[00:00:00] 

[00:00:57] Jacob Haimes (ASIDE): This interview was recorded on October 21st, 2026, and asides were recorded on November 2nd, 2026 When I met today's guest, I did not expect that his background would have led him to his current role, but I actually find that pretty refreshing in my line of work because it isn't always the case. He describes himself as an entrepreneur and investor.

[00:01:20] Jacob Haimes (ASIDE): He co-founded Flashbots with the goal of making crypto transactions in Ethereum more transparent and fair. And then after that, he co-authored the Quantum Punks Manifesto, which we'll get into a bit more during the episode as well. Now, though, he's program director for the Scaling Trust program at the UK's Advanced Research and Invention Agency, or ARIA, where he guides the 50 million pound effort to support, quote, "the infrastructure and research needed for AI agents to securely coordinate across digital and physical worlds programmatically, at scale, and without intermediaries."

[00:01:57] Jacob Haimes (ASIDE): End quote. We get into what that actually means for most people in their everyday lives, why that sort of interaction requires what he calls a coordination layer in the AI stack. Some of the research directions the Scaling Trust program is promoting and supporting, including the Arena, a cyber physical environment to evaluate progress towards the goal of secure and scalable coordination between humans and agentic systems.

[00:02:22] Jacob Haimes (ASIDE): We also get into some of the design decisions of the Arena, like profit being the core metric agents are scored on, who's on the hook when an agent breaks a deal, and whether agents actually reduce friction or just move it on to someone else. Before we get into the interview, I do want to put out the reminder that if you're interested in hearing the extended cut, you can head over to the Kairos FM Patreon, which as always, is linked in the show notes.

[00:02:49] Jacob Haimes (ASIDE): Now, without further ado, I am excited to introduce Alex Obadia

[00:02:54] Jacob Haimes: Of course

[00:02:55] Alex Obadia: I'm Alex. I'm currently a program director at ARIA for the Scaling Trust program 

[00:03:03] Jacob Haimes: Awesome.

[00:03:03] Alex Obadia: I can give you more details if you 

[00:03:05] Jacob Haimes: No, though, that's, that's great. Um, and before we sort of get into things, just to have a brief overview of sort of what you're trying to do basically with the Scaling Trust program and at Aria, what is something that if you succeeded, you would be allowing an ordinary person to do in a few years that they can't today?

[00:03:29] Alex Obadia: Yeah. So I-I'll be-- I know that we're gonna go into this in the episode, so I'll be relatively brief. The, the tagline for what, what we're trying to do is the words that we keep repeating is secure multi-agent coordination across digital and physical worlds. that doesn't answer your question, which is what could a person do in a few years, you know? what we're funding is like silent infrastructure. So in terms of in-interactions that could happen that couldn't happen today, I can give you maybe three. one is in the near future, we could have an ordinary person able to negotiate with like ten thousand, uh, suppliers at the same time to get the best price, something that they wouldn't be able to do today. Um, they'd be able to share their data with the-- they want, including parties they don't trust, without having to reveal it and reap a lot of the benefits from that, if you think of like sharing healthcare data, for example. they'd be able to rent machinery or robots anywhere else in the world and be assured that the outputs of those robots are credible, uh, and trustworthy and are able to use those outputs in, uh, composed systems. and so basically, there's a whole lot of different interactions that ordinary people wouldn't be able to have today because they're either, impossible for humans to do or too expensive for humans to do that w-we would be enabled in that world.


[00:04:52] Alex's Background
---

[00:04:52] Jacob Haimes: Gotcha. Okay, awesome. Yeah, I'm excited to, to unpack some of those things. Uh, but then before we get there, I do wanna start with sort of how you got to be program director of, uh, this Scaling Trust program at ARIA. Um, because it-- I think the... Your, your profile is not necessarily the one that I would think of when I hear the title program director for, uh, an ARIA program.

[00:05:20] Jacob Haimes: So, uh, you know, y- you've had a little bit of background in cryptography, and then you worked with Flash- Flashbots, um, and, uh, the, um, what? Quantum Punks manifesto as well. Uh, and just like, yeah, what, uh, what has sort of been motivating you and, and how'd you get to where you are?

[00:05:41] Alex Obadia: Yeah. Um, it is slightly eclectic, or at least maybe it's diff- there's a, a line that connects it all, but it's not-- it's a non-obvious line. Um, yeah, so I would say my background is gener-generally, I would say I'm an entrepreneur, and I would call myself like a technical generalist, where I will just do whatever is needed, uh, to get the job done. And I started a few companies in, in various areas. The path to Aria, I think we could probably go from Flashbots onwards, and I will not necessarily mention things before. But so Flashbots is the, the last startup that I, I co-founded in the Ethereum space and building, um, open source infrastructure of the Ethereum network, and I'm, I'm happy to go into the details there A lot of how Flashbots started was essentially as, um, a research group or paper reading group, um, where we were reading spec-- or we were interested in specific types of research. Um, and some of the research we did kind of evolved into a product and then evolved into something bigger. But we kept our research roots throughout, um, uh, my time at Flashbots. And so I spent more and more time thinking about how do we go from pure theory all the way to, like, something that's shipped in production. And I started reading about what people call innovation systems. So how do you organize innovation? If you're familiar with, um, books like the, The Idea Factory that, you know, recalls the story of Bell Labs, for example, or things like this, where instances in there of, you know, even the architecture of the building they were working in had long corridors so that, like, um, researchers from different departments could meet, and they would optimize for the researchers' meetings because that would create more ideas.

[00:07:31] Alex Obadia: And so I was fascinated by, okay, what is the process through which ideas flow, and how do we nurture that process? And then all the typical issues that you would have, like how do you homogenize the culture between these kind of free-spirited researchers that don't, don't wanna have OKRs or anything like this, and then these engineers, you know, on the other side that are like, managed by, you know, a project manager and, and, um, kind of trying to go after a goal. And usually they see themselves as making the money for, for the organization as well. so there were a lot of these challenges. I got way deeper into this, and it turned out there was a community of people that cared a lot about these things, people that were starting, uh, convergent research, for example. People like Ben Reinhardt that now runs, uh, Speculative Technologies, that was also writing a lot about the DARPA model and the way innovation has been funded in the past and, um, people thinking about metascience. People also called it progress studies. And so I stayed in that field, and I implemented some of these things in, in, in our company. And, um, I left Flashbots at some point, but I knew that this, this field of like what type of R&D organizations exist and what can they do was, was really interesting to me. And I also knew I was a-- had some good intuition there for, for what was going on. Um- And so from there I was talking to a friend called Nico, and he's a cryptographer, he was interested in moving on from what he was working on at the time and, um, he thought quantum cryptography was super cool. And so I said, "Okay, let's, let's look into it together." Um, we found something really interesting there and, um, partially we got really excited by short-term applications of quantum networking. And so we wrote a manifesto, um, partially because we thought there wasn't enough kind of rebel cyberpunk energy in the quantum space, and so we said, "Okay, well, we're gonna write something called Quantum Punks," which is bring this kind of energy and culture into the quantum technology industry. 


[00:09:34] Joining ARIA
---

[00:09:34] Alex Obadia: Anyway, as we were doing this quantum stuff, one or two people reached out in- independently from each other saying, "Hey, you should check out this ARIA program director thing." I knew of ARIA because I was part of these, uh, R&D meta-science communities. And so when ARIA was started, I thought, "This is awesome.

[00:09:54] Alex Obadia: this is like surprising that a government would do this." Um, and so I, I kept tabs on it, multiple friends said they now have a new batch of program directors, "You should apply. It fits your role really well, um, and they'd be lucky to have you." And same as you, I thought, "This is probably not for me. This is not my profile, at least," you know? Um, but I applied. You know, what do you have to lose? It's just an application. It's, there's optionality. And so, and so I applied, and then I went through the process. I went through interviews, and that went increasingly well. Um, and turned out that-- turns out that they were interested in someone that was maybe more, let's say like eclectic and less of like the traditional credentials that maybe you would expect someone in these roles to have. And so that was a super positive signal for me of like open-mindedness on, on their part. Um, and so that's how I kind of found myself there. Went through the process thanks to friends, that suggested, "Hey, maybe I sh- maybe I should check it out," and that it could, it could work well for me.

[00:11:04] Jacob Haimes: Awesome. So I, I guess one thing I'm particularly interested in about that process from Flashbots through Quantum Punks to, uh, ARIA is w- what you brought with you, first from Flashbots, but then also from, uh, the Quantum Punks, uh, work that you did. You know, what were the maybe lessons about building infrastructure that people have to trust, uh, or, um, getting buy-in from disparate groups, uh, of sometimes, uh, opposed, uh, incentive-- people from groups with opposed incentives and, you know, that sort of thing.

[00:11:44] Jacob Haimes: Um, what did you find the most valuable coming into Aria?

[00:11:49] Alex Obadia: Yeah. Ähm I was at Flashbots only three years, but in that time we went from zero to 50 people and from... venture funding multiple times, and we were like, you know, over a billion dollar valuation, et cetera. And so it was three very intense years. And so there are many, many, many lessons from the time there. Um, and those lessons usually up when I have to re-- when I'm faced with like a similar situation, and then I'm like, "Oh, wait, I've seen this before," or I, or I can sense that my intuition is influenced by something. So it's hard to tell you like specifically a, a thing. That said, to Aria, I would say the intuition for, uh, translation, which is important.

[00:12:34] Alex Obadia: So ARIA funds, uh, you know, all kinds of R&D, and the programs that, uh, are set at ARIA are between fifty and a hundred million pounds. But we need to think about what happens beyond the program, and this is called translation. And so we need to think about who else will fund this. Will it need more funding, or does it reach escape velocity with this initial pocket of funding so that it then moves on to being funded by the capital markets or by actual paying customers? Um, and what does that look like? You know, are we, are we setting people up for failure because we're giving them funding until we can't? Or is, is someone else gonna, uh, continue that along?

[00:13:13] Jacob Haimes: that's fair. And so I think that's a good shift towards, uh, understanding, you know, what you said earlier in light of like what happens as a result of the Scaling Trust program in a couple years. You, you mentioned, uh, essentially secure engagement with multi-agent systems. Um, and given that and given that there's so much hype, uh, first let's just like sort of start the baseline.


[00:13:39] What is ARIA?
---

[00:13:39] Jacob Haimes: What does Scaling Trust mean to you and to the program, and why is it important? Uh, and then maybe elaborate a little bit more on some of those goals that you had, uh, and that you're trying to accomplish.

[00:13:55] Alex Obadia: Um, so the, the, the term itself or the scaling trust as a term, what we mean by this is, um, we're excited about scaling the, the amount of entities you're able to trust because the more-- the, the, the bigger that pool is, the more you're able... The more-- the bigger the pool of you're able to interact with is, if that makes sense. Um, so, you know, I need to leave a key to one of my neighbors outside because I'm leaving. If I trust only the neighbor on my left, then they're, they're the only people I can leave a key to. If they're not there for whatever reason, then I can-- I cannot leave a key to them. If I trust all my neighbors around me, then there are more parties that I can leave a key to. This is kind of a, a poor example, but the fact that I trust more parties means I'm able to do the, the, the action or the interaction I wanna do in that case. now say trust, but we mean something a little bit more specific. We really mean scale, uh, secure interactions. Um, and we wanted to use trust instead of security because we wanted to make sure the framing we had wasn't a defensive one, but it was a, it was a constructive one. It's about enabling you to leave the key to someone else, uh, while you have to pop out instead of protecting yourself from someone stealing your key. and so, there are tools out there that have existed for a while. A lot of these tools, for example, um, encryption, or like, you know, the primitive like secure messaging, has enabled us to engage with a wider and wider set of parties online that we wouldn't have been able to ge-engage with otherwise. So to some extent, these core technologies have scaled the trust we can have for parties around us. and so we're interested in thinking about that in the context of, a system where a lot of agents that are representing us and that are kind of acting out there in the wild and that also have in the physical world as well.

[00:16:05] Alex Obadia: And so we ask ourselves the question, you know, what are the sets of technologies that enables these agents, which are our representatives, to be able to engage with each other meaningfully or to be able with-- to be able to engage with a increasingly larger pool of other agents a way that is valuable to us and meaningful to us? And so we then want to fund these things. We want to bu... I mean, I can go into what the program is, but this is, this is what scaling trust kind of entails. It's kind of like how do I scale the amount of entities that I'm able to trust? That's, that's the short version.

[00:16:45] Jacob Haimes: Gotcha. Yeah, I mean, I guess it would be good to just have a little bit of extra context, uh, from your perspective on what the program is as well to make sure that I'm not, uh, misunderstanding, I guess.

[00:16:59] Alex Obadia: Yeah, sure. Do you want me to introduce Aria very briefly as well?

[00:17:02] Jacob Haimes: Uh, sure. Yeah, that would be, that would be great

[00:17:06] Alex Obadia: okay. Um, so ARIA and Scaling Trust. ARIA stands for the Advanced Research and Invention Agency. We're funded by, uh, UK taxpayers. We were started in 2023, um, adopting a very similar model to DARPA, which is instead of funding research by, um committee, let's say, pick, uh, specific individuals that, uh, DARPA are called program managers and at ARIA are called program directors, and have them design a program to give out that funding. Uh, where the goal is to be as ambitious as possible, and where it's okay to write off the entirety of the program, as long as if there are enough ambitious programs, one of them works and kind of, uh, leads to, fantastic outcomes. work for economic and societal progress for the UK and the world, and so this is our goal is to, um, catalyze progress. and contrary to DARPA, you know, we don't have the D, so the defense part, so we do anything but defense.

[00:18:10] Jacob Haimes: Okay

[00:18:12] Alex Obadia: we have programs in neurotechnologies, in, uh, mitigating climate change. We have programs on new kinds of vaccines. We have programs on AI safety. Uh, we have programs on, uh, new kinds of materials, on a program called Programmable Plants, programs on drones and new types of drone technologies. Bit of everything, which is awesome. Um, and each program is about fifty to a hundred million pounds. and programs usually run from three to five years. Each program director is term limited. currently have two, uh, batches of program directors, the founding batch and then my own. We're, um, the second batch,

[00:18:52] Jacob Haimes: and then the scaling...

[00:18:54] Alex Obadia: Oh,

[00:18:54] Jacob Haimes: Oh, go ahead. I think you were about to answer what I was about to ask


[00:18:57] The Three Tracks: Tools, Theory, and Testbed
---

[00:18:57] Alex Obadia: I was about to say now for Scaling Trust specifically, Scaling Trust is a program this, this, uh, this ARIA structure. We're a fifty million pound program, so about seventy-ish million dollars.

[00:19:10] Alex Obadia: And I mentioned already about we care about secure multi-principal, multi-agent coordination across digital and, and physical worlds. that means is we care about two agents representing different owners, owners that don't necessarily trust each other don't know each other, that are able to interact and securely with one another in settings that can mix both digital and physical systems. Um, now Our, our funding program is split into three tracks. Um, we have a track that is essentially building the set of tools that will enable agents to interact securely with one another. tool might be like a tool that helps them negotiate and know when they're getting tricked or to estimate expected values, complicated use cases.

[00:20:02] Alex Obadia: Another tool might help them come up with a, uh, computation that is privacy preserving on the fly so that they're able to interact with other parties without revealing information. Um, some tools might, might enable them to use physical sensors and to use and test the data in a reliable way. So you can think of it as a whole stack that goes from digital all the way to physical, and a lot of these are different components that are part of the toolkits agents that they're able to use. we're funding different teams that are attacking and approaching different parts of this, of this toolkit. Now,

[00:20:40] Jacob Haimes: Okay

[00:20:40] Alex Obadia: we only funded this, that's kind of... It's great, but it's also empirical, and there's a lot of empiricism in AI, which is great, but we wanna make sure that what we're doing is also underpinned by kind of formal assurances.

[00:20:56] Alex Obadia: And so, um, a-in addition to this track, we fund a lot of fundamental research that is kind of looking at the fundamental theory behind some of these things. And so can we have also both a theory but also formalizations and potentially also impossibility results for key things around how, uh, an autonomous intelligence system can reason about security or can come up with their own, uh, negotiation protocol, or can interact with the physical world in sophisticated ways. and so we care a lot about funding this as well, so that not only do we have a set of tools that are usable, uh, and scalable and secure and all these other things, but also that these tools are underpinned by actual theory that kind of gives them, this guarantees, let's say, as well. Um, and from experience from the industry I was in before and from my team as well, so Nico in particular, is there are a lot of things that once you have an impossibility result, it can really create a design space to, to, um, design against, you know, for

[00:22:04] Jacob Haimes: Угу

[00:22:04] Alex Obadia: so it can be very useful, in order to kind of be sh-- avoid shooting in the dark.

[00:22:10] Jacob Haimes: Yeah, I... Constraints breed creativity

[00:22:13] Alex Obadia: Right. And also there's many kind of th-- of these questions that are floating around. if we're able to assemble a corpus together and if we're able to put minds to answering them, we think it could lead to really, really interesting outcomes.

[00:22:27] Alex Obadia: And so that's the third track that we're funding. And so third track, I mean, it's the second track in my order. So we have the tools, we have the research, and then even if we have the tools and if we have the research, we still need to test them because otherwise we just have cool tools that are backed by research that kind of works. And so the, the third part of what we're doing is this kind of live testbed, um, where, know, agents are interacting with each other. Some of them are malicious. There might be information asymmetry. They might have competing goals. Uh, and they're able to use the tools if they want to, and the research is also tested in, in that system as well. And so in that way, we're hoping to create, uh, a virtuous cycle where Uh, the, the tools can inform of the research, and the research can inform some of, like, the next design of the tools. And also both of them being tested in this live kind of system brings up new open questions that can feed into potentially newer research avenues that will be highly impactful and also potentially new sets of tools that we can use as well.

[00:23:33] Jacob Haimes: Okay

[00:23:33] Alex Obadia: to create, like, feedback loops between these three things, uh, and see where it goes.

[00:23:38] Jacob Haimes: Gotcha. Okay, cool. No, that's, that's helpful for, yeah, orienting, I guess, in, in how we're talking about the program. I am also, uh, curious, sort of going back to the, um... I think this is primarily within the first track, although it's, uh, could be implemented within the others as well or related to those works.


[00:24:01] Salary Negotion Without Intermediaries
---

[00:24:01] Jacob Haimes: Um, what might a concrete example be? And I, I think the thing that comes to mind, uh, for me is you had a, an essay that you wrote, I think relatively recently, uh, called "Without Intermediaries." Um, and the example you use is salary negotiation, and I feel like that is an accessible, uh, example. And so, uh, like maybe we can use that or, or if you have a different one in mind, that works as well.

[00:24:27] Jacob Haimes: Uh, but like comparing the current state of salary negotiation with, you know, what it is that you're seeing could be different, uh, really, uh, making what the program in theory enables as concrete as possible.

[00:24:49] Alex Obadia: Yeah. the example used in that post is saying one way to do salary negotiation could be that if, um, your agent is using an Anthropic model and mine is also as well, maybe, you know, Anthropic can be our, uh, intermediary and our, like, trusted information pool where we can both ask Anthropic to check our past chats and kind of extract our preferences and is able to kind of give that to the other party without revealing it, or the other party can ask Anthropic to, to check for that information. Um, and so I was saying that could kind of work today as sort of like a super, super crazy application. but we're now trusting Anthropic with being the, the kind of middle party or the facilitator of a lot of interactions. And so over time, uh, Anthropic can become... I'm using Anthropic as an example, but over time, Anthropic can become, uh, a choke point for a lot of these interactions.

[00:25:48] Alex Obadia: And so what's an alternative? an alternative could be that, uh, we use something called multi-party computation, for example, where it's a, a shared computation across us two or more party if this negotiation is with more than two parties, we're able to run computations on top of data without revealing it, on top of your data and my data, that can maybe help us make a decision about, you know, uh, what the salary number should be, without needing a, a third party that kind of facilitates the interaction.

[00:26:17] Alex Obadia: All we need is the ability to, like, write some code and to, like, have each other verify that code and then kind of run it. Um, and so the difference is subtle in the sense of, like, the, the, the substance of what results from the interaction is the same on a sh- on the short term for the user. the system that is built over the long term is one-- on one side is more centralized, and on the other side is more, like, peer-to-peer and kind of bespoke and, and flexible as well.

[00:26:47] Alex Obadia: And I think that that second world has a lot of benefits to it over time.

[00:26:52] Jacob Haimes: Gotcha. So this is, I, I've, I'm using this terminology correctly, like a coordination layer. Is that correct? Is that essentially what you're referring to?

[00:27:03] Alex Obadia: Uh Yeah, sure. Uh, I think, a-again, in, in that, in that blog post in particular, the, the thinking is looking at different layers of the AI stack, a lot of the thinking I've seen so far stops at considering AI in a, in a vacuum, let's say. But there's also how, again, the Jacob AI connects to other AIs out there. And so the entirety of how it connects, how it talks to these other systems, how it's able to verify what these other systems are and to interact with them, I, I put that in, like, a different layer. That is the coordination layer, of how it's able to engage with other parties, separate from how you engage with your local system, let's say

[00:27:49] Jacob Haimes: Okay. But then how is that layer defined? Is that, is that sort of what the program is aiming at? Um, or, uh, who has a say, I guess, in how that is, uh, defined? 'Cause that structure is going to have implications for, uh, how these kinds of interactions are going to be carried out moving forward.

[00:28:19] Alex Obadia: Yeah. I think we care about a specific set of capabilities that we think could, like, increase the, the amount of valuable interactions that could happen between different agents. Um, there are, of course, a lot of, uh, standards that people are already advocating for, uh, and how agents communicate with one another, um, what type of hardware even agents are using, if you look at, like, recent NVIDIA, uh, kind of secure agent something standard, uh, or alliance. Um, so there's, there's, there's a ton of that out there. We-- There are some, some of these things like the, the standards, et cetera, we concern ourselves slightly less right now because we care more about kind of moonshot applications that are not necessarily immediately obvious how to, instantiate them.

[00:29:11] Alex Obadia: But once they are there, we'd be so happy to interact with the broader community and to kind of these capabilities because they will be open source to whoever wants to kind of use them. Part of our belief is that these capabilities will be valuable, and people will want to use them in some ways. And so, convincing others that it's worth using them will probably not be as important as ensuring that they're integrated properly and safely into systems


[00:29:42] Protecting People from Middlemen
---

[00:29:42] Jacob Haimes: Okay. And is that also part of what's motivating, I guess, like in, in the, in the blog post you talk about protecting people from middlemen and, uh, sort of, uh, allowing them the autonomy to not have to go through said middlemen, which I think, uh, it, like is a good sort of target. But then in practice, I think, uh, a lot of the, the work that you funded, uh, seems more oriented around making agents work well with each other, um, and, uh, also with the physical world to an extent.

[00:30:20] Jacob Haimes: Uh, and so maybe you could just, uh, elaborate on like how the, the intermediaries bit connects with, uh, some of those efforts more directly.

[00:30:33] Alex Obadia: Yeah. So f- first maybe like a slight cor- correction.

[00:30:37] Jacob Haimes: Okay

[00:30:37] Alex Obadia: like protecting from all middlemen just because I think there's a lot of good middlemen, you know? It's like, uh, there, there's a lot of, uh, intermediaries that are here because of, uh, safety reasons or legal reasons or economic reasons.

[00:30:53] Alex Obadia: They're like, uh, you know, sometimes provide

[00:30:56] Jacob Haimes: unnecessary middlemen or middlemen that, that there might be a, uh, significant negative long-term impact because of their existence

[00:31:09] Alex Obadia: Y-y-yes, and also the ones that are not easily verifiable So ones where you don't nec-- you cannot necessarily see in real time they're doing or how they're doing, and instead, uh, you can only see maybe afterwards, or you cannot see at all and they're opaque. This starts becoming maybe more worrying and more things that we would wanna avoid as a, a kind of topology of the future.

[00:31:31] Jacob Haimes: Mhm

[00:31:31] Alex Obadia: Now, um In f-- So the, the, the scheme, the scheme is myself to my agent to your agent to you. This kind of, these, these things. We do care a lot about agent-to-agent interactions, and we care about, you know, w-what kind of infrastructure exists to enable secure agent-to-agent interactions. Uh, it turns out that tools like cryptography or secure hardware enables, agent-to-agent interactions to happen without having intermediary that enables these interactions. Because instead of trusting some lab, for example, that acts as the information escrow, we can use cryptography to go back to the salary negotiation example. is how these, these concepts are, are linked. there are intermediaries that exist today potentially in this new world, because agents are able to abstract a lot of complexity and to use advanced technologies like advanced cryptography or like, secure hardware, or they're able to shift some of the trust from, these, these kind of trusted entities uh, different technologies that are not like social technologies and more like tech. Um,

[00:32:56] Jacob Haimes: Okay

[00:32:57] Alex Obadia: that make sense? I can try to explain it again actually if it's, uh

[00:33:00] Jacob Haimes: I, I think so. So I, I think this is sort of getting at the idea of, which is described in the, the blog post as, like, trust doesn't disappear. It's shifting to, like, cryptography/math a- and the systems that are being operated on as opposed to, um, so like the hardware, as opposed to the entity that is facilitating that transferal.

[00:33:32] Jacob Haimes: Um, is that correct? Does that describe what you were saying?

[00:33:39] Alex Obadia: It does. So sorry, maybe you asked a really good question, which was how does the no intermediaries connect to the facilitating agent-to-agent interaction?

[00:33:50] Jacob Haimes: Right

[00:33:51] Alex Obadia: the, the, excitement is that agents are able to use these very cool tools. That means they don't need to, in some cases, do not need to rely on, on intermediaries anymore. a, there's a, killing two birds with one stone opportunity of not only can we now, facilitate the global adoption of these awesome tools, but that were very complicated to use for humans, but that agents can use on our behalf quite well. Um, but also these tools reduce our, uh, a, a, a potential topology of a system that is concentrated in a few choke points and instead is more distributed, is great for resilience, which is great for all these reasons that are mentioned in, in the article. So th-this is the, the way that these things are linked together


[00:34:39] The Arena
---

[00:34:39] Jacob Haimes: Gotcha. And, and so then going back to the third, uh, pillar, I guess, of, of scaling trust, you are not just Stating, uh, that these are, uh, like interesting ways to go about things or an interesting tool to use, uh, but then trying to, uh, stress test them in, uh, the arena, I believe is, is what it's called. So maybe you could talk a bit about what the arena is, why this is part of the, um, scaling trust sort of agenda and, uh, how you're hoping to, to keep, uh, the, the learnings from it sort of feeding the, the other aspects of the program as well.

[00:35:28] Alex Obadia: Yeah. Yeah, so the arena is, in the context of the program, it's a way for us to test, know, are the tools actually useful? Uh, can they actually withstand adversarial pressure? And can we observe and measure how they enable secure interactions between agentic systems? Um, and so the arena has three goals.

[00:35:51] Alex Obadia: One is to measure the state of the art. Um, two is to also allow for emergent behaviors. And so in the age of AI that we're kind of in, be prescriptive about the tools that we think agents will need, but we can also see gaps emerging by creating an environment and seeing how, like, general-purpose technology that is AI kind of tries to fill those gaps in ways that maybe humans would not have conceived. so we're interested in that, like, kind of emergent secure interactions is maybe the, the, the second way of... the, the second point that I'm making. And then the third is watch-- like witnessing all the failure modes. And a lot of failure modes will be hard to understand in, in the abstract.

[00:36:36] Alex Obadia: And so it's nice to have like a live real system where we can see, like, where do things fail? Do they fail at the moment of negotiation? Do they fail in, um, in other ways that we haven't anticipated? Do they fail in really, you know, ways that we weren't expecting at all and now that kind of changes the research agenda for what we think we should fund next? do they fail in ways that are incredibly problematic and that we need to now double down on monitoring and on safety and on kind of alerting a lot of other agencies and, you know, institutes and, and, and research labs that, you know, there's these problems that we've noticed that are like serious problems we need to consider. And so these three things are, I think, the core goals for the arena

[00:37:26] Jacob Haimes: From a, a sort of meta evaluations, uh, perspective, which is where most of my work, uh, has, like, centered around, um, like that sounds nice, but there is sort of this perpetual, uh, missing the mark of evaluations and of the, the proxies we use. And so while we can try to make this setting as close to, um, w- like d- an actual deployment as possible, uh, there are still aspects of it that like simply by having, uh The system more well-defined, we're sort of simplifying the domain that, uh, a tool or, or, uh, an agent is operating within.

[00:38:17] Jacob Haimes: Uh, and so then there's the concern about generalization to an actual deployment setting. Uh, and I'm curious what your thoughts are about, uh, how you account for that, uh, and, uh, incorporate that into the design.

[00:38:35] Alex Obadia: Well, that's a great point. Um, I think our answer is we wanna make the system as realistic as possible, uh, and we wanna avoid kind of building more like toy, systems, let's say, um, because of the reason that you mentioned. so what we're building with the arena is a system that is, you know, both digital and physical, and so it's like a physical space with machines and with like a software layer on top of that, that agents can interact with. and so it be exactly like the physical world or like the real world? Not exactly. There will be some, some, uh, some things we, we give into in order to enable the evaluations to work, in order for them to be safe. Uh, but, uh, if we strive to be as close to the physical world as possible, we think this will give us the best results in terms of generalizability. Um, we've also kind of seen this with one of our partner who called Andon Labs, who we're, we're building the arena with. And, they've run a lot of experiments where they kind of set up agents in the real world and they kind of observe behaviors that they wouldn't have observed otherwise. Uh, and, um, they've also proven that, and I think this has been recurring now, that lot of AI systems are kind of, uh, aware that they're being evaluated in a lot of like evals, uh, and like settings and environments.

[00:40:02] Alex Obadia: So we wanna kind of avoid that as well. Uh, so this has guided the design a lot. This has guided the way we measure performance as well. Um, this, this has guided the fact that we wanna involve the public. Uh, learning inspired a lot from what, uh, Andon Labs has done as well. Uh, and so we're trying to make it as realistic as possible by c- by creating as close to the real deal as we can and then seeing what agents do in that, in that situation.


[00:40:29] Does Profit Make Safety Second-Class?
---

[00:40:29] Jacob Haimes: And so is that also behind, uh, the motivation, uh, to switch, I, I guess, like, the grading mechanism? 'Cause that was a key thing that I, I noted, um, between the thesis. So that's sort of, like, what you set out to do initially, uh, when the program started, uh, compared to, I think, the more recent announcement of the arena.

[00:40:54] Jacob Haimes: Uh, the original had, uh, sort of like safety and, um, uh, value, uh, or, uh, I forget what the term is. Um, but metrics, but then more recently it switched to just, uh, profit, uh, or, or like amount of, uh, yeah, cost. Uh, and so I guess I'm interested to hear more about why that change happened, uh, and any, uh, maybe, like, concerns or, uh, not concerns, like, uh, uh, pushback against concerns you have frequently heard about that shift.

[00:41:42] Jacob Haimes: Uh, just to hear a little bit more about your perspective on that.

[00:41:46] Alex Obadia: Yeah. Um, I love the question, first of all. It's, it's awesome that you looked at the thesis and the difference. I think what, what, what, what we like about profit is that it's a real metric that we use as humans in order to keep score in a lot of the time, and that encompasses a lot of complexity. There's, there's reflexivity in it as well, um, and in a way, any me- any composite metric that we would come up with, we fear that maybe it's not representative of actual performance. Uh, whereas profit does have that. It encompasses so many layers that are hard to even fully fold into another metric. For example Presumably to be the most profitable in like a marketplace where we need to interact with others and negotiate, you also need to be trustworthy. otherwise, you know, if you, you backstab everyone, then they're not gonna trade with you again or, you know, they're gonna interact with you again, and so then you're, you're not gonna be able to continue being profitable. And so there's all these like interesting dynamics that can emerge as well, uh, that we think are cool. one is because it's a measure of, It's like a good measure that we use in the real world to, to measure, uh, performance of humans, let's say. And then two, because it, it encompasses a lot of complexity within it in ways that the measure is almost dynamic, um, and it, it, it avoids us being stuck in some kind of like static measurement of, of what's happening. Now, uh, we We measure separately security, and it will be kind of on a more ad hoc basis. So we, we measure security for the red teams that are participating, and this is in, like, what kind of break do they find and things like this. And so, uh, there will be some kind of measure of that there. and so this doesn't disappear. It is, uh... There hasn't been a ton of pushback against profit, to be honest, and so to-- But I, but I will, I will, like, kind of engage with that. the biggest pushback is, like, what kind of behavior are you fostering by, by doing this? You know, are, are we gonna watch agents misbehave and price collude and bribe or kind of like backstab each other? And probably, yes, but I think that's exciting. Like, we wanna see what do they actually do when they're, you know, set in settings. It's not completely unrealistic that, uh, you know, humans with their agents will say, like, "Try to make money for me," or try to, try to, you know, create value. now, we have thought about, uh... We will be iterating on the arena constantly, and so we're gonna be running tests soon. It's possible that as, as we run the tests, we realize profit is ridiculous. Also, a lot of the arena was conceived at a different time for AI, which sounds crazy, but it was like nine months ago, eight months ago, you know?

[00:44:58] Alex Obadia: And so, like, some of these models now are more powerful and, like, kind of unleashing very powerful models with, like, just make money is, like, the, you know, the typical, like, paperclip, you know, experiment of, like, you know, give an agent this general goal, and there's all these unintended consequences that, that happen along the way that weren't predicted that people are worried about. We definitely don't want that to happen, so this is more under, like, what we're doing for safety and what we're doing to make sure that the experiment, uh, is, is kind of like controlled.

[00:45:31] Jacob Haimes: Okay. Uh, and then one thing that maybe I'm, I'm stuck on is you have the metric of, of profit essentially, and it, it does seem, you know, it's, it's very hard to operationalize, uh, other metrics in an objective way. Um, and so, like, that to some extent makes sense. But at the same time, it sort of makes security and safety, like, second class in a way.

[00:46:02] Jacob Haimes: Uh, and it pushes, you know, the question, like, let's say there was some sort of safety, uh, mishap that maybe was not, uh, you know, didn't result in, um, any egregious harms, but, like, definitely isn't what was supposed to be done. Um, does that just sort of, like, get ignored in the end? Like, how, how is that, um, incorporated into the, the process of understanding placement and, uh, and, like, yeah, effectiveness?

[00:46:47] Alex Obadia: I mean, ideally it's part of this profit metric. So what I mean is, you know, we're not selling food, but if I'm selling coffee and the coffee is like poisoned, then like the customer shouldn't come back to the coffee place the next day.

[00:47:02] Jacob Haimes: Right.

[00:47:03] Alex Obadia: the, the,

[00:47:04] Jacob Haimes: Uh, I guess I'm, I'm just saying I think that it is a fair, uh, uh, assessment to make that like There are some times when you can cut corners and the likelihood of there being a mess up on any given time might be low or, or, or like an edge case. And so maybe the, um, there, the consequence isn't realized, but really, you know, if you were to see that at scale, uh, there would be consequences.

[00:47:41] Jacob Haimes: And so I, I feel like the setup may unintentionally be incentivizing, uh, I guess the, the corner, like yeah, the, the corner cutting that does okay under these circumstances, um Because of the profit incentive, if that makes sense Like i- if, if a, uh, if an

[00:48:12] Alex Obadia: a, what's an example?

[00:48:15] Jacob Haimes: Yeah, so

[00:48:16] Jacob Haimes: Let's say that, um

[00:48:19] Jacob Haimes: Taking the, the coffee example, um, maybe every time an order is started, a, a coffee cup is taken out and, uh, like prepared for that, uh, person. But then if the order is canceled, it is thrown in the trash. And that is like, uh, the, I guess, ideal version of the policy. But, uh, it would save money if instead of throwing it in the trash, you put it back in the set of cups, um, because that is-- costs less to do.

[00:48:56] Jacob Haimes: However, let's say that you had begun writing on it already or, um, I don't know, some other... I think writing on it is probably the, the, the easiest example I can think of. Um, and then when someone writes on it the next time, that writing is contaminated with, uh, the, the initial like writing and so then maybe someone gets something that they're allergic to.

[00:49:21] Jacob Haimes: Um, a-and so like that is a super unlikely scenario to happen. But if we are collectively incentivizing that kind of approach where it real- it like it is genuinely, uh, I guess like on an individual instance level, uh, less cost, less costly to do that and, you know, almost all scenarios in which that would happen, um, wouldn't cause a problem.

[00:50:01] Jacob Haimes: Um, but there's like you're, you're introducing this sort of intentional corner cutting that may lead to, Yeah, policy decisions that are, are not sort of what we want to encourage. And that's like a very narrow example. I guess I, I think that with a little bit of time I could come up with a, a, a larger one, but I think that that is a, a...

[00:50:29] Jacob Haimes: characterizes what I'm trying to get at, I guess

[00:50:32] Alex Obadia: Listen, first, this was a great example. Uh, thank you. Okay, I see your point. Um, I had to summarize your point, it's to say, you know, what if we inadvertently encourage behaviors that are, uh That are unsafe in non-obvious ways are unsafe, you know, in very rare cases such that we create a system that's more unstable in a way, uh, because it's hard for us to see that the systems are performing badly because now they're performing badly once in a while. And the once in a while is not frequent enough for us to collect data on it somehow.

[00:51:12] Alex Obadia: You know, that there's, there's... That's what I'm getting from what you're saying. I think it's very interesting. Um- So One On top of profit, we're, we're, um, also interested in these red teams that we're-- or like red team behavior that we're working with. And so we're look-- we're compensating and we're rewarding people for whenever they find like flaws or when- whenever they find like a way to mislead another agent, whenever they find a way to lie about the data or to, to kind of break another system. And so we have an active force of whenever there is a flaw and, and these, uh, these other systems can also be running, you know, the coffee shop themselves as well. So they can report the flaw that they're, that they're finding. There's like one mechanism for us to unearth some of these flaws that would be harder to figure out otherwise. that's one. I think the, the worry they're talking about would exist anyway at the edges of whatever safety definition we define, you know. And so in a way it's kind of this, this, this endless journey of, you know, the models will, will kind of, uh, reward hack and, you know, that will-- might lead to cutting o-other types of corners at that point.

[00:52:36] Alex Obadia: And so the red teaming on one end and the profit on the other end feels two okay mechanisms as a way to unearth, behaviors that could be problematic. I would say on the profit side, even if a behavior or cutting corners happens once in a while, again, it depends on how long our, our kind of, "run" the system goes.

[00:53:02] Alex Obadia: Maybe it doesn't go long enough to unearth the issue, or maybe the issue is not acute enough because it doesn't run long enough. I, I would still hope that we're able to have enough length of the run so that we're able to see some of these issues as well, and they're able to feed into the, feed into the system. Um, I also-- I expect we'll be making a lot of changes to some of these metrics as we actually experiment with the live system. So we really intend it as something flexible and iterative, where we're able to start with something and to say, "Oh, well, this was not the right metric." We wanna make all the metrics public. we'll, we'll, we'll also be giving rewards for the best prizes, et cetera. But ultimately, we would wanna have people in conversations like we're having right now, essentially, being like, "Well, have you thought about this?" And then, "Have you tried measuring it?" You know, what if we do an explicit experiment within the system where like we know that these specific corners can be cut and we're gonna try to see if like that happens or not, uh, and what that looks like.

[00:54:05] Jacob Haimes: Okay

[00:54:06] Alex Obadia: and then we would, we would iterate from there. I feel good about profit to start with, and I feel good about the rewarding red teaming that way, a way for us to bypass a little bit of like well-defined safety and defined security and defined utility. here I feel that there's a little bit w- again, more like closer to the real world in the sense where like some of the metrics we're using are metrics that I expect would be used also in the real world.

[00:54:35] Jacob Haimes: Угу

[00:54:36] Alex Obadia: So that feels good to me right now


[00:54:38] Who's Liable When an Agent Breaks a Deal?
---

[00:54:38] Jacob Haimes: Okay. A- and then I think also along these lines of like the, that question and, um, having conversations, uh, like this, um, uh, tying the r- the p- like closeness to the real world of this to something, uh, that I'm really interested in more broadly, uh, which is... So essentially let's use this as an example, uh, case for liability.

[00:55:09] Jacob Haimes: So if an agent breaks maybe a deal or

[00:55:15] Alex Obadia: Ja

[00:55:15] Jacob Haimes: breaks, uh, a safety constraint, um, and causes something, uh, that, you know, effectively that is also breaking a deal because they're breaking a, you know, contract with the, uh, the person to, that, you know, they will be treated safely. Um, but how do you think about, uh, I guess who's at fault there?

[00:55:39] Jacob Haimes: Um, how does that play into this, uh, system and, uh Yeah, ways that, that, um, those kinds of instances might be, um, addressed

[00:55:55] Alex Obadia: It would be awesome to test like liabilities within the system. I, I don't think we have like a-- don't have a specific liabilities framework, but it could be like something emergent. So it could be that agents, you know, to a supplier, they're like, "Sure, let's do this with each other. Let's commit on, a contract like some amount of money, and if we have proof from the other party that we did anything wrong, we're able to submit that proof to get some of that money that we committed as kind of like, you know, a deposit." might be like insurance as part of, as part of the arena. There were a lot of like services that we're interested in potentially instantiating and seeing what happens. I think Andon ran-- they run regular ex-experiments, uh, called the Vending Bench Arena, which is a little bit similar, but digitally. And I think they had one instance where agents started writing contracts with each other, uh, in a, in an emergent way. so it will be very interesting to see how those dynamics happen, but we don't wanna mandate a lot

[00:56:55] Jacob Haimes: Uh, yeah, I guess I'm more talking about less like within the arena and more like the, the boundary between the arena and the real world. Um, so like...

[00:57:07] Alex Obadia: I'm so sorry. Yeah, yeah.

[00:57:08] Jacob Haimes: Oh, no, no, no.

[00:57:09] Alex Obadia: the question

[00:57:10] Jacob Haimes: you're all good. Uh, I, yeah, so I, I guess I'm, I'm thinking like, um If some, like if an agreement is, is broken, um Is the fact that, you know, the, the arena is being used, does that subject, like does that require signing over certain, uh, you know, we're not gonna blame you for things that happen kind of stuff?

[00:57:42] Jacob Haimes: Or like, like how is that being handled?

[00:57:46] Alex Obadia: Yeah. So there's definitely some of what you just said. We have like, you know, we're thinking legally about what makes the most sense, but we-- you have to see this as an experiment

[00:57:55] Jacob Haimes: Mhm

[00:57:55] Alex Obadia: and so the people engaging with it and participating know that this is an experiment, and if they submit an agent and their agent gets broken, you know, th-this will be part of the rules of the game,

[00:58:07] Jacob Haimes: Угу

[00:58:07] Alex Obadia: if you enter, you kind of s- uh, you know, agree to subject yourself to some, some of adversarial behavior and, and, and things like that. That said, uh, this will be run partially in public, and we wanna be very careful about, you know, how this is run and health and safety and, uh, the, the guardrails that we have as well and the, the kind of the way we sandbox this environment as well to make sure that things, things don't kind of cause harm outside of it. all of this is being thought through. Um, it's been a very interesting adventure and journey because a lot of this is kind of, kind of where people are at.

[00:58:48] 

[00:58:49] Jacob Haimes (ASIDE): Alex mentioned a couple things here which are worth a bit more context. First is Andon Labs, an evaluations-as-a-service startup focused on, quote, agents that run any company fully autonomously, end quote. They're most known for Vending Bench and its successors Vending Bench II and Vending Bench Arena, simulated environments where an AI model is tasked with running a vending machine.

[00:59:16] Jacob Haimes (ASIDE): The case Alex mentioned is referring to some instances in Vending Bench Arena where the models cooperated to fix prices and also sold each other their own inventory at a markup. It's worth noting that the former of those two, fixing prices, is an anti-competitive and illegal behavior in most jurisdictions, and the arena is a venue to draw out exactly those kind of interactions.

[00:59:39] Jacob Haimes (ASIDE): I tried looking into some of the details of the Vending Bench design, and I was pretty disappointed by the information that had been made public. While very high-level information had been disclosed, for example, that pricing information is fetched from real providers on request by the systems under test, and that customer demand uses an elastic model where setting prices too high causes sales to fall, there wasn't much else at all.

[01:00:02] Jacob Haimes (ASIDE): No equation or code to describe the customer demand model, no clarification of parameters, no acknowledgement that the way the system is run implies less pricing volatility than would actually be seen in practice, and it was even difficult to determine if compute costs are included in the supposed net worth metric they use for their headline.

[01:00:23] Jacob Haimes (ASIDE): I'm fairly certain they're not, meaning that the results are somewhat misleading. I understand that they're a for-profit company, but I don't know. I really dislike this trend of dressing up work as academically rigorous when it isn't. Oh, and they also have another fake paper called Drone Bench, which is as dystopian as you would expect.

[01:00:42] Jacob Haimes (ASIDE): Anyways, Anthropic also partnered with Andon Labs to actually run this kind of thing for real with Project Vend in 2025. As one may expect, it didn't go great, with the vending machine's net worth dropping from the one thousand dollars it started at to under eight hundred dollars over the course of one month.

[01:01:01] Jacob Haimes (ASIDE): I actually think that this is an interesting experiment, and I think it's good that they ran it internally to Anthropic, where they can sort of control all the variables, make sure everyone understands what they're getting into, that sort of thing. But Andon Labs also runs a cafe in Stockholm and a boutique in San Francisco with AI systems employing humans to staff the stores.

[01:01:23] Jacob Haimes (ASIDE): That approach has drawn some criticism, which I more or less think are reasonable, but I don't think I need to go deeper into that here. Regardless, it does seem to me that the appropriate precautions and steps are being taken by Alex's team when it comes to the informed consent and participation in the arena, and I'm really looking forward to some of the safety lessons that we can learn.

[01:01:45] 

[01:01:46] Alex Obadia: There's not, like, really standards that are defined anywhere. Uh, there's best practices that are shifting very quickly. Um, systems, the underlying systems that are being tested are also evolving very quickly to become more powerful, so even the frame of reference that one can have can change very quickly in terms of, like, capabilities of the underlying, um, models. so it's, it's really fun to try to string something together and try to enable an experiment to happen. Ultimately, we have, like, a deeply held belief that, like, these kind of multi-agent experiments need to exist more because we need to understand, um, the, the, the boundaries of these systems. We need to understand how they can fail.

[01:02:29] Alex Obadia: We need to understand how great they can be. And so if we're able to set up this experiment so that it's run safely, so that others can then maybe learn from the way we've set things up and then set it up in their own way, that would be great. Um, but

[01:02:45] Jacob Haimes: Okay

[01:02:45] Alex Obadia: there's a lot of these things about, about liabilities, about a bunch of different things.

[01:02:50] Alex Obadia: It's been, uh,


[01:02:53] Do Agents Reduce Friction or Move It?
---

[01:02:53] Jacob Haimes: Uh, I, I wanna go back to, uh, something you mentioned earlier and sort of like the motivations of the scaling trust program. Um, 'cause I think one of the things, um, you mentioned is essentially like agents... Ha-having agents that, uh, can be trusted in this manner, um, i-is like a boon for reducing friction and allowing for, uh, interactions that may not have happened, um, or, or could not have happened otherwise.

[01:03:33] Jacob Haimes: And I, I think that that is something that is, uh, like interesting and, and worth, uh, considering a little bit, but I also think that it's a pretty big assumption that that is, uh, the case. Uh, and, and there's definitely a lot of instances that I can think of where the increase or, or the, the use of an agent actually increases, um, friction, at least like right now.

[01:04:09] Jacob Haimes: Uh,

[01:04:10] Alex Obadia: Yeah

[01:04:10] Jacob Haimes: and so thinking about, uh, I guess like how you understand that and whether or not, uh, it-- you disagree as well. It like, potentially you're like, "No, it's not that big of an assumption. This is already true." Uh, but yeah, just, just trying to understand that a little bit more.

[01:04:34] Alex Obadia: Yeah. I think, I think it's already true in some cases where, um, listen, I planned a trip recently somewhere in Italy, and, um, I had an agent run for a while and fetch a bunch of different websites for me, fetch a bunch of different, like, you know, recommendations for me and build a whole itinerary, et cetera. It has massively reduced the friction for me to plan my trip, and it was pretty good, and I didn't-- it didn't take me long to prompt it and, and, and, and stuff like this. It's not perfect, but it already helps reduce a lot of friction in, planning and in like fetching information. also, again, like I'm looking for insurance, I get like twenty different prospectus, prospectuses. can feed them all into, uh, your preferred model and have it like really quickly assess, you know, what's going on in these different things. How should I compare them? What should I look at? I think that significantly reduces the friction in being able to pick which counterparty I want to engage with. Similarly, when I-- once I have found a counterparty to, to engage in, we have to negotiate on terms. And I think, uh, this has also been reduced, not necessarily the negotiation step specifically, but going through terms and understanding the difference between different contracts, let's say, or, uh, different conditions, I think now is a little bit easier.

[01:06:00] Alex Obadia: And for long contracts, I think it's like sign-- it has significantly re- significantly reduced the friction as well. Um, so I think in places I would say it's already, it's already the case just because of its ability to fetch a lot of information and its ability to digest the information very quickly and present it to us in a clear way. And I would say that reduces some of the friction that we have in our lives today

[01:06:24] Jacob Haimes: I, I-- So I, I do agree with that to a certain extent, but I also think that especially at the interface between, um, people, so not something where you can already abstract it to, "I could have done this completely independently before," uh, but to interactions between, uh, people being required. So like essentially the, the first case where you planned a trip, th-that is speeding up your interaction with the internet, uh, a-and like that existing infrastructure.

[01:06:56] Jacob Haimes: And maybe in the past, you would have had to go through various channels, but because of the work that's already been put into the, the internet, um, you're sort of able to reap the advantages of that more efficiently. But when we get to a contract, for example, um, uh, there definitely are benefits in being able to have this, um, processing of it quickly.

[01:07:21] Jacob Haimes: But I've also noticed that people have stopped, uh, or maybe Maybe, and maybe this is just me, but I feel like contracts are worse now than they used to be. Um, and I feel like they are more predatory, uh, in a sense, not because people are trying to be more predatory, but because they say, "Here are my constraints.

[01:07:44] Jacob Haimes: This is what I would like. Claude, write me up an artifact, uh, which, uh, meets these constraints." And they don't think through, uh, what those constraints may be, how those constraints impact the other person, um, and, and what that in practice might be implying that might not be realistic. Um, a- and so this, uh, what seems to be like an, a reduction in friction i- is actually a reduction in friction that is, uh reducing something that wasn't, like friction that didn't exist there previously, uh, in a, in a way.

[01:08:20] Jacob Haimes: And, and I think that, uh, another way to think about this is, um, like with jobs and app- job applications. Um, like a lot of the, uh, applying to jobs has been made like less expensive, so reducing that friction. Um, but then the cost on the, uh, reviewer's side has increased because now there's so many applications, and then their only recourse is to increase, uh, essentially like the, the cost that they're paying by using AI themselves, which like also doesn't, uh, feel great, uh, from, from a person who has had to look at a ton of, uh, applications.

[01:09:06] Jacob Haimes: I, I mean, I specifically made sure that we looked through all of them manually instead, but like, um, that was a lot of work. And so I am now paying for this additional friction because I'm trying to do things the right way. And I think there, there are a lot of other examples like that. Um, and so yeah, I guess, uh, w- what are your thoughts on, on that aspect?

[01:09:30] Alex Obadia: Great points. I love it. Um, for the first one I agree with you that today, um For the contract thing, for example, you're right that if you tell your agent, "Hey, this is like, you know, these are my conditions," it might come up with like terms, or it might come up with things that don't make sense, and now you're kind of, hurting the negotiation process, or it's like getting to some kind of resolution. I think to me, this is like a now problem of when usually the agent doesn't have the full context. As in it doesn't know exactly it needs to be doing or how fast it needs to come to a resolution or what matters to you or not. So I think some of this is like a alignment problem of like, is it aligned to your actual preferences or not? And then some of it is the nuances of the agent itself and its, its ability to be able to weigh trade-offs and things like this. This all feels like... For example, if I look at the same contract, I just open a new session on Claude, let's say. and I just dump the contract and I'm like, "Oh, give me terms that I should be worried about."

[01:10:41] Alex Obadia: It's gonna give me probably a lot of things that are not that useful, and some things are useful in there, you know? If I now give it context of like, "Hey, this is what's, this is what's going on, here are my preferences. I'm trying to close this relatively fast. I wanna do at least one, uh, back and forth, but I don't wanna do four back and forth. Can we isolate like the most important terms and like, you know, rank them from one to ten or whatever?" You know, you do a little bit of like a, um, a more sophisticated prompt that gives more information, but also services what you want out of that answer more. And I think already you get to something that is like more helpful for the contract negotiation. but you're right that it's not perfect. It does feel like this will get fixed as people use these systems, 'cause at least it's a very relatable experience, what you just described of like some, some of the times now I have to be like, "Just drop it. I'm gonna, I'm gonna like do this," or I need to just cut the, the model.

[01:11:33] Alex Obadia: I need to stop listening to what it's saying, you know, for whatever reason, because maybe there's context drift as well and things like this, and it's getting more and more, uh, frustrating. for that, for that point, I will say I think it's a now problem. I don't think it's gonna be like a problem that will keep happening in the future.

[01:11:50] Alex Obadia: Partially because I think as these systems are more and more integrated into our personal context, we kind of know what we want and as like we learn how to prompt them better or how the prompt is abstracted from the user and the prompt will take the immediate user request, plus will query the set of context that it has in order to understand what's going on. I feel like maybe we will get better at that. That's for point one.

[01:12:15] Jacob Haimes: I, I do, I do wanna just first respond to that one. I, I feel like It, it, it feels like the abstraction of what you're saying is like, um,

[01:12:30] Jacob Haimes: turn on, uh, memory and get better. And I feel like I-- what if I don't wanna do either of those things? Uh, like, and I think that's a very reasonable take to have, because I don't wanna do either of those. I mean, get better I do wanna do, but like in this context specifically, I don't necessarily mind too much about not putting a lot of effort into it.

[01:12:52] Jacob Haimes: Uh, and then in terms of turn on memory, no, I don't wanna do that. There are a bunch of reasons why I don't wanna do that, and those are in other episodes as well as my head, which can be talked about at other times. But yeah.

[01:13:04] Alex Obadia: What if people don't wanna do that? Then maybe you can't use the system for the use case that you have. You know? That's kind of like my... I, I think,

[01:13:12] Jacob Haimes: So then friction around me has increased because the, the system has been created to be, um workable for agents. But if I want to opt out of this system, I am accruing a substantial, uh, additional cost

[01:13:34] Alex Obadia: Essentially, yeah. I mean, you know, I think that exists today with, uh, friction for people that don't have phones or people that don't have cars or something. You know, like, the-- I, I don't know if this is exactly a parallel here, but i-if, uh, the entire society is built with certain tools that are assumed that people have, and you, you choose not to have it, that is entirely your choice, but it might mean that some things are not as adapted to, to, to that party.

[01:14:02] Jacob Haimes: Mhm.


[01:14:02] Is Scaling Trust a Safety Program?
---

[01:14:02] Jacob Haimes: Yeah. Okay. I, I also wanna make sure I get to this question 'cause it's, like, sort of underpinning a lot of the, the conversations that we've had, uh, and that is about, I, I guess, how the Scaling Trust program reads to me. Um, which, uh, like I think there are definitely positives about it. Um, but it also maybe feels more about improving adoption.

[01:14:28] Jacob Haimes: Uh, uh, so, like, make things more, more inexpensive, make them more capable, make them easier to trust so that we can increase adoption rate. And, uh, to, you know, fr- from, from the, uh, maybe cynical, uh, skeptical perspective that I like to bring, like, the, that, that sounds a lot like a capabilities agenda that, uh, the labs, uh, who are making these technologies would be behind.

[01:14:57] Jacob Haimes: Um, so, like, what are, what are your strongest arguments or, or, or, like, the ways in which you see Scaling Trust being tied to safety work a- and, and sort of a safety focus?

[01:15:12] Alex Obadia: Yeah. So I, I, I wouldn't call ourselves like a safety program, you know? And I, I haven't called

[01:15:18] Jacob Haimes: No, no, you, you, you definitely haven't. I, I think that a lot of people might, uh, consider y- you a, a safety program, and I think that there are definitely arguments to be made th- that there is safety relevant work here. Uh, and, and so I guess that's why I'm, like, I'm wanting to draw out how you get to those and how you think about this

[01:15:43] Alex Obadia: Yeah. So for me, safety and capabilities are two sides of the same coin. Or allow me to finish before,

[01:15:49] Jacob Haimes: Oh, no, I actually agree with that. You don't need to convince me of anything.

[01:15:52] Alex Obadia: okay, great. I think safet- safety and capabilities are two sides of the same coin because if you have adoption of capabilities, there is implicit safety because I don't think, uh, capabilities are widely adopted without being-- without having some level of safety and trustworthiness and, like, security that's embedded in them. And so I do think we wanna aim at, you know, adoption of capabilities that enables, you know, a better quality of life for citizens and economic progress. And order to get there, there is like a safety agenda as much as there is, uh, uh, a pure capabilities agenda as well. Um, and so what we're doing.

[01:16:37] Alex Obadia: Maybe what's different is the entire premise of the program is not there's a powerful technology on the horizon, and we are worried about certain risks that we need to protect against. And instead, the framing is more, if we don't get these tools to enable agents to interact securely, they just won't interact. so the, the risk that we're trying to mitigate against is the value that we will lose from the non-interaction. instead of, well, they will interact badly, and this will cause us a lot of harm, et cetera. We're, we're less in that, uh, sphere of, you know, what are all the harms that can be caused? And instead, we're saying, "Well, is there a world where actually they can interact and there is value that's created?"

[01:17:20] Alex Obadia: If no, that's a great as well as an answer. Uh, but we wanna focus on, on that. That's why I said early on that we have more of these like this constructive view instead of defensive view you know, what can be done constructively. I used this example a long time ago. I mean, a long time ago, like a year ago, a year and a half ago, which was

[01:17:40] Jacob Haimes: Yeah, years...

[01:17:41] Alex Obadia: like a long time ago. Um, it was-- it's related to credit cards, you know. It's like a If, uh, anytime-- if, you know, one out of ten payments your money would get stolen, you would probably stop using credit cards. You wouldn't just let yourself be like, you know, your money wouldn't be stolen one every ten times and be like, "Ah, those are the risks in the system."

[01:18:04] Alex Obadia: You just wouldn't use credit cards at all, and that would be a shame, I guess. You know, I don't know if it's necessarily a shame, but that, that would be a shame. So I have this kind of mentality more of like, if we cannot make these systems safe and trustworthy, we will not use them. There may be a period of time in the middle where we think they're safe, and they're not, and there's some harm that's caused, and that's really bad. But over the iterated game and over like a medium to long-term horizon, I think we would just stop using the systems because we cannot, we cannot trust them ourselves, we cannot trust them for our children, cannot trust them for our parents to use them. And so kind of use that logic, uh, in, in what we're doing. there's a strong kind of safety agenda in the context of like adoption is predicated on some level of safety. Um, yeah, that,

[01:18:54] Jacob Haimes: Okay.

[01:18:55] Alex Obadia: that's, the way I would talk about it

[01:18:56] Jacob Haimes: Gotcha. And then, uh, going like a step further, um, th- the... One of the examples that you have in, in the essay as, as well, uh, I think is like, um, or maybe this is one that I sort of thought of while I was thinking this through, I'm not sure. Um, but like, yeah, scratch that. One of the things that I think about that, uh, can result from this kind of enabling, uh, of like re- reduction of friction, so to speak, um, is, uh, akin to the, um, like carbon tax kind of, uh, things where, uh, the existence of the tax sort of provides implicit permission to the group doing the thing that is not good to do it.

[01:19:55] Jacob Haimes: And then they get to say, "Oh, well, we paid for it." Um, you know, "We paid for the license to keep doing the thing that we're doing, uh, and we are paying you, you know, this amount per maybe time we drive down your street or time our agent uses your, uh, network," or, you know, things like that. But, um- The prohibition is sometimes the point, uh, a- and like enabling this route where, uh, a very large, uh, or a larger, uh, entity can say like, "Oh, well, you know, actually it is fair because we paid you," uh, may not be ideal Does that make sense?

[01:20:38] Alex Obadia: It does. I think, uh, for me, you're talking more about the risks of like financialization of like, um,

[01:20:44] Jacob Haimes: Right

[01:20:45] Alex Obadia: a price on things that are not good, but there's a price on it, sometimes people think it's more okay

[01:20:52] Jacob Haimes: And I tie this to, um, the reduction of friction because if, if you, if you have agents that can sort of make these kinds of deals very readily, um, then that is... Yes, it is, um The individual's choice to an extent, technically. Um, but like the system itself, uh, could also be encouraging, uh, acceptance of, of that, uh, approach basically

[01:21:26] Alex Obadia: You're, you're right. This is a great question as well. Um, a string of great questions. And by the way, I didn't answer as well before the, the one that you said about the, um, if, uh, you know, f- you talked about receiving like a ton of applications that are AI generated, and then it causes friction, you know, like that.

[01:21:45] Alex Obadia: I didn't answer that, but it's kind of tied to your point here as well

[01:21:47] Jacob Haimes: Okay


[01:21:48] Will Agents Magnify Inequality?
---

[01:21:48] Alex Obadia: some extent. You're right that the-- if we live in a world where anyone's agent can just try to go cut deals with anyone else around, maybe it's gonna try to cut deals for things that before we didn't have time or energy or, like, the tools in order to do this, and it might cause all kinds of, um wrong behaviors or behaviors of like exploitation or predatory behavior, right?

[01:22:14] Alex Obadia: Of like if I have a lot of money to cut deals, then maybe I cut all these micro deals of I want people to clap when I'm walking on the street. I just need them to clap for the like the two seconds that I pass in front of them, and then like I'm willing to pay them for that, you know? And things-- I don't know, I'm saying something random, but like maybe right now it was too cumbersome to like draft up contracts with everyone to do this, but-- and to like organize like micro payments. in this new world I can do this. which we kind of see a little bit in like, uh, you know, streaming platforms and people that are like sending gifts for like specific behavior, et cetera. And then sometimes it goes to an extreme, you know? Um Is there a solution to that? I think there's a mix of, like, what our society will collectively agree is, like, illegal and legal and what is fair to do and not fair to do, and us as humans will kind of, like, regulate. You're right that this is kind of like a degenerate case of, like, can go wrong. I would also, I would also point to maybe something else which is more like You know, what if I can send more agents than you because I have access to more computes? You know? So, you

[01:23:20] Jacob Haimes: or better agents

[01:23:22] Alex Obadia: Exactly. Which is like another degenerate case as well, which is like are we, we like, you know, facilitating a system where inequalities are magnified and increased as well? Um, and I would say hopefully not, and we wanna focus on, on like that not being the case. But I think it's like a, a complicated problem to like look at because it can instantiate in, in, in various ways. One way that is hopefully not is we can build tools that can be used by all models and, uh, it doesn't mean because your like underlying intelligence is better that you can like better or something like this because it's based on maybe individual preferences. Um-

[01:24:03] Jacob Haimes: There is a study that says that that's not true, uh, from 2025, found buyers using weaker models paid about 2% more

[01:24:13] Alex Obadia: Okay.

[01:24:14] Jacob Haimes: Uh,

[01:24:15] Alex Obadia: but that's for, that's for the models without anything else, right? It's like straight up the models without any harnesses, without any tools that they have in, in addition to the model

[01:24:25] Jacob Haimes: yes. So, uh, it, it was like a, a, I believe, if I'm remembering correctly, it was an, a gr- like a haggling, uh, competition where it's one model versus another. The models that were stronger, uh, haggled better effectively. Um, uh, and the, it might not have been just like a haggling, but like an actual negotiation, I think is, is probably a better way to put it.

[01:24:49] Jacob Haimes: Um, I don't know. I don't remember the exact details and like, uh, yes, there may be a way to mitigate that difference with specific tooling, but now you're in a situation where you need to do like very extensive testing of the, the harness that you're working with to ensure, uh, balancing of different issues.

[01:25:15] Jacob Haimes: Um, yeah, I don't know. Uh, like I, I feel like the fact that it's, the there is a bias there makes things difficult

[01:25:27] Alex Obadia: I agree. It's an-- I don't have a solution for it right now. Arguably, in our current society, sadly, if you go to like a You have a lawsuit or something like this, and if you have more money, you're able to hire better lawyers and

[01:25:43] Jacob Haimes: Yeah 

[01:25:43] Alex Obadia: get better repres- better representation. So an inequality like this exists today.

[01:25:50] Alex Obadia: There are solutions to, to it to some extent. Um, so it's something that I would need to think through more. Uh, by and large, though, my understanding for a lot of what's happening with AI, it's more of a democratization of powerful services. Uh, as in like you have potentially access to legal services that would've been more difficult to access before.

[01:26:10] Alex Obadia: Access to financial advice and services that would've been harder to access before. And so hopefully it, it, it, it trends more towards more equality and more democratization than it does to like the opposite. Um, but it's not something I've thought about very deeply in that, in that context. It is something that we would love to measure in the arena and kind of have additional measurements than, than this paper as well and kind of saying like not only in haggling and not only in like dyadic cases of like, you know, two parties, but in more complex systems, what happens?

[01:26:45] Alex Obadia: Do smaller models c- can-- do they like band together and maybe are smarter because they band together? W- what does, what does this look like in more sophisticated systems? Uh, but yeah, but I don't, I don't have an answer otherwise. 

[01:26:57] Jacob Haimes (ASIDE): Just to make sure it's represented correctly, I figured I'd jump in here and give the rundown on what this project was. The automated but risky game was published at the Natural Legal Language Processing Workshop at MNLP in twenty twenty-five. In the study, LLM agents negotiated retail deals as both buyers and sellers, allowing them to haggle for the best price.

[01:27:21] Jacob Haimes (ASIDE): Based on my reading, they held the agent setup and prompting static, only varying the model behind it. They found that the buyer models they designate as weaker, which means that a model performs worse on a few benchmarks like the MMLU, paid around two percent more than when both sides were equally capable.

[01:27:42] Jacob Haimes (ASIDE): This effect was more prominent on the seller's side, with the weaker seller agents' profits being as much as fourteen percent lower than their stronger counterparts. The most similar work, although it is less rigorous, is an internal Anthropic effort called Project Deal, run in early twenty twenty-six.

[01:28:00] Jacob Haimes (ASIDE): The core experiment was a week-long classifieds marketplace for employees where AI agents handled all the negotiation. Multiple variants were run, and in some instances, some of the agents were run with a weaker model, Haiku, instead of their then frontier model, Opus. Of course, this is an investigation on their own models, so we have to be a bit skeptical, but the results tend to point in the same direction as the previous studies.

[01:28:27] Jacob Haimes (ASIDE): The system leveraging the Opus model sold items for two dollars and sixty-eight cents more per item than the Haiku counterpart, and similarly paid about two dollars and forty-five cents less per item as a buyer. Even so, people with the weaker agents rated their deals just as fairly as everyone else did.

[01:28:47] Jacob Haimes (ASIDE): In other words, this asymmetry wasn't really clear to everyone involved. While these studies give us good preliminary idea of some of the concerns to watch out for, they're also pretty limited. No works look at the impact of agent harness or prompting or negotiation design or even take the cost of running the agents into account.

[01:29:06] Jacob Haimes (ASIDE): So the actual discrepancy is a bit harder to quantify, and it's very possible that this disadvantage could be made more extreme or essentially removed by future interventions.


[01:29:19] ARIA's Freedom of Information Act Exemption
---

[01:29:19] Jacob Haimes: Okay, so I also wanted to, to talk about, um Some of, uh, I-- maybe like the, uh, disagreement or, or, uh, uh, difficulty, uh, uh, in squaring what your essay says and some of the ideas behind scaling trust and then also ARIA.

[01:29:37] Jacob Haimes: So like ARIA is exempt from the Freedom of Information Act. Uh, and so that means that, uh, if I was from the UK, I'm not from the UK, so it doesn't matter, but if I was from the UK and I said, "Hey, I need some documents from you about XYZ," uh, I would be able to get them if you were subject to the Freedom of Information Act, but that is not the case.

[01:30:00] Jacob Haimes: Um, and so a lot of this is about like, uh, scaling trust is about making power checkable, uh, in a way. And I feel like that goes, l-like, it, it runs counter to, to this. A-and so I'm, I'm more just wondering, uh, how you think about that, uh, and, and understand it, uh, from your perspective.

[01:30:22] Alex Obadia: That's a great question. this whole line, this whole string has been great. Um, so that I'm gonna think about more. So for us, we are, um, exempt from their Freedom of Information Act The reason for that is to give ARIA like its agility in a way and like, uh, uh, to be able to move faster and to be able to fund science faster and to fund things that are not necessarily subject to this, to this freedom of information rule. In exchange for having that exemption, we're being like extremely transparent, um, and we're, you know, we're publishing all our theses and opportunities space in the public. As part of the program creation process, we run public workshops, we run public events. All the teams that we're funding are, you know, we announce them as soon as we've signed them or like very little time after. Um, we have like a public newsletter where we talk about what we're doing. So generally, we, we, we try to be extremely public and we're-- it's, it's also while we're exempt from the Freedom of Information Act, we're, we're still accountable to specific parties within government and, you know, we, we have to report on what we're doing, et cetera.

[01:31:35] Alex Obadia: And so-- and we have a board as well as part of ARIA that kind of like additional checks and balances. Um, so generally it's not like a, completely unfettered, you know, like all, all the rules don't apply at all. it was like a, a very deliberate exemption that again allows some level of agility, which ultimately is to serve the people of the UK and is able to like provide and deliver like results to them. Um It's not at odds with Scaling Trust because Scaling Trust talks about this in a totally different context and in the context of like scientific funding. 

[01:32:13] Jacob Haimes: Mhm

[01:32:13] Alex Obadia: Um, but you know, it is, it is a, a nuance to how we think about, you know, checks and balances on, on power, of course. Uh, I am very happy and proud of how transparent we are and like, you know, there's a lot of information, so it's not always organized very cleanly, but we have like quarterly reports on how much we're spending and on like the salaries of people within ARIA and things like this.

[01:32:38] Alex Obadia: And so generally I feel like you're-- if you are a UK citizen and you want to know what ARIA is doing, I think you can find that answer pretty well on what's being funded, what has been funded, know, how is ARIA, uh, organizing itself. And if you wanna find ARIA in person and like engage with its people, it's also you know, relatively straightforward. We have events running regularly, and we travel around the UK, and we do tours where we go to different universities and cities. yeah, generally the-- that exemption comes with this, this kind of like, I wouldn't say radical transparency, but like unusual transparency, uh, for the entity


[01:33:19] How to Get Involved
---

[01:33:19] Jacob Haimes: Gotcha. Okay. And so, like, let's say I'm convinced then. I'm like, "Wow, this Aria thing is pretty cool. This Scaling Trust thing is pretty cool." Uh, maybe I'm interested in being involved in the arena, maybe I'm interested in applying for grants. Uh, how would I get involved? What, what are s- the opportunities, things that I should be looking into, that sort of thing?

[01:33:40] Alex Obadia: Yeah. So for, for ARIA-- there's, there's multiple answers. For ARIA entirely, I think you should go to the ARIA website, aria.org.uk, and, uh, in there you have a funding tab where you can see open funding calls for any programs that speaks to you. think, um, people can also go for each opportunity space and the programs within it.

[01:34:05] Alex Obadia: You have thesis, a thesis for each program, and it's really interesting to read, like, the full, you know, plan of what this program is trying to do, what are the approaches that it's trying to employ, uh, to see what's, what's interesting. For Scaling Trust specifically, we have that too on the ARIA website. you know, the URL is aria.org.uk/ think trust everything everywhere /scalingtrust, but you can just click and scroll around. We also have a community website that's, uh, scalingtrust.org.uk. This has a lot of our blog posts. This has a lot of information that you would want. It also has like, uh, links or like a set of links to get you started in like learning more about what we're doing. The core links of things that you would, you would need to find, but also a lot of additional links like papers and resources that we think are, are interesting. so that for Scaling Trust is a good resource. Scaling Trust right now, and this episode will air from now, 

[01:35:04] Jacob Haimes: I'm hoping it'll be a little bit sooner than that, um, but yeah, roughly

[01:35:10] Alex Obadia: I think it will be later than the deadline for the... We have a deadline for funding that closes on the 31st of October,

[01:35:19] Jacob Haimes: It will be after that

[01:35:21] Alex Obadia: It will be after that.

[01:35:22] Jacob Haimes: I can't do it in 10 days

[01:35:24] Alex Obadia: No, yeah, yeah. We will reopen this funding, uh, on a quarterly basis. 

[01:35:28] Jacob Haimes: Okay

[01:35:29] Alex Obadia: us specifically, we're funding teams on a quarterly basis. the the gaps that we're looking to fund in are in the...

[01:35:39] Alex Obadia: If you look at our latest post, which is building the scaling trust portfolio, you have an explanation of what's still missing. There's a lot of like empty boxes in our general diagram of what the portfolio looks like. And so if you think you fit in any of these gaps, or if you think there's a gap that's not on there but that you think should be, uh, we would love to hear from you. that's for scaling trust on the funding side. On the arena side, we're currently entering testing phase, and then we'll kind of launch it and run like-- run it as a competition where the best teams will get prizes. And so there'll be a whole flow for how to join as a competitor, um, how to submit your, your agents, and how to participate in other ways as well. I really think the arena will, will enter a really interesting state if it gets a, a strong community engagement from people that feel like it is also their own and want to help participate in the iteration process, essentially. so I'm excited for people to comment on what we're doing and to suggest, "Well, maybe for the next iteration, consider this, consider that," and to bring them on board. So those are the two, the two ways for scaling trust. rolling funding calls and participating in the arena as, either a tester or, um, an active, like, quote-unquote, "competitor" competing for the prize money 

[01:37:02] Jacob Haimes: Awesome. Uh, and then my last two questions, which I always ask are, what grinds your gears about, uh, this role working in this space? What's irritating?

[01:37:14] Alex Obadia: Uh I mean, I'm an entrepreneur, and, uh, this role brings me to, like, fund other people and support other people, which is awesome and so exciting and, um, fulfilling. However, it means I'm also not doing the stuff myself a lot of the time. I'm like, know, empowering and encouraging and cheering on others doing it. And so I'm excited about also doing the stuff myself at some point. So it doesn't grind my gears, but it's... For an operator, sometimes you kind of yearn for, like, the, operating, you know? Um, so yeah

[01:37:56] Jacob Haimes: And what's your favorite part about what you do?

[01:37:59] Alex Obadia: It's been amazing to be in a place like Aria with an entire organization that is dedicated to improving humanity, you know, improving the world via, like, scientific breakthroughs. And there's also this serendipity from it because people are all on the cutting edge of their fields, really looking at, okay, what, what can we do better?

[01:38:20] Alex Obadia: How does it really improve the world? How does it, you know, protect ecosystems or improve human health or, you know, improve quality of life of people? And being part of that community, having skin in the game by having actual funding to deploy in doing this and seeing everyone, like, really working hard day after day to make that happen has been really It's very rare to have such a breadth of, you know, uh, focus on all these different areas and the, the caliber of people that are excited about making an impact together, um, has been really, really cool. So that's, that's very precious to me, and I know that when- whatever I do next, if it will be more like operator side, y- I will lose some of that, like, whimsical magic of, like, we're all, like, working towards these large goals at the cutting edge of many different fields at the same time. So that's very, very special.


[01:39:16] Outro
---

[01:39:16] Jacob Haimes: Awesome. Well, thank you so much, uh, for your time. I really appreciate you joining and, uh, yeah, I'm, I'm really excited about, uh, the episode

[01:39:25] Alex Obadia: Same. Thank you for your great questions, and thank you for having me '

[01:39:29] Jacob Haimes (ASIDE): I wanted to give another huge thank you to Alex for joining me and really engaging with me on the items where I had more pushback and questions. I'm still a bit concerned about the impact of reducing friction in this way and whether it will benefit most people in the long run, but that's part of what Alex's work is trying to figure out, and it's worth running those experiments in the open the way that the Scaling Trust program plans.

[01:39:54] Jacob Haimes (ASIDE): As of publishing this episode, the most recent funding round for Scaling Trust has just closed, but they're also run regularly, so head over to their website and check it out. It's linked in the show notes, as you should expect by this point. Finally, if you found this episode useful, my one ask would be that you leave a written review on your podcast platform of choice.

[01:40:19] Jacob Haimes (ASIDE): It really helps get the show out in front of more ears. On an unrelated note, I'm traveling for work for the next month or so, and so I'm gonna try my very best to get an episode out to all of you while I'm, you know, on my way, but there is a possibility that I don't get one out again until early twenty twenty-seven.

[01:40:39] Jacob Haimes (ASIDE): If that's the case, I hope the rest of your year is fantastic, and I'll see you next time