travis oliphant - ceo @ openteams.com

Welcome to another— we have Travis Oliphant for round 2 at the podcast. Yeah, re-invited Travis for a reason, super interesting guy. Uh, just as a reminder, creator of NumPy, SciPy, Numba, uh, some coding languages I used to use when I was learning Python, I guess. Uh, Travis connecting companies with communities, so he's CEO and founder at OpenTeams. He's also chair founder at QuantSite and head of open source at Zyphra. Travis, welcome to PODS. Tell us a bit more about yourself and what you've been up to.

Hey, awesome to be here. Thanks so much, Charles. Yeah, been up to a lot, it feels like, and sometimes not as visible as other times in my life. But although recently I just announced a new project where I'm experimenting with coding agents to essentially recreate the foundations of SciPy using something I'm calling PostPython, which which is performance-optimized, statically typed Python. So I'm working on a kind of a new project, working on pushing open teams to help every company own their own intelligence, their own intelligence hub. Lots of people are leaning into the open weights world, and I'm excited about that. We've been pushing that for a couple of years. And then still investing in early-stage startups with my business partner. We run a technology incubator., and we're starting a studio version of that at Open Team. So still the same story, just working closely with open source communities. Quansight, the company I founded 7 years ago, still doing well. In fact, you know, new CEOs over there, Rolf and Tanya, are doing great driving that story. I feel like that's a success story. Anaconda was a previous success story where that's going great. My co-founder Peter there is the new CEO. They're actually, I finally turned it around, I feel like. There's, Anaconda did some wonderful things and now they're continuing to do some wonderful things building for the OpenWaite community and really the same mission of helping everybody build intelligence that's geared towards what I call the distributed AI economy, a world where we're not just a few AI overlords, but there's AI as a capability that enhances all of our lives.

Yeah, uh, fascinating. We'll get more, we'll get on that, uh, in a couple of minutes. But first, I think last time we spoke was around GPT-4. Okay, what, what changed since then in, uh, in your world?

Yeah, well, you know, like everybody else, you know, uh, I've been doing machine learning for a long time, done AI for a long time, but you know, like everybody else, when GPT kind of 5 GPT-4 was pretty good too, but this world of, oh, these AI chatbots, the attention, there's several things. It wasn't just the introduction of the attention algorithm. It was also the improvement on the GPU side of massive parallel computing enabling lots, and then hardware improvements on the data side. Just lots of data could be processed all at once in parallel. And, you know, at that scale, all of a sudden, you know, these neural nets organized a particular way, started to actually make predictions about human text, you know, human-based text. Those predictions started to match our experience with each other. Like, oh wait, this is like close to what we find when we talk to each other. That's pretty wild. Now we've kind of replicated that, at least language loop. And then what people could start to do with that just intensified. So lots of people everywhere kind of experimenting with this space and going, well, what can we do with this and how do we go farther with this? And so incredible research. I would say over the past 8 months, incredible progress on coding agents, on this loop, being able to be useful to developer communities who are used to coding. All of a sudden, AI-supported coding is everywhere.. And I don't know of a single developer who isn't touched by that at this point. And it's been a rapid growth, actually, really rapid growth. I was at PyCon recently. I spent the time going around talking to everybody I could find about how they were using coding agents. Some of my good friends in the open source communities have been using it. Now, some— there was a tiny few at PyCon, 4 months ago that were still resistant. And even inside of Open Teams, there are a few holdouts of people like, oh, you know, the AIs do this. And look, AI is not perfect. These coding agent bots, they're not like they're, you know, just let it, it's everything, they'll do everything for you. In fact, a lot of it's like really smart teenagers who just went to MIT, Stanford, and learned everything possible, but they still don't have much street sense. They come in and they're still a little bit unruly. They're still a little bit arrogant sometimes. You gotta go, okay, they don't have the experience they need. But man, everybody I know has been using them somewhat effectively and that's gotten a lot of people excited.

How are you using AIs to code? I guess you would be a fun use case. Yeah, describe a day in your life coding.

Yeah, exactly. So I've leaned in pretty heavily too. In fact, for years I would say, you know, I started my career early as a scientist and then I coded, then I released that code on the internet and started SciPy and NumPy and did I did a lot of coding myself. Since I started companies and started, NumPy was the first example. I wrote the first version and then I started essentially coding through other people. And really for the past 15 years, the code I've been writing has been through others. So Conda, NumPy, JupyterLab, Dask, these other projects I've been involved with, I wasn't the one coding, but I was heavily involved in the conversations and kind of I was prompting other coding agents So for me, it's kind of been, oh, well, I've been doing this for a while with other people and now I have this coding agent that's a little more responsive, doesn't have any, doesn't SaaS, kind of doesn't have their own story. On the other hand, not as creative, not as, you know, they kind of do what you tell 'em. A little bit of, you know, they'll bring things to mind that you may not have thought about, which is cool. So I decided to start with a couple examples and I just started to build a few examples. Myself using, oh, let me see if I can code a little bit. It was incredible how much I could code. 3 years ago, I did the same thing and it was not quite there. I built some examples of, well, let me try this piece of— it was an interpolation strategy that I've learned to build years ago and I tried to get AI to do it and it missed. I had to fill in the details, use vectorized programming and wasn't as good. Fast forward today, I said, wow, this is pretty good. It actually gives me back a story that I'm looking at. I go, that's not bad code that it's writing. Let me— it's a great first start. I decided actually to finally take the reins by idea, but it's been in my head for about 3 years. I call it PostPython, P-O-S-T Python, Performance Optimized Statically Typed Python. It's not just a compiler for Python. In fact, it's a reference implementation of a standard. Years ago, I wrote NumPy to unify NumPy and NumArray. Then I was like 11 years ago now, 3 years ago, 4 years ago, we started something called Array API at data-apis.org. When more arrays showed up, instead of saying we're going to write the one array to rule them all, we're going to write a standard for what arrays, what the API for an array should be in Python. Libraries like SciPy and scikit-learn and these ecosystems around it, they could use the API and then if they wanted to switch from NumPy to CuPy to TensorFlow to JAX, it'd be easy. They could write the long down-term libraries and then they could just put the backend however they wanted if they followed the API. So we wrote that and that actually has led to, you know, recent blog posts just where people are like, oh yeah, scikit-learn now runs on CuPy because they use the standard API and it's, It works really well. Given that, when I have this concept of really Python has a subset that I first explored in the project Numba. Numba is a compiler for Python, but it was a subset of Python. One of the differentiators between me compiling Python and other people compiling Python is some people are like, let's compile all of Python. I've been like, well, I'm not sure that's a non-goal for me. I don't really care about compiling all the dynamic nature and all the scripting ability and all the capability of integration that Python has., but I do want to be able to compile a loop. I do want to be able to compile something that could be a kernel in a Ufunc called a generalized universal function, a generalized Ufunc, or a kernel in a PyTorch module right now. I want that. And in fact, today, 14 years after the first version of NumPy was released, there are 20 or 21 different Python compiler technologies that are in this space. Uh, Taiki Lang, Codon, Mojo, you could consider in this space. It's a little bit— if you squint, it extends it. Cython, mypyc, Numba. Like, there's all these, but I'm like, what, what standard Python are they compiling? It actually varies. Just, there's no standardization around, well, what is it that they're compiling? So the idea of PostPython was to just create a standard for, well, here's the language that defines this subset of Python that's typed, that can be performance optimized, and let's just define it. And so, well, if you're going to do a definition, you got to have a reference version of it that says, okay, so here's— so that's the idea. And I said, well, let me use agents to help me. And turned out with Claude and Codex, kind of both interchangeably, kind of use them both as co-competing agents on the same problem, I could make progress in the first version of that in 4 to 6 hours. And actually have something, kind of a working compiler in that timeframe, which was amazing. It was like, wow. So something that otherwise would have taken me several weeks to months, I could get in an afternoon. And it's not perfect, but I could start. And then basically a series of sessions allowed me to create not only PostPython, but something I'm calling PostSciPy as a— and starting to create modules. Part of the reason for that is SciPy was how NumPy got built. NumPy got built as a reference implementation for SciPy, and how I knew NumPy was far enough along as it could implement all of SciPy. I'm doing this as an experiment. PostPython is an experiment in constructing a standard using agent-first, I call agent-first coding, where agents are participating in the conversation, human oversight for, on behalf of humans. Then 14 modules that I'm writing for, I call pp_special, pp_stats, pp— SciPy has a whole host of modules underneath it. Instead of having a single pp_scipy or something, it's post_python_special, post_python_constants as separate modules. Those become now testing grounds to make sure the compiler standard is sufficient to support that use case. That's a project I've started. I've got a few interns working on it. It's open to anybody who wants to participate, but it's agent-first. Development, which also gives a chance to experiment with what that means. There's a lot of open source communities out there that are experimenting with what that means. I think a lot of my friends are— some are excited, some are frustrated because they get AI slop pull requests coming in that they then have to manage and they're too busy anyway. Here we just go, you know what, it's the output that matters. We're going to do documentation-driven, test-driven development. And if, you know, it's more about does it work and does it give us something we're looking for. And then the ultimate, ultimate result is a spec file and then an implementation of that spec. And so the spec file is something that humans can review and look at and say, yeah, that's something we can support. And ideally, if we're successful, we'll get, you know, Numba, Codon, mypyc, cpy— like all the rest of these compilers implementing that spec, or at least portions of that spec. We got it subdivided in different portions anyway.

So, but how do you code on a daily basis?

Like I code with agents. So I prompt agents.

You prompt agents and then you watch them work and then just come back and then—

I tell it, oh, that wasn't right, do this differently.

And what do you use, like Cloudwork mostly?

I use Cloud Cowork and Cloud Code and then Cowork both. Sorry, it's actually Codex, which has been merged recently in ChatGPT to a single, which is better. They might as well have it together. But I've been using those interchangeably. I've started to experiment with the GLM models. We have at Open Teams a new desktop app called Colab that we're just putting out there that's essentially like a Co-Work, but it's a free download that lets you use open models. And what the main purpose of it is to help maintain the context. Now, I'm a big proponent of privacy and a big proponent of owned AI. For public open source development, I'm not that concerned about using the foundation models. It's totally fine. I'm happy to use them. And so far, my usage hasn't gotten so much that it's too expensive. Yeah.

Did you get an aha moment with all these new AI tools? Because what's the difference between this and being a manager of many talented software developers like 10 years ago? There's no difference.

Yeah, there's not. Well, there's some differences because coding agents are different than humans, but But in terms of the ability to produce, there's— it's, um, no, I did have that aha moment. That's like I started this by saying I've been kind of—

speed, right? You get your results way, way quicker. You— the feedback loop is closed way, way quick, way quicker.

The iteration loop is so much more so quick. Yes. So it helps me realize that, man, the future is going to be, I think, small tools. I feel like it's— we're going to go back to the Unix philosophy. Of, yes, there's gonna be a lot of open source, but the kinds of open source that people will rely on are modules that do what they said they'll do really well and that they can be integrated together with and you can depend on it. So kind of the big framework ideologies are kind of, I think, gonna be less interesting to people because they're just more complicated and more surface for that. They'll just, well, I'll just recreate that for my own needs. Right.

So, and can we trust AI agents to vet like what is the best underlying framework or is that still a human?

No, not entirely. Yeah, yeah, I think so. Ultimately, those AI that what we're trusting is actually the previous documentation that oriented the AI in the first place. Like the AI, how does it get that information? Well, it gets it from Reddit and from the internet and from all the different blog posts that are out there. So you're essentially interpolating the past 5 to 10 years of decision-making, kind of that's what AI is telling you to do. Can you trust it? No, not yet because we're not entirely clear on why it makes the decision it does. I think we definitely need human in the loop, we definitely need validation, but everybody should be leaning into it. Like, there's no reason to run away from it or be scared by it.

Like, if we depend on human bias, right? Like, even Wikipedia to some degree is great, but it's very biased, right? There's a bunch of editors that have like political sentiments and the full information is skewed, right? Code is way less political, fortunately. But it's like there's still a lot of brain power needed. And I feel yet like the best humans will— are sort of merging with AI and are intertwining in a forever spiral of growth. Um, do you feel that is the case? Do you feel that AIs will replace like most jobs because most humans are not super hard workers and want to just chill? Do you feel that the full economy is endangered if AI is let loose. What are your thoughts on that?

So, yeah, a couple things to unpack there. I don't feel the economy is in danger unless we allow AI to concentrate power and wealth too much. That's the real danger.

But how can we stop that? It feels unstoppable, this AI arms race, even if politics are Intervening.

A couple of things are happening. One, the hardware providers don't want that either because they want more people to sell to than just the, you know, 5 oligarchs, you know, 5 monopolists. You know, so you see recently, you know, Jensen, they're supporting open weight models. Qualcomm, NVIDIA, AMD, they're all leaning in. You've got this massive open source ecosystem out there that's starting to experiment. I think I don't think it's unstoppable at all. I think in fact we're at the precipice of a— I think it'll take a little bit because there'll be kind of a bit of a dip as a lot of people rely on this easy button. And then as, oh, these tools emerge and people start writing on their own tools, you'll have this resurgence of a more distributed economy. That's what I'm pushing for and what we're trying to help encourage and support. So in terms of the economy, yes, I think AI will replace or modify many jobs people do today. However, I don't think it means people won't have jobs. I mean, I think people will. And the reason I think that is, is multifold. One, an economy is simply people serving each other. As long as there are humans, there'll be needs those humans have and want. Now, if many of them can be provided by machines, great. But fundamentally, there's a lot of things that machines won't provide to humans and humans will still want. Not the least of which is connection, community, relationships. And so ultimately the question really amounts to where is the money dissipated? How much, if people, if the money is distributed broadly and there's lots of businesses in it, then lots of people will be able to participate in the economy the way they want it. So it's more about, it's less about AI replacing jobs and more about AI concentrating wealth. That's the thing I worry more about. And what I try to encourage by helping create more owners, creating more builders, creating more owners. That's in fact my life goal right now, is to create as many owners as we can.

Yeah, well, I mean, it's also my life's mission to create like as many founders as possible, right? Uh, and humans are gritty, but I feel with 9-to-5s and, uh, corporations the past decades, we really lobotomized grit in humans. And it's like, oh, be comfortable and don't take risk, right? And just chill. And I'm not sure how we're gonna adapt to that shock, right? I do feel that the future is fully entrepreneurial, people owning pieces of businesses and rev share and all of that, and the hourly wage I feel will just dissipate, right? And if we talk about agents, it's going to be per usage, but it's just like, I have a hard time to see how we're gonna have humanity pivot another time away from sort of comfort and embrace grit again.

You know, it's a great point. I think it'll be fits and starts. Frankly, there are a lot of people in humanity who have already embraced grit or still embracing grit. Yes, there are a lot of people who don't. That's going to be— and either they're going to be independently wealthy and fine, or they're going to be not and not fine. And so I think there's going to be— if I am a firm believer in the more wealth there is, the easier it is to support that. I'm more concerned right now about the geopolitics, frankly, supporting a stable economy than I am right now, specifically over the next, I'd say, 2 to 3 years. That's the thing that concerns me the most. I think on the technology side, I was concerned even a year ago, but now with all the movement towards open weights that's happening through very large organizations realizing that's the future, and then also realizing there's massive numbers of people out there who will lean into this. I'm actually more confident. So I'm very optimistic on the 10-year horizon, the 5 to 10-year horizon. And then the 1 to 5-year horizon, I'm like, okay, it's gonna be a bit of a, we're gonna have to be all working together.

It's gonna be dicey.

It's gonna be dicey.

So do you feel government will survive that? Like as a coding guy, as a systems guy, as a scientist, a mathematics guy, what would be your ideal version of a governing body for human beings?

Yeah, that's a great question. I would say generally, and I don't have exact answers, but I have general answers, is kind of the inverted pyramid, right? The more people that you're trying to govern, the fewer the rules should be and the more exact they should be. So if you're trying to govern a massive group of people, then effectively AI should be able to govern it, right? Because it's a very few set of rules and it's a very clear set of decision-making. The fewer people you're governing, the more flexible it can be. The more judgment can be involved. The more, you know, essentially it becomes until all the way down to the tribal level, you essentially have, you know, just good judgment. People interact with each other. So I like to see many overlapping systems of governance with lots of, you know, power corrupts and absolute power corrupts absolutely. And how do you push against that is you distribute it. You make sure that there's no single autocrat that has too much power. And that's something that I thought the internet would help us with so far, and it's not just the internet 'cause this has been going on for thousands of years. This is not just this tension between the ruling class and the bourgeoisie has been happening for a long, long time. I'm still very hopeful. In fact, my hope right now is we can use AI assistance to help more people understand how they can lean into their own if not themselves, at least their tribe, their group, their organizational group, can actually have more ability to push up against the tide, the tide of essentially lack of autonomy, the tide, more independence because you've got support. I mean, right now, even today, I've got attorneys who I work with who are like, you know what, just go talk to Claude. I mean, they whisper me that because they can't officially say that, but honestly, you get pretty good advice. Just from these, you know, interpolating the legal information that the AIs have access to. Like, that's very powerful because all of a sudden, you know, you have the ability now to get better advice and better ideas about how to interact with folks. So I think that's exciting. I think there's the opportunity for that, but there's still organizational work. There's still, you know, I would say the ecosystem to support that is still very anemic. We don't have, you know, For example, a metaphor I'll use is in the United States, homeownership is supported by Home Depots and Lowe's on every corner, hardware shops in every corner, and then also services to people to help support your ownership of your home. We need the same thing in the AI ecosystem. We need an ecosystem of support for owning your own AI, right? That actually has many places you can go for services and support. And because you want to enrich your data, you want to enrich solutions, you want solutions that support your ownership of your AI, not ones that are secretly trying to take it from you. Right? And that's kind of the battle we're in, I think, right now, is that ecosystem support needs to grow and change.

Yes. Um, and it's interesting that you have these philosophies as a guy that's, uh, wealthy, right? Like me, I tend to lean more on a libertarian side, and I do believe in the wisdom of the crowd. And, and I think we talked about it last time, but, you know, you said like, uh, more rules when there's more people, or more precise rule. I think that's what we're getting in this right?

And well, fewer, fewer rules the more people you're trying to govern. More, more very clear and fewer the more people you're trying to govern.

Okay, uh, yeah, I see why you're thinking this, because it's simple and, uh, people with less education can understand, uh, faster these rules and can end up less in jail. Um, but that's not what you said. You said another— yeah, but you said, you said another thing, right? Like more people to govern the, the system. I think there's a trade-off there, and it's that the decisions are slower if there are many. If there's more people, of course there could be a technological solution to that. And also there's less expertise. You're, you're rounding out, you're rounding out sharp corners. Experts know their shit. Like me, I'm more on the technocratic side of things. You just put an expert in power, you give them a lot of power, but you do put mechanisms because I do agree with you that power corrupts, right? Um, and my system also, we would— people would be able to choose it from scratch. But you mentioned something about tribes, right? I think that's smart. I think a fundamental flaw is that you're born into a country, taxes are imposed on you, and you didn't even choose to be there. It's like that, that is not, uh, and a lot of people end up in prison also is like from a country they didn't even choose, right? So I think people should be able to choose that in the future. I think last time we spoke about the network states, right? Network states from, uh, Balaji. I think that would be like the concepts, uh, the concept I'm hinting at here.

I love it. I think that— I think it's exciting because I think it gives us a chance to talk about some of the ideas. I think AI gives us the ability to actually have simulations better because all of a sudden, if— especially if you can have a bunch of trusted AIs— we don't— what we don't want is like, okay, 5 overlord AIs that tell us everything. That's never gonna work. But if you have a million AIs, essentially paired with every organization, they have an AI, they go, "Oh, this is our AI." And then you have— Sovereign AI. You have a sovereign AI and then you have aggregators that say, "Hey, here's a policy we're testing. Let's future-proof this policy and see what your AI says. Let's play this out, simulate this in a model economy where," Individuals or groups can say, here's an AI that represents my worldview, and I'm going to see how it plays out in your future world. Imagine if you had a world where no law could be passed unless it was simulated on AIs, where instead of a representative in government, you have your AI representative that plays in the simulation to see the impact on your way of thinking that that policy would have. And that's got to basically be done and reported on before any law is passed. Imagine a world, that's a positive thing that could happen where we could actually get better, just better data around laws. 'Cause right now the problem with most, even democracies, organizations, you end up with a bunch of people who don't really know. And then frankly, now they're mostly bought out by dollars, by the donors who have funded them, and they're not really representing people. And even if they were, how would they know? They're trying to pass a law and they're like, oh, we don't know the impact, implication of this law. And history and governance is just littered with unintended consequences of bad laws. Some complex. Right? And you could argue, some would argue that, well, some of those laws could have been contemplated if you just had the right framing of understanding.

Yes.

But I could see, actually, if we have a world of many AI systems where people truly feel represented by that model, and if you don't, guess what? You can move to another group where it feels like that AI represents you better. Ultimately, yeah, exactly. Digital simulation of cognitive views. At Open Teams, we have a concept called COGS. I'm promoting. A COG is not just the model because a lot of people are talking about models, cool models, but that's barely the story. The model is just the interpolator of the information. The other part is, well, what information are you giving it? Well, how are you orienting that model? What's the expertise? What's the context? What's the prompting? What's the tools, the skills? If you combine a model with the context, you get what I call a cog that could be specialized. Now there could be billions of cogs. Now, that's a lot of infrastructure to support that, but let's start with tens of thousands. Let's start with hundreds of thousands of cogs for every line of work, for every expertise, for every person, for every— hey, I've got this history cog. I've got this thing that then can be used and pulled into your AI workflow. So that's what we're promoting at Open Teams, a world of millions of COGs, a COG for every org. In fact, most orgs will need like 100 of them. You need one for your accounting department, one for your HR department, maybe 3 or 4 for your sales department. You need some for your engineering. And of course, your engineers will need several because you want a COG that's an expert in some area. So I actually— this is what gets me excited because I do see a world where you have this distributed AI economy. If we can take the AI capability built on open source and apply some of the same principles that made open source thrive, which is putting tools in the hands of people, allowing these communities to self-organize and form and creating support infrastructure, companies that support that infrastructure in ways that people can self-select. Ultimately, markets work if people can say no. Markets don't work when they become monopolies. And it's just, you know, top-down, you know, forcing. But they work when people can say no and then can opt in because then the companies that are serving that market have to serve the market in order to survive. They have to give people what they want. But if people can't choose what they want because they're essentially forced to take it, that's not really a market. And unfortunately, that's kind of one of the things that we've been moving towards is sort of non-market forces. Um, yeah, a lot of that.

I like the test, like, I have a 3-month-old daughter, right? And it's like, what is she gonna think in 15 years from now of what's archaic, right? Yeah. Oh, 9-to-5 model, hourly payment, and me enslaving someone to 9-to-5— archaic. Me eating meat— archaic. Um, me having, like, us dying of cancer or just dying from natural cause— archaic. But dude, the political system, it'll be like super fucking archaic. Like, oh my God, Dad, how could they pass laws without like testing them, right? Seeing the effects on humanity, right? So archaic, it's crazy, right?

I totally agree with you. Yes, exactly. So, so these things give me optimism. The current structure give me some pessimism, you know, so short term. But long-term, I guess, keep everybody excited because those who survive the coming challenges are gonna have the opportunity to build something incredible.

So what will be these challenges? Do you see, like, I don't know, wars, civil wars, and upheaval, like of people getting their jobs slashed or like—

Unfortunately, I do see some potentially, you know, one, you have the, You have the increase of physical violence just in the world that could, you know, and physical violence props physical violence, right? And that's, you know, that's what's happening. And so, we'll have some of that and so then, you know, different people in the common reaction to that is escalation. And as you escalate, of course, it just creates more and more and more unless you can find a way out. Then I see companies making bad judgments.— and doing things like you already seen it a little bit, like companies went and they did mass layoffs thinking that AI would help them and it was way too soon. And then they had to hire back because AI wasn't anywhere near where they thought it might be. So I think you'll see some of that. And then you'll see honestly just a little bit of fear because one of the challenges people in a— when they're in a state of uncertainty or fear, then they'll kind of pull in and they won't lean in because Progress requires people to lean in with faith, lean in with a vision for where they want to see the world in order to take action towards that vision. And, you know, in fear, people don't. They just sort of— it's sort of— and then the effect of that is like jobs are harder. People end up being a little less open for new experiences. So, but I think there are going to be pockets of the world that are doing very, very well. And so I think there's places— there are going to be places to go and places to engage that are going to be doing very well all through this. Yes.

So granted that prediction, how does that change where you invest your time, money, and energy?

Yeah, for me personally, it doesn't change much because I've been, you know, committed to this, to doing as much as I can to build a world of distributed ownership for, you know, 15 years and still doing that. And so, uh, for many people, I think, frankly, I, I think, you know, build meaningful relationships. That's still the most important thing you do in your life. Yes, and part of building meaningful relationships is what drives you then to do things in the broader world. My own life, the reason I got excited about business was because I had a family when I was excited about open source. The reason I got excited about open source, because I love science. I loved the idea of learning about the world, and that pursuit of science led me to open source to help pursue that science. And then that led to just the, oh, these relationships I have with people all over the world building open source is phenomenal. And then my relationships with family, with people that I've committed to, and my children, my wife, my extended family, like, oh, they need me to earn a living, to earn money. So that led me to, well, how do I do that? And so kind of all of my choices have been led primarily from the people closest to me and then the ideas that I get excited about. And I think, you know, that's— recommend that to everybody. You know, everybody thinks you kind of just do that, then you end up building strong connections, you build tools that matter to you and the people around you, and that'll inspire you as to what to build next, I think.

Have you ever been excited like you are today? Has there not been ever a time to be alive as today? It's like great threats, but great opportunities as well, right?

Yes, exactly. Pray you live in interesting times. It's never been so much so true now. It's like there's dangerous times, but it's also super exciting times. Like just the ability to take an idea, and I've not just watched myself do this, I've watched other people do this who otherwise weren't coders, but they have an idea they've wanted to bring into fruition. And now, oh, all of a sudden with a coding agent, they can bring it to fruition. And this is what we've been doing with Python for 20 years, is Python was that language that let people who weren't necessarily coders do enough coding to give an idea of what they wanted. Then there's still plenty of room. I think there's still plenty of room for expert coding capability guided by agents because you're still going to need to support that. Cool. Now you've got an idea. I'm watching that happen already with PostPython and I'm experimenting more with what's going to happen. Where will humans still be needed? Because they're going to be needed.

Yeah. It's like the architecture, the engineering part of it, Do you feel that the human mind could be an infinite box in which AI would fit in, or can AI actually replace that, um, that infinite, uh, yeah, it's a good question.

So, uh, definitely the, the current incarnation of artificial intelligence with the attention model and the transformers will not replace human capability. It's just not possible. So I'm definitely in the camp of the Yann LeCun, the other folks out there who don't believe that —which is pretty contrarian in your position.

It is relatively contrarian.

However, the current intelligent systems we have are very powerful, and they can be used to do lots of things. In fact, I'm interested in what they teach us about ourselves. For example, humans are a very complex organization. We barely have scratched the surface of what neuroscience means. What does it mean to be a human? What does it mean to have our brain? What is consciousness? We don't know the answer to these questions. All of a sudden, with the success of some of these language-based intelligence systems, these attention-based intelligence systems, I'm convinced there's a part of our brain that just does that. It essentially just plays out language in our head and is a little module in our brain. Now, it's not all that's going on in our brain, but it's a big part of it. And so—

But isn't this like a God-given argument and an argument that we're all so special while in reality we're just like a data filtering machine?

Not necessarily. I didn't, I didn't, I actually didn't say that. I personally believe we are special, but I'm not claiming that because of—

we're special, like statistically we're special for sure, right, right.

But, but I don't know the— I mean, but I'm not trying to claim that. That's more, that is more of a religious argument, um, but it's, it's also an argument.

Why couldn't we— I don't have another explanation for it. Like, why don't we reproduce it?

I'm not saying we couldn't. I'm saying the current systems The current systems won't do that. Attention is not all you need for that. That's all I'm saying. Attention is not all you need to reproduce the human machine. Now, could we do more? I believe we could do more to replace, to mimic the intelligent machines that biology has constructed. I believe we could. I'm not in the camp that, oh, we'll never see— I'm not that camp at all. I actually think— in fact, I think most humans primarily live their life as a machine anyway. Like, that's my belief, is most humans live their life as a machine. So of course to be replicated by a machine as NPCs, to put it not in a nice way. Correct. It's like, um, we're all— many humans live their life as NPCs.

Yeah, as like, I'm watching World of Warcraft gameplay as we speak. That maybe that's why, uh, I mean, the human brain is infinitely complex, right? It'll take us a while to understand what's in that box and how it works. Um, and, and I'm excited, right? Like, emotions are crazy complex as well, and biochemistry, right? Like dopamine, serotonin. It'll take a while for us to, to understand all of that. Um, okay, and, but you believe in super intelligence, right? Like in 5, 10 years we're gonna get it?

I, I believe— I don't know, I don't know if in 5 to 10 years we'll get— I think we'll get better models in 5 years. I think some of these world model ideas are good ones, are good directions. I think we have, there's other feedback loops. The human brain, I believe, is a complex system, like several systems. I'm a believer in the 1,000-brain hypothesis, where there's actually many systems that are integrated into our human brain. And so I think we'll be able to— it is crazy. You talked about the emotions. It's one big frontal lobe.

It's mostly the frontal lobe. Which is like computation ultimately.

The frontal lobe is what differentiates us from the rest of the animal kingdom. That's correct, correct. But we still—

it's most of her— it's most of her brain. It's most of her, uh, liquidware. Uh, it's most of like her brainware. It's like, uh, 80% of it, right? Yeah, 80%.

80% frontal lobe, is that right? I thought it was only like 20% is frontal lobe. Really? 25%. Well, we've got neocortex, we got visual cortex, got basal ganglia, got cerebellum. Got lots of brain structures. So, neuroanatomy is something I studied in university at Mayo Clinic. I thought I'd be a neuroscientist at some point. I'm just better at computers than lab equipment, but the brain is fascinating and it's the frontal lobe that differentiates primates. And then, particularly, our frontal lobe is bigger than most other primates. And so, we have the ability to— and if you believe Yuri in Sapiens, which I tend to think he's onto something there, what separated the Homo sapiens from Neanderthals was our ability to reason about mythology. To reason about things that aren't there, basically to make assumptions and have an ideal. So yes, that's also the difference. It's also what ends up killing us because of, you look at wars and fighting and at scale, what are we fighting over? Typically it's an ideology, it's a concept, it's a— It's emotions, right?

It's testosterone. It's like emotions then undergird that and give it energy.

Right? But the logic that constructs the ideal that the emotion then is a feedback loop to support comes from the frontal lobe.

And it's just— and generates cortisol and testosterone.

Generates cortisol, generates all those things. That feedback loop is what keeps it stronger. Yes. And of course, sometimes that feedback loop gets out of bounds. I mean, for me, for example, I noticed my weight gain go crazy when I was CEO of Anaconda because I wasn't managing the cortisol levels well. You know, cortisol, that stress, you have to exercise when you're feeling it. And I didn't do it.

Oh yeah, yeah. Exercise is my drug. I couldn't do anything without it. Last question for you, which is very much related to that, right? And you do serve defense in your company. You talked about geopolitical conflicts like, are we going to blow ourselves up till we get superintelligence? Is that the real danger and not AGI?

That is one of the dangers. Yeah, that— yes, I believe the danger is not superintelligence. It's potentially blowing ourselves up before we get there.

So how do you prevent that? More dialogue, more data, more, less emotions? Yeah, you have to lean in.

Less humans? Well, I would say more people at the table. I mean, I'm a believer in the US Constitution that it's separated powers or tried to, right? And so anytime, you know, so anointing kings and anointing single points of power is always a danger. Right? Even if they're benevolent for a little while, you can never rely on that for a little while until they get corrupted. Exactly. Or until somebody else takes their place, their son comes in. I mean, history is replete from this. This is the lesson of history. So ultimately, you need more people. You need people to acknowledge that they need to step up into their power, what they have, and they need to, you know, you're a leader. If you're a leader of a tribe or an organization or a group, you have an opportunity to be at the table and say, no, this is my group, this is our representation, is what we need. And you respectfully, you know, look for ways to, for peaceful dialogue, use your creativity and your brilliance to come up with satisfying conditions, understand this is what they want, this is what they want, how do we find a place of cooperation, even if what that means is we just go our separate ways, like we can't trade because that may be the case. It might simply be You know what? We don't really align, and so we're going to live in different parts of the world and have very— trade through an intermediary. That may be the answer. But OK. That's worked for thousands of years to create peace.

But has humanity gone better? I think we have shocks every now and then. Last one has been Second World War, and then we tried to learn from that. There's an argument meant to be made for that. Also, I don't think— that generally we're killing machines. I mean, we do eliminate like competition, which is pretty scary. Imagine if superintelligence would do that, right? But we don't tend to exterminate one another all the time, right? Like only a couple of pathological few that probably needs to be controlled in some, in some context. Right. So there is an argument to be made that humanity is getting better, but it feels that we're at a tipping point, right? Like we're— I agree.

I agree. Conflict. I think communication is critical, really getting information out. Usually, in order to support antisocial behavior, you have to restrict communication because by default, humans are actually pretty social, like, except for a few pathologies. Very, very vast majority of humans are social. This is what open source taught me, just, and it was actually breathtaking and super optimizing, you know, optimist-producing, just to go to different countries. I mean, even in country, Persia or Pakistan or countries I've never been and seeing those groups of people there. And they have the same hopes and dreams I do. Different emphasis, different things they care about, but at the end of the day, there's a lot in common. And so building on those common pieces, open source has been a place that people could cooperate. I'm excited because AI means everybody can now participate in open source because, okay, you don't code, that's fine. You can prompt an AI to tell it what you want, and it can build a thing you can then share with the world.

Okay, last question. I know you need to run, but the AI arms race right now, good or bad thing? I probably know the answer to that. And if it's a not so good thing, what would be a better way to sustain innovation?

Yeah, I think a lot of money has been in the flawed idea that we have to go support a couple of big labs. Like, I think that's a flawed idea. We need, we need AI development to be happening in hundreds of places, thousands of places. And so it's better to distribute that both knowledge and investment. So that's, that's my main story. It's like, don't call it an arms race. An arms race is when you have, you know, concentrated power trying to go after something. If you treat it more like science, then you have—

Well, it's US versus China is what I mean right now, right? It's like US—

Yeah, US versus China. I think we need to de-escalate that. Yes. Right? And so I'm a big believer in Jensen's note he just wrote. He wrote a great note on Twitter or on X today, posted the AI, you know, the US needs to really, like the US policy, the US government, the US research labs, they need to think hard about, oh, we should be promoting OpenAI as a science and as a gift to the world. The United States should absolutely be doing that. We've got the talent, got the capability. We've got people who are already doing it with very little money. It's a no-brainer to me that we should have massive—

Populism is so, so much easier to be populist, right? And just go with the emotions and manipulate people and take low education people and just convince them that China is evil. I just encourage people to visit China, go step in China and talk to the people. You'll see They're just like, you know, they're very smart. They're smart business people. They're great. And, uh, yeah, it's, it's, uh, a bit more complex than it sounds. Like, I don't want to sound like a peace and love hippie, uh, which I'm not. I'm a super capitalist. But it's like, there's good business to be done with China, and probably better business to be done with China vs. than just, uh, against them. Uh, Travis, where can people find out more about you?

Yeah, so I'm on X at @teoliphant. I'm at LinkedIn on @teoliphant. You can go to openteams.com, our YouTube channel, DM me on X or on LinkedIn.

So that was Travis and Charles.