for you, but in finance in crypto, there’s always this notion of you want to be co-located, you know, the hedge fund wants to be close to the New York sock exchange. That’s right. I understand. Well, I don’t really understand all the technical reasons on the I, but I, but they track, um, is there a notion of me being closer to an ambient node than I am to an open AI node at a data center?

Or is that not, I might have missed that.

Travis Good: Yeah. So actually, you know, it’s helpful to talk about, um, like the other use case that we want to enable also. Uh, so, you know, there’s the, the hyperscaler, uh, use case. Um, but then we’re doing something a bit more, uh, custom, uh, to support, uh, the OG Bitcoin, uh, user use case, uh, which is that you wanna join [00:35:00] like a mining cluster, uh, you know, you wanna take your, uh, disco light, GPU.

And, uh, get on board, uh, and just, you know, show up and mind something. Uh, and, you know, traditionally that has been very difficult, uh, because it’s hard to spread large language models across, uh, many GPUs and have any sort of level of performance. But, you know, recently there have been, uh, some papers and, uh, research, uh, that has really, uh, I think, uh, opened the door, uh, for really performant, uh, inference, uh, in a distributed setting.

Uh, and so that is our, uh, next challenge that we’ve undertaken to make that type of inference, uh, competitive and cost competitive and, uh, timing competitive, uh, with, uh, some of the, uh, scaled up, uh, inference. Uh, and so then that relates back to your question because, uh, yes, you have the, the hyperscalers who are in data centers all over the world, uh, [00:36:00] potentially who are in bunkers.

You know, uh, who are in like, my, my, I I, my dear father, uh, I, I love him. Uh, he was really into Bitcoin and he rented out a police bunker, uh, that was not used in St. Louis and was running a Bitcoin mining operation.

Tommy Shaughnessy: Dead Family Dynasty is nuts. I, I, we could do a whole nother episode. This is,

Travis Good: I have this

Tommy Shaughnessy: picture

Travis Good: with my, uh, my brother and me in an elevator shaft of the police command center was, so he had Jeep used just in the bottom of that bunk.

Yeah. Like, they were bare, like it was just open and he had a, like all these wires that it, it was, it was pretty great actually. That is no. Does he still have it? Is it still? No, no. Well, we could talk about this, but the Bitcoin mining at that scale actually became unprofitable, which is a whole older, yeah.

Um, but, uh, at least with GPUs, right? Like everyone went to Asics. Um, but you know, the, relating it back to [00:37:00] the data center concept. There are two key things there. Uh, first of all, yes, absolutely we expect there, there would be closer nodes to you. And so there could be some, uh, latency optimization that goes on.

Um, but also, uh, like if you are in a part of the world where there are no GPUs and there are no data centers, uh, you know, you are not, uh, at the mercy of data centers, uh, anymore, uh, because there’s a whole distributed network that wants to take your business. Uh, and, uh, you know, that that should give you, like if the whole economy depends on machine intelligence.

Like that should give you a little peace of mind, is that we’re gonna build sort of a, a safety net for you.

Tommy Shaughnessy: Uh, Travis, one thing that went over my head, I just wanna clarify. So you are, we are focused on one model, but you, you mentioned that there will be specific fine tunes. Is that like per domain, per application, or per user?

And apologies I missed [00:38:00] it, but I wanna make sure we clarify. Well, you know, it depends on what

Travis Good: the network comes up with. Uh, you know, we’re gonna actually seed, uh, the network with some really interesting, uh, fine tunes that we’ve been clicking up, uh, which we think are a lot of fun. Uh, you know, I kind of, uh, share a view that I think you expressed on X, which is that, uh, large language models can be very boring to deal with.

The AI slot is a real thing. Uh, and so, uh, one of the things we’re doing is, uh, putting together some very opinionated and bossy models, uh, that will disagree with you, uh, and give you very specific perspectives and feedback on things, and they will disagree with each other. Um, and we, we think that, you know, it’s, it’s funny on the one hand, but it also gives people, uh, you know, a little bit of a canvas to paint on if we publish those techniques, uh, show how uh, models can be tuned in a different way, uh, that it will open up, uh, sort of a creative, uh, license, uh, for [00:39:00] others to, uh, create, uh, stuff That’s interesting.

Um, to, like, to get a little bit more, uh, specific if you’re talking about useful domains. Uh, you know, I think that, uh, a lot of like current jobs that require lots of manual effort are going to be automated. Uh, you know, different types of. Uh, you know, I’m gonna call it coding jobs. And a good, a good example would be insurance coding.

Uh, so not like coding coding, but where I’m looking at this thing and I’m like, what category does this fall into? And, you know, somebody is spending like an hour like pouring over documentation and photos and trying to say, is this, you know, does this qualify for catastrophic flood insurance or something?

Um, and like, I hope, and I think it’ll be to, you know, to the, uh, great benefit of everyone that all of those jobs go away and that people are able to do higher value [00:40:00] work than that. Uh, but you can imagine it’s great

Tommy Shaughnessy: to get rid of the menial work, right? Yeah. I don’t know why people are against that. It’s way better quality of life.

Yeah.

Travis Good: And you know, this actually, you know, this kind of goes to. The world that Ambient, uh, thinks about, uh, that we want to support that, you know, we think our network can help bridge to, uh, which is a world of abundance that runs like a Swiss watch. Uh, so to give you, uh, an example, uh, uh, of what I mean by that, uh, you know, I am a pretty modest traveler.

Uh, I, I think probably a lot of us are, you know, I book economy flights, uh, uh, fairly mediocre hotels. Uh, and you know, I, I feel like a slight pride when I do this because, uh, you know, I, I’m being, I’m being frugal. I’m, I’m spending on the things that are really important. Uh, but I’ll, I’ll tell you, like occasionally in my life, uh, people have put me up in really fancy hotels.

You know, it’s usually for a job [00:41:00] interview. Uh, and, uh, you know, this, uh, it’s a swank place that I would never pay for and that you would pay. Yep. Yeah. I, I, I just, I feel like this tremendous sense of luxury. But then the thing that I always notice when I go into the hotels is that the elevator is scuffed.

There are cracks in the mirror, in the corners. Uh, there’s dust, uh, there are fingerprints in places that probably should have been wiped down. And it always makes me think like there’s not actually even in places of abundance, enough human labor to make things perfect, you know? Uh, there’s always more polish that could be done, but there are not enough hours in the day for people to do that.

Um, but imagine a world where we had enough productive capacity to. Where, uh, robots could be polishing the heck out of all [00:42:00] the elevators, uh, like the customer service could be perfectly tuned on every dimension. Uh, someone could be thinking about solving your problem in the deepest possible way and working with others in a cooperative, cooperative fashion to solve collective problems, uh, in a smart way that no person could have ever thought of.

And if you build that kind of world, and I think that that world is best supported by ambient intelligence, am intelligence that lives everywhere and is not influenceable by, uh, special interests or, or different parties, like, if you build that world, uh, everything just becomes so much better. You know, uh, you can, uh, you could, uh, really create that, uh, sort of Star Trek, the next generation.

Feeling, I think for everybody just by focusing in on the details of every [00:43:00] decision of, uh, every little action, uh, that was taken in the world in a way that like people, people on their own could never afford to do, companies could never afford to do, but in a world of abundance, like properly directed, that becomes possible.

And that’s, that’s what we, we wanna build.

Tommy Shaughnessy: That’s such a great example. It’s, uh, I can’t understand tons of robots in AI working so hard on every little tiny thing to make life perfect. I, I don’t think I can understand that until experience it. You probably can, but it, it’s, uh, yeah, it’s tough. Yeah. It’s like, uh, it’s kind of

Travis Good: like, uh, what if we could turn every place into Disneyland, you know, di Disney pla Disneyland is like a, a place that every corner is scrubbed, you know, they’re very particular.

Um, but like, what if we could turn. Every corner of every street into something like that. And what if we could, uh, think at a [00:44:00] level where the contours of our national discourse and our communications among each other, uh, were pristine and at a level, uh, of perfection that Yeah. Is almost unimaginable.

Like that’s, you know,

Tommy Shaughnessy: Travis, the first hotels to get this are gonna be the rich ones. So you’re gonna have to break your rule. And,

uh, no, it’s a, it’s a great example. So maybe just going back, I have a couple nuanced questions for you before we get into the future. Sure. But, um, in ambiance, you know, bio marketing materials, et cetera, there’s a focus on the 600 billion parameter model, right? Mm-hmm. And in crypto, we deal with so many different governance votes, Dow decisions, way back when it was the Maker Dow interest rates.

Now it’s select improvement protocols. Take it back for you. How does the network decide, you know, hey, I wanna focus on a, a new type of foundational model, right? Or I want to expand the parameter [00:45:00] count. Like, how do, how are we gonna make all these decisions for the base model? It’s a

Travis Good: great question, and I think that the answer has to be in a properly decentralized fashion.

Um, and, you know, the drivers of this network are the miners. Uh, they’re the ones who have the most rational economic interest in the network, uh, being used and in being highly secure. Uh, and so, uh, we’re gonna give them the ability to vote on, uh, what model to adopt. Uh, you know, initially our thought is that that’s gonna be like a multiple choice vote.

Uh, you know, that’ll kind of come up. Like, Solana had this economic change vote, you know, I, it was, it was voted down. But, you know, that’s a good example, right? Like, uh, you’ve got two options, yes or no. Uh, and, uh, someone can vote on that. And so initially, you know, we will curate the options. We’ll say, all right, there are only, you know, maybe four open weights, models in the world that we could all adopt, uh, that would be competitive, uh, with the state of the art.

Uh, like, here are the [00:46:00] options. What would you like to do? Uh, and people will vote. And then, you know, just as when you have a validator upgrade in Solana, uh, there’s a scheduled downtime, uh, there’ll be a scheduled downtime. Uh, the nodes will go down. Uh, the model will be changed, the nodes will come back up.

Uh, and we’ll be in a slightly different world.

Tommy Shaughnessy: That is pretty cool. It, it’s, um, it’ll be interesting to see if these hardware guys agree with the AI experts, right? Mm-hmm. That’ll be interesting because there has to be some communication between, Hey, look, this is a literal leading edge model. We want to change something in ambient, but I know this might impact incentive short term.

I don’t know how it’ll play out, but it’ll, it’ll, yes. Yeah. Agree.

Travis Good: No, I think you have to, um, you have to really be candid in your communications about the pros and cons and, uh, you know, probably there’s an element of the whole community expressing its preferences as well. [00:47:00] Uh, just so that the, the miners get grounded, you know, when they ultimately vote, uh, in what the community’s interest really is.

I think those are really important points.

Tommy Shaughnessy: So, Travis, two more quick questions to, to make sure we cover as much as we can about ambient before the fun future stuff. You’ve mentioned verified inference a lot through the podcast. Yeah. And I haven’t asked you why that’s important. Why is verified inference such a core part of Ambient?

Travis Good: Because composability is fundamental to crypto. If I want to, uh, build useful things on crypto, I need to have a trusted way of putting the pieces together, a trustless way of putting the pieces together. I. And if a large language model inference is going to be one of those big pieces, uh, in the future of crypto ai, uh, then I need to make it trustless.

Uh, which means that the inference must be verified. Uh, because if you use unverified inference, uh, then, you know, there are all sorts of shenanigans, uh, [00:48:00] that can be pulled. And we’ve, we’ve seen them in non LLM world with logic exploits. Uh, but you, you know, it’s almost unimaginable what will be accomplished, uh, in unverified LLM model world.

Tommy Shaughnessy: Do you think that there has been, um, weird or funky things with inference behind the scenes of the web two labs that we cannot view or understand or see and like, what would they be? Uh,

Travis Good: so, uh, I, I do, and I think it’s a huge problem, and I’m gonna relate this to software deployments because ID I’ve done different types of software deployments my entire career.

Uh, and, uh, you know, when you deploy a piece of software you normally do, particularly one that’s safety critical. You know, I, I used to do like large scale safety critical system deployments. And so when you, when you deploy those things, um, you want to do unit tests and integration tests, uh, you want to make sure that the [00:49:00] individual little pieces of, of functionality are checking all the boxes that you want.

And you want to make sure that the different architectural commo components are communicating with each other in the expected way and producing the expected results. Uh, and, uh, when you insert an LLM, uh, into the middle of something like this, uh, you immediately run into some severe problems. And, you know, one of those is related to the non-determinism, inherently of the output.

Uh, you know, there’s just a fair amount of randomness baked in if you, even if you’d set the temperature, uh, to, uh, zero. Uh, but another, uh, problem is, uh, that, uh, it’s very hard, uh, to predict, uh, what will happen when you give a novel series of inputs, uh, to the LLM itself. Uh, and so, you know, it’s, uh, because of this phenomena, uh, because [00:50:00] like, that’s already very hard.

Um, closed source model providers are absolutely causing chaos when they deprecate models really fast. So, you know, GPT-4, it lasts, like the, the version that we care about lasts like a year. As a software shop, like I’m trying to figure out all the behaviors of these models in different conditions, and if I get it wrong, my car dealership might sell someone a Chevy for $1.

Oh. And there’s just no way in hell that I can test all this stuff at the Cadence it’s being released. And so, and they’re not being transparent with us about the change in the behavior of the models either. They’re just like, oh, it’s smart or not. And it turns out like it doesn’t answer any political questions anymore.

Uh, or like, oh, like the model is better at problem solving, you know, except for in your coding use case where it absolutely face plants every time now, uh, you know, and, uh, so on a very like, visceral [00:51:00] level, like they’re absolutely changing everything all the time without being properly transparent. And I think that that is really dishonest and also dangerous, uh, when your economy, uh, depends on these models, uh, because people feel a certain amount of pressure to keep up with the Joneses and to switch, uh, and, uh, the moment they switch, uh, they’re not incurring risk for open AI or anthropic.

Uh, they’re incurring risk for their business. Uh, so I think that’s one whole category where, yes, some shenanigans are being pulled, uh, but then there are some subtler things that people over time have really picked up on. And, and those include the fact that, uh, the models are being dumbed down, uh, silently, uh, for no good reason other than the provider’s infrastructure is a melting down, right.

Um, you know, if you’ve used Claude, uh, you have probably gotten the responses temporarily abbreviated due to high [00:52:00] load, uh, message, uh, which is honestly better than it used to be. Because you used to just get stupid responses and wonder why, uh, and, and now you get stupid responses. Uh, and you know that it’s not just that, uh, the model is under heavy load and they’re shortening what’s going on, but actually if there’s thinking going on in that model, they’re shortening the reasoning traces associated with that model as well.

Uh, and that is causing the model to be noticeably dumber.

Tommy Shaughnessy: So you could, you could just get, be way smarter if you use the models at the right time of day. I mean, I don’t know when that day time of day is, but,

Travis Good: well, you know, I’m, I’m a strange night owl. Like I, I’ll stay up till two and three in the morning.

And boy, those models are smart at like 2:00 AM

Tommy Shaughnessy: They’re, they’re job at 10:00 AM Eastern. Yeah, they’re, yeah. Yeah, I, I was trying to fix our Delphi bot I built to just handle inbound plays, and it was hallucinating on [00:53:00] the number of plays it reviewed. Mm-hmm. Which I made a tweet weekly, a very basic thing, and I asked XJDR how to fix it.

And he is like, just lower the temperature and such an easy fix for the Yeah. Just such, yeah. It’s less creative, but, yeah.

Travis Good: Well, so I mean, that’s a major problem is this dumbing down on models, like, without being transparent about that, but there’s actually a final class of problem, which is related to all the censorship, uh, that they have put into these models and the way that they silently change the censorship, uh, which actually breaks use cases.

Uh, so you get these patch notes on versions of Claw that are like, oh yeah, like more responsive to queries now. Um, and it’s like, oh, like, uh, what actually changed in the behavior? Like, which queries are, is it more responsive to. Is it going to respond in what you and I would think of as a truthful way now?

Uh, or is it gonna hedge in what we might think of as a deceitful way? Uh, like

Tommy Shaughnessy: that is a, [00:54:00] it, it’s crazy ’cause I always think of censorship as that system prompt when you ask it a question that lives in the shadow, in the invisible realm that, that tells me that Ken and cannot do. I also think of censorship in the data that goes into train the zeros and ones.

You’re talking about a totally different level of censorship.

Travis Good: Yeah. It’s, uh, the kind of censorship where the model is actively trying to prevent you from getting into whole spaces of thought. Uh, where like, Claude has a, it has a morality built into it and if you try to lead it along certain, uh, perfectly valid, uh, philosophical lines, uh, it will try and steer you into another zone.

It’s kind of like, uh, in those, in the old video games. Like, if you would walk to the edge of the map, uh, they would like turn you around. Um, it’s, it’s like, that’s the best analogy I can think for it. Um, but it is, uh, [00:55:00] but, but

Tommy Shaughnessy: why though, like, why wouldn’t they let you go off? Like what is the downside for philanthropic saying, Hey, go explore the matrix.

Like, why don’t they let you? Uh,

Travis Good: that is a great, great question. I, I think there are probably a couple of reasons. Um, and I don’t wanna get myself into too much trouble. Um, but I think that the world has embraced a culture of safety, uh, where, uh, it is dangerous for people to even encounter certain ideas.

Uh mm-hmm. So, uh, you know, we used to think that, uh, you sh you know, shine the light of truth on bad ideas and they evaporate. Uh, and there are a lot of, doesn’t affect Yeah. It, it, it disinfects. And you can argue against them, like it’s very easy if you expose them. Uh, but I think that, uh, you know, the culture has shifted in a way where, uh, these companies have this [00:56:00] almost paternalistic, um, uh, contempt, uh, for their users.

Uh, they think I need to protect you from dangerous thoughts. Uh, and, uh, you know, that actually makes us all stupider, uh, because then we can’t, we can’t reason about dangerous things if we never encounter them. And, you know, the reality is, yeah, life is full of dangerous things. Uh, and like you wouldn’t want to, uh, try to learn to ride, uh, a bicycle, you know, without, uh, like, uh, or go down a highway on a bicycle, uh, without ever having fallen on a bicycle in your neighborhood.

I.

Tommy Shaughnessy: I mean, imagine never seeing a gun and then seeing it for the first time. Like you don’t, that’s life or death. Yeah. Yeah. And like if you’re just using some of these

Travis Good: image gen models, you might never see a gun, you know?

Tommy Shaughnessy: True.

Travis Good: Uh, and so I think that [00:57:00] there is this patronizing moral attitude that, uh, we have a better idea of what morality is than you do, uh, and we are going to protect you, uh, from all the bad things in the world.

And I think it’s one of the insidious dangers of, uh, corporately controlled LLMs because, uh, that attitude, while extremely frustrating right now, can morph into totalitarianism very quickly. You know, all of a sudden dangerous ideas start to include, uh, you know, opposing the existing political order or the economic interests that are sponsoring the LLM today.

You know, you can imagine like Microsoft Azure chat, GPT sponsored by Pfizer, like it’s gonna do deep research, but is the deep research gonna return, like any results about adverse reactions to the drug that you’re taking? Uh, like, maybe not [00:58:00] that sure would be helpful to know.

Tommy Shaughnessy: Don’t, don’t research a new drug if it conflicts with one of our main selling drugs.

I mean,

Travis Good: that’s right. It’s so easy. And, you know, we’ve seen, we’ve seen the commercial incentives, uh, like skew our news media. Uh, we’ve seen them bend all sorts of institutions, uh, in really unpleasant and unhealthy ways, uh, for the populace. And I think that, uh, the same thing could happen, uh, with, uh, LLMs in, in really scary ways.

And, and so that’s why, you know, I would, uh, just offer an appeal, uh, to the people who are out there, you know, thinking about something like ambient. Like, please come on board with us, you know, advocate, advocate for this kind of decentralized model, uh, so that we can make a better world, uh, together.

Tommy Shaughnessy: Well, one of the, one of the reasons I’m so happy for you is because you’re, you are building at a time when the [00:59:00] world isn’t yet, all the apps aren’t yet powered by these AI models, right?

That’s, there’s still a chance that we can build the model on ambient, right? Mm-hmm. Um, yeah. And it’d be a lot safer. Um, well, not safer. I, yeah, it’s now, now we’re getting in a weird vocab, I guess, uh, more democratic or

Travis Good: open. It’s, uh, it’s more free. And, you know, that comes with, that comes with upsides and downsides, but.

At the end, like honesty is really important in society. And if you can’t trust anything, uh, that is presented to you, uh, you start to live in a low trust society and that has really negative impacts on your quality of life and on your civilization. And so, you know, I think that creating a network like Ambient, which is censorship resistant and which is designed to be economically, uh, competitive, uh, with the players that could dominate, is what creates a wedge alternative.

Uh, that [01:00:00] prevents us from getting into the totalitarian dark futures, uh, that are possible. Uh, but there always has to be an alternative or, uh, those players are going to exploit and try to control as much as they possibly can. Uh, and yeah,

Tommy Shaughnessy: no, no, I, I, preaching to the choir, I totally, totally agree. One, um.

One question I have for you is taking the behemoth open ai, um mm-hmm. We can talk a lot about a lot of different ways we can take this, but OpenAI is, I think Sam, uh, mentioned they may open finally open source some model. Right? Right. Um, if is there, let, let’s like play the other side. Is there a risk that open AI ever goes full tilt open source?

And if they do, do you think that would negatively impact Amio? Um, so,

Travis Good: uh, that’s a great question. [01:01:00] And, you know, I, I think I have a few different answers to it. Uh, like I, I, we should play the Steelman version of this where they do and we can talk about it. Um, but let us look at the history of it. Uh, so if you look at the emails, uh, that they sent back and forth.

You know, I think there was a conversation between Sam and Ilya at one point where it’s like, we’re going to use open source as a recruiting tool, uh, but we’re never going to actually give the technology to the public. You know, if I’m badly paraphrasing this, maybe, but there’s one of these emails that feels like a smoking gun to me, uh, where they were just, um, I think abusing, uh, the community, uh, to be quite honest, uh, to generate enthusiasm for their product in a very cynical way.

Uh, and that’s baked into their DNA, uh, you know, that level of cynicism, uh, like it or not, exists at the founding point, uh, of the [01:02:00] company. And the company still has the same leader. So if you’re a, if you were to ask me, like, do I believe that, you know, Sam Altman has had a, like a wave of benevolence, like wash over him and, you know, um, like, uh, I think he’s, uh, you know, an excellent capitalist.

And he will do some things for temporary advantage. Um, but he will never, uh, give a true economic advantage to other parties. Uh, because for, for someone like that, that would be to seed the field, uh, to the competition, uh, he will always maintain the strongest possible edge, uh, for open AI while trying to create a narrative that is palatable, uh, to the public.

Tommy Shaughnessy: I think that’s a great way to put it. The it, and this is a great, a great segue if you have time, but like when we first did a podcast, you wrote a phenomenal long form response to Leopold’s situational awareness post called situational blindness. I’ll link it in the show [01:03:00] notes. And Leopold argued a bunch of things that.

Um, increases in computational power and, um, un hobbling and algorithmic improvements would mm-hmm. Increase the, the power of these AI models, yada, yada. Talked about web two and scaling and data and hardware and everything like that. Um, since then you’ve made considerable progress with ambient open. AI has made insane progress with everything they’ve done, but we also have open source, we have deep seek.

We have Quentin, we have Mattis, we have, take your pick. Have your views changed at all from your long form repost on situational blindness? And I know we can’t go into the whole post here, but I’m curious if anything specifically changed in your view since you wrote that post.

Travis Good: It’s a, it’s a great question, Tom, Tommy.

I, I appreciate it. Um, you know, let me, let me go over some of the assertions that, uh, Leopold made. Uh, so, uh, he said we live in a bipolar world. Uh, it’s the US and China, uh, and, uh, they’re [01:04:00] in, they’re locked in economic competition. He said the only expedient thing to do is to lock down all AI models, all AI research and form a Manhattan project, uh, that is devoted to defeating China.

Uh, and, uh, we should clear the way, uh, for everyone regulatorily, we should, uh, favor the incumbents. Uh, we should declare open AI specifically. It might have had a little bias there. Uh, we should declare them the winner of the AI race, and we should essentially give them the keys to the kingdom along with Microsoft and say, you know, go build as many coal-fired or nuclear power plants as you possibly need, uh, to, uh, compete with the Chinese.

Um, and, uh, he very strangely, uh, justified this, uh, in the name of democracy. Uh, and, uh. You know, interesting. It, [01:05:00] it’s a, it’s a bizarre way to look at the world, uh, that you have to empower the most secretive and undemocratic institutions in a society in order to

Tommy Shaughnessy: achieve democracy with the totalitarian lock-in that comes from a single model winner.

Travis Good: That’s right. Um, and, uh, it is been immensely frustrating for me, uh, to watch, uh, people, you know, who have nominally democratic, uh, you know, principles, uh, support this stupidity, uh, because it is just plainly stupid. Uh, and anyone who thinks about it for 10 minutes should be able to come to the conclusion like, this guy is a very smart idiot.

Uh, you know, he’s, he’s right about certain things. He’s right about the scaling curve, uh, to some degree, although we can talk about that. Uh, he’s, he’s right about some of the un hobbling things, although he doesn’t do a very good job of explaining why, uh, which I think is kind of crucial. When you come in with strong claims, you need to sort of [01:06:00] justify those.

Um, but he’s very stupid about the implications of what he’s doing because if we followed his plan, we would end up in a totalitarian hellscape. Um, and you just have to follow his own assertions to their logical conclusions, uh, which is what that essay does, uh, to come to that. Uh, so, but then let you know.

Let’s sort of take that and let’s reflect on that. What has changed, uh, in the world since that essay was written? Um, and I think that, uh, what is maybe becoming clearer. Is that secrecy is not the biggest weapon. Uh, openness is actually the biggest weapon. You know, Nvidia, when Deep Seek was released, experienced this horrific plunge, uh, because all of a sudden people thought NVIDIA’s not gonna be selling as many [01:07:00] data center cards now.

Uh, and you know, I, I personally think that’s, that’s probably healthy. Like, I don’t think that one company should, you know, dominate the entire sector of the economy. Uh, uh, and, uh, but, you know, uh, people were aghast about this and, you know, so it turns out that, uh, Leopold’s fundamental assumption was that the, uh, form of economic attack was going to be secrecy and, uh, competition in secret.

Uh, but actually, uh, it turned out, uh, to nullifying advantage. Through open source releases like that, that was the power move. Um, and that China pulled, yeah, that China pulled. And I think what that did, uh, is it changed the patterning for the world because now, you know, different groups in the world, like, you know, to give an example, uh, [01:08:00] minstrel, uh, stopped kind of releasing models for a while.

Uh, uh, you know, they juiced their stock price, uh, their, whatever their internal stock price was, the valuation. They juiced their internal valuation. They got a lot of funds from the European entities, and then they kind of shut down their open source program for a little while there. Um, and uh, then they came roaring back recently with like, uh, this my small model, which is awesome.

Uh, and I think what they saw is that, uh, you can lead through inspiration. And, uh, if you inspire the most people, and if you give the most, uh, then uh, the world might come along with you. Mm-hmm. And, you know, it’s something that, uh, I think would be predictable. If you think about how business relationships go, like you might, you might meet someone at a conference.

The person you stay in touch [01:09:00] with is the person who’s offering you some value in that business relationship. You know, maybe they have some crucial perspective or insight, uh, that they’re providing to you for free, essentially. Uh, and when they do that, like you feel positively, uh, about them, uh, and like they may influence you, they might change your mind about some things because they’re giving you a gift.

They might change your mind about some unrelated things, uh, that, uh, you know, you hadn’t really deeply considered before. And, uh, I think that what we’re seeing is hard power versus soft power. So in a world where the US is abandoning its soft power, uh, China is embracing a different kind of soft power.

It’s the soft power of open model release. Uh, and Leopold’s premise is based on hard power. Um, [01:10:00] but I think he got it wrong. Like the, the most powerful thing is soft power, uh, when exercised carefully. Uh, and to be clear, like I think that, uh, you know. That’s not typically how a lot of countries act like, I don’t think that countries play the soft power game, uh, very well.

Uh, but I think that they’ve discovered in this arena that model releases are soft power. Uh, and I think that that really screws up Ashton Brenner’s calculus, uh, because now you’re gonna have, uh, countries jocking for soft power via model releases. Uh, and I think that’s awesome. Like, I think that’s really healthy.

Um, but you know, if you were asking me if I changed my opinion, uh, I would say that, uh, I, I don’t think his argument has gotten any less wrong. I don’t think the implications of his argument, uh, have gotten any less dire if people were to follow [01:11:00] his assumptions. Um, but I do think actually that the technology picture, um, that I was thinking about has gotten significantly less bleeker.

Less bleak, um, you know, in the, in interval because, you know, at the time that I wrote that it was very unclear if anyone was ever gonna catch up with open ai.

Tommy Shaughnessy: China didn’t even have a, did they have a great open source model when this was written? No. Right. So much.

Travis Good: And, uh, you know, I think, uh, one of the big funds had just released a piece about how, uh, you know, models were going to be like mainframes forever, and we would all just have to rent from them.

I don’t know if you remember that. I,

Tommy Shaughnessy: I forgot. Yeah.

Travis Good: Um,

Tommy Shaughnessy: the crazy part, um, also with Leopold’s argument is that. Um, he argued, I think that we’re racing for like this trillion dollar cluster, but the reality is like deep seek, like they really innovated, right? [01:12:00] Like they use old generation hardware, they only turn on parts of the model mixture of experts.

They did this absolutely nuts. If you read Sam Lehman’s posts innovation on how to train their new reasoning model. So like to think that we’re just competing in this just direct linear fashion Yeah. With one model type in architecture, it just, it, that like, just doesn’t make sense. I, I

Travis Good: totally agree with that.

And I think that there’s going to be a lot of, um, innovation in the field, you know, both in software and in hardware. And, you know, one of the really important things for, for Ambient when we were looking at the design of the system, uh, you know, is that it was future proof. So, you know, we designed our algorithm in a way that, uh, yeah, the, the model architecture could change and, and the verification would still work.

The underlying hardware, uh, could actually change significantly and things would be totally fine. Um, and we expect the hardware to change. Uh, we expect the model architectures, uh, to [01:13:00] change. And I agree with you like this idea that, uh, one group had cracked the code for a field as broad as, uh, machine learning is very arrogant.

Um, and, you know, possibly only explainable by, you know, someone being in their early twenties and only ever have really having really worked at one company.

Tommy Shaughnessy: No, I, I’m glad it. At, at the time I was happy with how forward looking his article was, but I, I’m more so happy with how, how correct your, your response was and how solid that I, I recommend people, um, go look, look at it.

Um, Travis, we have like 15 left. Um, yeah, we use it. We use it. But one, one of the other things I want to ask you, I know we’re jumping around a lot, but there’s just Sure. So many different things to cover, um, in that world that you’re describing where, where these countries are releasing open source models mm-hmm.

Um, maybe I’m being overly bullish here, so you could back, you could agree or disagree, but do you think that there’s a world where they say, Hey, look, you know, we don’t have [01:14:00] the expertise to do this. We don’t have the hardware, the energy, the data, whatever. Let’s buy ambient. Let’s let, let, let’s join an open source alternative that we can just buy the asset of.

Like, do you think that that’s a viable sell to these countries? Right.

Travis Good: I do actually, because I think it can be a meaningful expression of soft power. Uh, you know, if you have countries that band together in the world. Uh, to contribute the, to the, uh, pre-training of an alternative model, uh, you know, who become part of, uh, a global expert discussion about, uh, how to improve, uh, collective capabilities.

Uh, I think those kind of forces are always gonna be more powerful than unilateral action on the part of one country. And I think a balance of power in the world and a spreading of wealth around the world. Uh, and there’s gonna be plenty to go around because I think that we, we can have abundance, uh, with these technologies, uh, is, [01:15:00] uh, what we need to have global peace.

Tommy Shaughnessy: That is a good point. And then unrelated, but the usage of ambient, the on chain nature versus the off chain nature, right? Mm-hmm. The privacy aspect comes into play, right? Yeah. If I type a response into ambient, is that on chain? Do I have to worry about what I’m, what I’m typing into an ambient front end?

Oh,

Travis Good: yeah. It’s

Tommy Shaughnessy: a, it’s a great

Travis Good: question. Um, so no, uh, so the way it’s set up is, uh, it happens via an auction. Uh, and on the auction you say, I have this query that’s this long and I need it back in this amount of time, and I would like to pay 30 ambient, uh, for that. Uh, and then miners come in and in order to make a bid, a miner puts, uh, some money in escrow.

Uh, so we don’t have civil attacks and spamming, uh, and, uh, they say, Hey, I can do your query, uh, for, uh, 15 ambient, uh, and, you know, 20 seconds or whatever. Uh, and, uh, the second lowest bid wins is kind of a reverse, [01:16:00] uh, auction. Uh, and, uh, you know, what that, uh, enables is a kind of economic rationality in the sorting of queries to different miners.

So, uh, and that helps, uh, dynamically adjust the price, uh, to match the actual value, uh, that’s being delivered. Um, but what it also does, uh, is, uh, it’s a mechanism for, uh, obfuscating, uh, the source and nature of the queries, uh, and only revealing those, uh, to the miner, uh, and the validators of that inference.

Uh, so, uh, you know, if you’re the miner, uh, you’re necessarily going to see the query and its output, uh, but you’re not gonna know where that query came from. Uh, and that query is gonna be in an economically similar block of queries, queries that are all kind of the same links that need to be, uh, delivered in the same timeframe, uh, that, uh, you know, is, is an unintelligible [01:17:00] jumble, uh, to you.

Um, moreover, uh, in times of low traffic, and there are those on every network. We’re mixing, uh, programmatically generated system queries into the blocks. Uh, so it’s gonna be really confusing. Um, and you know, our model is essentially like the same security model as Run Pod. Uh, like, you know, you, you show up on Run Pod, you run different operations, uh, yeah, like the, um, the user could, uh, the operator of that server could, uh, look at individual things you were doing on there.

Um, and because they have control of their, uh, machine. Um, but what we think with is a little different in LLM World is that unlike in run put, uh, where you have observation of someone’s continuous, uh, usage, uh, it’s very, very difficult, uh, to build any sort of useful profile of someone. Uh, if you don’t know where the queries come from.

If the queries are [01:18:00] all, uh, mixed, uh, together, uh, and if you can’t do any flow on the queries to correlate, uh, the queries to each other. And so we’re, we’re taking measures to address all of those pieces. Uh, and we think that, uh, that achieves effective, uh, anonymity. Uh, and, uh, you know, it’s going to be, it’s gonna make it, uh, easy, uh, for, for people to, uh, adopt the network because they can have, uh, they can have confidence that the design of the system, uh, makes it incredibly difficult to glean anything useful from any particular query.

Tommy Shaughnessy: That is awesome. No, it’s great to know that you’re approaching it from three main angles to maintain that anonymity. And you, you mentioned like 30 ambient, 15 ambient. I’m trying to think through a world where I price intelligence and. Ambient, right? For the small inferences, the little, the little asks, you know, what, what the sky is blue?

[01:19:00] What’s five plus five? Those I won’t really worry about. But do you think we’ll ever get to a world where you and I are sitting across from each other and you’re saying, Hey, Tommy, you can go start a business with 10,000 ambient. Do you think we’ll get to a point of that?

Travis Good: I do. Because I think that, uh, you know, to to, to make a fun example of this, uh, let’s say that, uh, you wanna start a coffee shop, let’s make this very concrete.

You remember, love coffee. Yep. Yeah. Like with Bitcoin, it was really hard, uh, to buy a coffee or a pizza, you know? And like, you probably have regret for the rest of your life if you did that, because it’s worth now like $300 million. A million

Tommy Shaughnessy: dollars.

Travis Good: Yeah. Um, and, you know, but the, the thing about ambient is that unlike, um, Bitcoin, uh, you can set up a, uh, sort of perpetual motion machine of value creation.

And in order to do that. All you need to do is make it so that, uh, your [01:20:00] intelligence profit, uh, exceeds your intelligence costs on an ongoing basis. Uh, so, you know, let’s look at the coffee shop. Uh, I have a coffee shop. I have some physical fixed costs, uh, you know, that are probably unavoidably in fiat, uh, at some level, uh, like, you know, my place and my machines.

Uh, but then the rest of this, um, might be denominated in Abient. And, and the reason for that is like your whole back office is a series of agents. Uh, your accountant is an agent. Um, your bookkeeper is an agent. Uh, you have a coffee supply chain agent. Who is negotiating, uh, the best prices for the best brews.

And, you know, reading hundreds of reviews of different brews and looking at chemical analyses of, you know, product, uh, and, you know, picking things that are gonna align with the exact preferences of the type of people who are coming into your shop. Uh, and, uh, you might also have [01:21:00] a coffee barista robot. Uh, and, uh, you know, if you’re an upscale place, uh, it might be conversational and that conversation, uh, might be powered by ambient, uh, because, you know, people like having a fun, intelligent, uh, interaction.

Uh, you know, when they, uh, go into the coffee shop, maybe, uh, their, uh, barista has a particular personality and is quippy. Uh, and that’s, uh, that’s powered by ambient. And so, uh, you know, we could go on and on, but. Uh, it starts to look like the entire cost basis of your business is denominated in the machine intelligence that really powers it.

And the question you find yourself asking is, if that’s the majority denomination, why would I do any conversion? You know, why not just deal directly in that? Uh, and that’s, uh, what ambient aspires to be is that facilitation. You know, that that currency that’s facilitating that kind of, uh, direct economic [01:22:00] exchange, not an abstraction on something, someone else’s value proposition, uh, but a direct way to act upon the world using the dominant economic system of the world, which is machine intelligence

Tommy Shaughnessy: and the, the robots themselves on the network between each other.

These agents, it’s actually an extra step for them to go from a. Ambient to dollars, right? That’s right. Like if you’re talking like, why would they, like they’re talking, they say, Hey, I need you to do this job. It’s gonna probably be this many tokens, this much time, whatever. Why would I then, why would they then do the extra step?

It’s a small step, but why would they, it’s extra.

Travis Good: Yeah. Well, it’s extra. And if you do it, you suddenly become beholden to the powers that be. So payment processors are notorious, uh, for enforcing their will. There are whole categories of businesses that simply cannot exist under payment processing [01:23:00] regimes.

Uh, and they make it hellish up until recently and still, uh, for, uh, you know, people to get into crypto. Uh, right. Like the KYC can be ridiculous. Uh, it can, the whole thing can feel encumbered and, uh, they can rug pull you at any moment. And, you know, the, the crypto onboarding firms just have to deal with.

Incredible amounts of paperwork in every country, uh, to be able to function. And so yeah, anytime you enter into fiat, uh, it’s probably bad for you. Uh, and we should avoid it. Uh, and yeah, I, I, I think Ambient enables you to do that in a world where machine intelligence is what provides most of the economic value.

Tommy Shaughnessy: I don’t know if it was Vitalik or Anatoli or, or somebody else, I apologize, not remembering whoever you are, but they said like the whole startup world under $1 transaction fees never, definitely Anatoli probably never happened, right? Just because of those costs. [01:24:00] Um, so it’s just cool to see what could happen.

I mean, for you it’s under what, a millionth of a penny or something for these robots to talk to each other. But yeah, um, it’s pretty cool. So Travis, just to close out, I want to get your overall thoughts on. Um, when we talked, like, you know, a long time ago and when we even started on the crypto AI side, there was a pretty considerable leap on this all starting to pan out, right?

Mm-hmm. Like we didn’t have this whole concept of global reasoning traces in a distributed manner, improving a model, right? We didn’t have this notion of owning the piece of a model because everyone said there’s no moats and you you’ll own something and it’s worthless. We can copy the weights. It seems like the whole economic totality of open source, decentralized crypto eye, whatever you want to call it, is really taking shape and ambient, sort of at the forefront of like just capturing that image.

Are you, is there certain aspects that you’re much more confident under the economic lens of ambient being [01:25:00] successful for the user, for the minor, for the investor, for the person that wants to build an app? Are you feeling much better about that now than let’s say a year ago?

Travis Good: Uh, honestly, yes. Um, because.

You know, I think, uh, like everybody, I was wondering, uh, what economic leaps we would need to take in order to deliver the same value as the closed source, um, providers. Um, and, uh, you know, what we aspire to do, what we’ve been talking about is to be, uh, the biggest and the best open, decentralized provider of machine intelligence, uh, to everyone, uh, to uh, to web two, uh, to Web3.

Uh, there are gonna be lots of competitors. There are gonna be lots of different, um, approaches to the economics of this. Uh, but we, what we believe in the beginning was true, we still think holds very [01:26:00] true. Uh, which is that. People always need recourse, uh, to a big open weights model, uh, that is economically and functionally competitive with the closed source options.

Uh, you may reach for other tools. Uh, you may do lots of other interesting, uh, cases, um, but that bread and butter sort of needs to exist, uh, in order, uh, for, uh, you know, the economy to function smoothly. Uh, you just have to be able to reach for that. You can’t, um, denominate yourself, you know, like a lot of the early, um, x.com agents that we were seeing, you can’t be using open AI for those things, uh, forever if you want to maintain credibility as a decentralized uh, solution.

Um, but overall, like I think, yes, we were all. Probably very early to this trend. Um, and there were a lot of, uh, leaps, uh, of faith and, uh, you know, uh, much credit to you [01:27:00] guys on the, the Delphi, uh, side. Um, I felt like we were speaking the same language almost immediately, like five seconds into the conversation, we were reading the same things.

We were, uh, using the same vocabulary. Um, it felt very fluent. Um, and I think that, uh, we’re still early and that for everyone who’s getting into this, uh, you know, it’s not too late actually, uh, to come on board. This is the, this space is the most fun, and I think it’s gonna just continue to be the most fun, uh, probably for like 20 years.

Uh, so yeah, join us.

Tommy Shaughnessy: No, I’m, I’m totally with you, Travis. What’s the best way for, you know, somebody listens to this, they really agree with your view. They agree with ambient like. If they’re a hardware provider, if they’re a tech genius, if they wanna leave OpenAI and come over, like, what’s the best way for everyone to get involved with Ambient?

Travis Good: Yeah. Uh, well, I’ll answer that in, in part. So, uh, [01:28:00] for, uh, the folks who want to develop applications, uh, on Ambient or who want to even develop, uh, you know, applications on another chain that need a really good verified inference solution, uh, that’s cost competitive with other providers, like, please, uh, you know, send me, send me a message.

Uh, if you look at our website, which is, uh, ambient xyz, uh, we’ve got an email address, uh, hello at ambient. Uh, and, uh, we read all those emails. We’ve responded to hundreds of them, you know, in the past, uh, a few days, uh, in the wake of the announcement. So, you know, builders are, uh, welcome. Uh, if you’re someone who has been sitting in big tech purgatory.

And you would like to join, uh, this, this side of, uh, the equation and help balance out, uh, some of the oligopolistic uh, tendencies, uh, that exist there. Uh, we welcome you. Uh, we love research collaborators. [01:29:00] Uh, we love, uh, skilled programmers and machine learning experts. Uh, you know, we’re a bunch of hardcore techies and, uh, you know, we, we respect fellow hardcore techies.

Uh, so, you know, we think you’re gonna be in, in good company. Um, if you are someone, uh, who, you know, feels the same way that we do, uh, that something like ambient, uh, is necessary, uh, you know, we would ask you to, uh, help advocate for us. Uh, you know, we’re, uh, in the scheme of things like pretty small. Uh, we’re just at the beginning of this.

Uh, and it can make a huge difference to us. Uh, if you support us, uh, if you talk about, uh, ambient goal, uh, which we think is returning to the roots of crypto and delivering on the promise of a world currency, uh, if you, if you believe that also, uh, we appreciate your advocacy and your public support, uh, because [01:30:00] that’s the only way, uh, that we can build critical mass and a movement to counterbalance, uh, some of the other forces, uh, in the world.

Um, and, you know, we, we wanna interact with you more, uh, as a community. We’ve, we’ve just come out of stealth and we wanna build this with you. We want to build something that serves your interests and, and a currency that ultimately serves the world’s interests instead of the interests of a few.

Tommy Shaughnessy: Thanks so much, Travis.

It, it, it’s just, um, it’s just such a more open vision, right? Like non-investment advice. But if I have, I. A couple of bucks abroad and I wanna invest, or if I want to build an application or if I wanna do a fine tune of your foundational model, which you really can’t do with a closed source model like that, like the whole stack is just open for every value creator to come in, which is just huge.

So Travis, thank you so much for coming on the show for the third time, and thanks for having us as investors. We’re extremely excited for, uh, for ambient. [01:31:00] Thank you so much,

Travis Good: Tommy. I as always really appreciate, uh, the perceptive questions, uh, and your understanding of the space. Thank you so much, Travis.

Tommy Shaughnessy: See you soon.

Delphi Oracle

Latest on Delphi