00:00:00 · Bernardo Aceituno, Ruby Justice Thelot
[00:00:00] S0: I always say and I stag by this that AGI was achieved in February 2026, around the time Claude Opus 4.6 came to the market. And today there's hundreds of open weight models that perform better than this closed form model, just like running in a local computer. And that's mind blowing. [00:00:19] S1: Today, we are joined by Bernardo Aceituno. He's a returning guest today, and he is the co founder of Stack AI, which is now part of Asana and open weight models and how enterprises are actually adopting AI agents is what we're going to talk about today. So Bernardo, welcome to MTS. [00:00:36] S0: Thank you so much, Sofia. Pleasure to be back. [00:00:38] S1: Pleasure to be back. I think one of the reasons I like talking to you is because you know everything that's going on when it comes to open source and enterprise. Let's just give a quick recap. What happened in the last few weeks that really changed the game?
00:00:38 · Ruby Justice Thelot, Bernardo Aceituno
[00:00:38] S1: Pleasure to be back. I think one of the reasons I like talking to you is because you know everything that's going on when it comes to open source and enterprise. Let's just give a quick recap. What happened in the last few weeks that really changed the game? [00:00:50] S0: Well, many, many things. I think that more than the last few weeks, I would say the last year has been really like the advent of open source. Past few weeks we've seen just perhaps a Cambric explosion of open source models. We've seen Meta, Reflection, we've seen pretty much Chinese labs with new chemists coming out there that now are surprising benchmarks and beating frontier models. But what I think is actually you see the most impact is when you look at the ability to provide cheap inference, that is also frontier in performance when you look at enterprise tasks. I always say, and I start by this, that AGI was achieved in February 2026, around the time Claude Opus 4.6 came to the market. And today, there's hundreds of open weight models that perform better than this closed fork model, just like running in a local computer. And that's mind blowing. Yeah. I would say that, of course, the frontier will always like, you know, set the pace and, you know, always the open wave model will lag behind. But the open weight community will catch up. And, yeah, like, from you, we'll always be ahead. But at one point, you think start realizing, okay, what's my trade off on investment, governance, and residency? Relax.
00:00:50 · Bernardo Aceituno, Ruby Justice Thelot
[00:00:50] S0: Well, many, many things. I think that more than the last few weeks, I would say the last year has been really like the advent of open source. Past few weeks we've seen just perhaps a Cambric explosion of open source models. We've seen Meta, Reflection, we've seen pretty much Chinese labs with new chemists coming out there that now are surprising benchmarks and beating frontier models. But what I think is actually you see the most impact is when you look at the ability to provide cheap inference, that is also frontier in performance when you look at enterprise tasks. I always say, and I start by this, that AGI was achieved in February 2026, around the time Claude Opus 4.6 came to the market. And today, there's hundreds of open weight models that perform better than this closed fork model, just like running in a local computer. And that's mind blowing. Yeah. I would say that, of course, the frontier will always like, you know, set the pace and, you know, always the open wave model will lag behind. But the open weight community will catch up. And, yeah, like, from you, we'll always be ahead. But at one point, you think start realizing, okay, what's my trade off on investment, governance, and residency? Relax. [00:01:59] S1: Yeah. A lot of enterprises recognize that they should start using open weight models for different parts of the stack. But how do you actually do that? How do you think about what to integrate when you should still be using the closed frontier lab models versus the open source models?
00:02:14 · Bernardo Aceituno
[00:02:14] S0: Yeah. A 100%. Well, if you look at the the way an enterprise will adopt AI, they will have, I would say, two ways in which they bring AI to the operation. There's, of course, a more assistive AI usage, where they have like a AI assistant that is available to employees, whether it be Cloud, Copilot, ChatGPT Enterprise, etcetera. And at this level, it's very normal to rely on the Frontier Labs. Not necessarily for the Frontier model per se, but because of the application layer that exists there. Now we're seeing an advent of new tools that allow you to operate under Frontier... Under your own selection model. And the fact that you can translate that usage from a closed way model that lives into an unknown cloud, that lives in an unknown region, and where you don't have a complete visibility on what data it operates with, depending on your arrangement with that provider, makes it so that it's very attractive for enterprises to switch their usage to open weight. We've seen plenty of enterprises start switching their usage, especially after they start looking at the token bill, when it starts crossing $7.08 figures a year, it starts raising alarms. And that's the first category of enterprise usage. The second category of enterprise usage is what we call back office automation or workflow automation, where they take processes that are persistent across the organization, and they place agents in different places of those processes or translate just a full process into AI agents in order to transform parts of the operation. This is what we do at Stack AI. And in there, the switch to open weight was very rapid, because most of these systems are built either with custom tooling or with AI development platforms, where often you have the ability to pick your own model, and you have every control on how you use tokens. When I look at some of my top customers in Stack AI, I think that there used to be a major trade off between open weight and having a closed force model, especially when I look at my military clients. They had to decide, okay, do I want good performance? I have to make a trade off and use OpenAI hosting an Azure GovCloud. But if I want full resonance and control in my data center, I can use a Mistral model or a Kimi model or whatnot.
00:04:09 · Bernardo Aceituno, Ruby Justice Thelot
Today we got into a point where for the purpose of these workflow automation tasks, they can go to, like, their latest open way model. Let it be let it be InkLink. Let it be Muse. Let it be Beam. And they can have just as good a performance as they would have expected in February this year when, for the purpose of enterprise automation, these models achieved this general intelligence capability. [00:04:30] S1: What's still not working? I think that
00:04:30 · Ruby Justice Thelot, Bernardo Aceituno
[00:04:30] S1: What's still not working? I think that [00:04:32] S0: when you look at, you know, going beyond the domains of enterprise world automation, and you look at operating in domains where, like, long term memory is needed, when you're looking at domains where very deep visual understanding is needed, which often are not pressing in the world of automation space to a level a human does, at those set points, both the open way and the frontier level models, that's not available. I think that there's a big topic on benchmarking as well. I would find that over the past three months, there's been a corrosion and the trust towards benchmarks in the enterprise just because of how much different providers benchmark their open way models to perform well in the benchmarks. And that's scary for them, because they don't have a good way to operate on it. And unfortunately, I have to rely on, you know, my favorite benchmark, is the BiTES benchmark. And at that point, you know, there's a big trade off to make into, like, do you trust the results of this benchmark, or do you, like, actually rely on a model? And I think open close frontier models and frontier labs still have that trust into what state of the art performance looks like. But I think that's gonna get closed, especially as better benchmarks and better consensus of the community is achieved.