Well guys, uh, thank you everyone for joining. I think for those who don't know me, I'm Júlio Vasconcelos. Uh, I'm a founding partner here at Atlântico. We've been organizing these tech talks here, uh, mainly over the past few months. So I'm very happy to have my long-time friend, Mike Krieger, join us. I'm here in San Francisco, at his house, so we have people here from all over the world, which is going to be great.
Mike, let me give a brief introduction for those who don't know. Uh, I've known Mike since 2006, almost 20 years. Here's a fun fact: he made the first version of Peixe Urbano. Uh, but then his career just went up from there, right? Uh, Mike, as many of you know, was the co-founder of Instagram. After that, he set up another company, then joined Anthropic as Chief Product Officer, a role he held for a while, until he decided he wanted to get his hands dirty again and start building products. So he went to lead Anthropic Labs there, and today continues building new products within. We can go into a bit more detail about that.
Uh, and thank you for participating. Mike is from São Paulo, so our great Brazilian representative out here on the AI frontier. So I'm very happy. And for everyone who's curious about who else is on this call, who are the other three beautiful faces here with us, uh, what I also did, given that the idea here is for us to do a tech talk and have, let's say, a more technical level than I could reach, I invited three technical leaders, entrepreneurs from companies in our portfolio, great friends, who will help ask questions and dig deeper. Another thing I did, I have my computer here and everyone who sent questions before, I was able to summarize several of the most interesting themes and I'll bring some of that to everyone here.
Just to introduce each one a bit, Lucas Smaira, for those who don't know, is one of the co-founders of Veto and he runs all the AI research there. What I think is interesting is that Lucas spent many years before, I think almost six years at DeepMind as a Researcher, and has been working more in this AI area. For those who don't know, Veto is a research lab building reinforcement learning environments for recursive self-improvement. So mainly looking at coding, computer use, optimization, lots of science stuff. They sell these environments to the main frontier labs in the world and also apply these same techniques to bring self-improvement to agents within large enterprises. So, Lucas, thank you.
Mike, for those who don't know, uh, I think he's the only Brazilian who founded two unicorns. Uh, Mike was co-founder and CTO of Wild Studios first, and now he's CTO and C-founder at Enter. For those who don't know, Enter is a company that sells to large enterprises a legal AI solution, end-to-end, that has had incredible success in Brazil and now even outside of Brazil. So I'm very excited about Mike.
And finally, Víor Olivier, also for those who don't know, Víor is founder and CEO of Decate, an AI-native wealth management platform for investments, and before that spent many years as CTO of Nubank. And he was there from the beginning, since the first five or six employees. So I also think everyone here has many different perspectives, but I'll start with the first question and then pass it to you.
Great. Mike, thank you. Uh, again, let's start with a soft one here, just to warm up a bit. If you were to build Instagram today, what would you do differently, and also what would you do the same, which I think is interesting both in terms of the profile of people you'd look for, how you'd organize the team, and how you'd even think about perhaps the MVP you'd build, the size of that MVP to find initial product-market fit.
That's a great question. Hi everyone. It's funny because Instagram, more than most companies of that era, was already kind of... it wasn't AI-native, because that didn't exist back then, but we already built the company with very few people. It was just me and Kevin for the first few months. And the two of us, what we wanted to do was contribute from design all the way to implementation. And the good thing about that was that we were able to move very fast, right? Up until the launch, and after the launch we started growing users rapidly. It took about two months before we started hiring people, but that really helped in terms of alignment, not having to have so many conversations. We saw this also with the second company - at the time, Arcti was just starting, but we hired a lot of people from the beginning. We started the company not just with K but with six more people, and that was a mistake, because the first few months of any company you're trying to decide, well, this is going to be the direction, let's change the product like this, and every time we made that change it was like, well, now we need to see how it's going to be received by this person, this person disagrees, now it's going to take longer. So in a certain way, I wouldn't do much differently in terms of the team. I think one or two people with the ability to do any part of the product process, I think that's a good recipe for starting a company.
Now, in terms of the MVP, I think about this a lot because the MVP we built for Instagram and delivered to the first beta testers, that took about two months of us working on another product. We decided to pivot, we took a bunch of things from that first product, put them into the second one, but that process was part of the product. So I think if you have a product that isn't working, understand what's working in it, try something different. AI just accelerates the building between these moments. I think you still need to have that contact with the real world. So I think it compresses certain parts, but some parts have a kind of natural duration. But in terms of social products with AI, I think we're still at the beginning of this journey, so I don't know exactly how different Instagram would be with these capabilities. I think part of the value of Instagram and products like it, at least in the beginning, was connecting people without much technology between you and the person. So maybe the MVP would be similar, but built more quickly.
I have a question, continuing on the Instagram topic. I think Instagram found the great form factor of images as a big angle for new consumer products. Then comes TikTok with videos, and now in the AI world, what do you think about consumer products, right? I think chat interfaces are one thing, but do you think that's the end, or what are the next form factors in your mind?
It was interesting because what we saw during Instagram is that sometimes these other apps would appear that were like a variation of Instagram. So, Instagram but with sound, this was before video. Then Instagram with machine learning editing, in the very beginning when you could transfer your photo to look like a painting. Nowadays, what gets me really excited is something between apps and pages created by people. Something I've been working on a lot at Labs - we can talk more about this too - are our artifacts, which are basically things you can create with Claude, everything from a static page to an app with a lot of functionality, with a database, real-time, multiplayer. I don't know if this will be as common a creation as a photo, which is so easy to make and distribute. But I think there's something interesting - I saw this with my wife who was doing a collaboration between three different companies, and instead of using an existing solution, they designed an artifact together that became almost the operating system of that collaboration. After the collaboration ended, maybe they'll never use that product again, that product they created. But I find it very exciting, this idea of software that exists to solve an immediate problem and maybe doesn't need to exist forever, kind of like a photo on Instagram doesn't need to exist forever. It could be something more ephemeral.
There's also this angle, Mike. If you think about it, with AI and Claude, you can go from zero to a product that's extremely adapted to you and resolves your problems almost immediately, right? How do you see in the future this separation of value creation between Anthropic itself with the end consumer who can generate the products they need quickly versus an intermediate application layer that will try to specify a very specific use case and build things around it to sell to the consumer, versus even bringing intelligence to companies and being able to train models internally to make it more customized? There's this customization aspect both at the application level and in the models.
Let's talk about applications first. I think there's still value. Us at Anthropic, we don't want to make a product for each use case. Making a really polished and refined product for each of these verticals, I think it's very complicated to maintain that quality. So what we need to do are things that are very horizontal. For example, we launched within Artifacts last week, Claude Slides. You can make presentations within Claude. And what's interesting to see, if you go outside of our product, just outside of Claude, and ask Claude, just using Claude Code or CLIFEX, to create software for me to make presentations and show them and maybe export to PowerPoint, Claude will come up with something relatively good. But there was a great post within Anthropic last week that captured this really well: now with AI, we do the first 90% very easily, but the last 10% is still quite difficult. So what we ended up needing for Claude Slides is, well, what interaction do you do, the transition from one slide to another, how can you import other data, what should the UI look like exactly? So I think it's still worth investing in these products that are very horizontal.
And then, of course, you also give these tools to companies that are creating even more specialized or more vertical tools for them to do those last 10% on their side. In terms of model training, Anthropic's focus has always been on the more general models, and of course there are opportunities within companies or within workflows for a more specialized model. I think one ends up using the other. Jeff, JV, that got a lot of attention - I don't know if it was last week or two weeks ago, everything moves so fast in AI that you completely lose track of time - is a good example. I saw a demo just yesterday that was pretty cool, where someone had created an entire video game inside Claude, but in terms of the behavior of the characters within the video game, they used GV to make it faster. So I think this orchestration of models, even within virtual worlds, is something we'll continue to see a lot of.
Following the line of AI-created software, models today are coding and basically learning from everything humans have done, following the same patterns. They use the same interface we do and probably are applying the same patterns they used last year. Model generalization allows us to open interfaces that can now adapt, allows creating software that doesn't need to handle edge cases, nor anticipate future requirements, because it can generalize by either creating the code in real time or maintaining the code. But inside Anthropic, have you seen any emerging patterns of model usage that really change how software works, not just writing software the way we write it?
Super good question. We've discussed this - the eternal debate inside Anthropic is whether the future will be more like MCP, where the model uses tools that are kind of created for the model, or using interfaces. And I, every six months I kind of shift my perspective more toward one or the other. What I find interesting is that, exactly as you said, when it's using an interface, it usually follows that human path, but not always - sometimes it needs to be creative. One thing I remember, of course, that thing with Claude trying to solve a problem even without good tools - we had a copy-paste task, and it needed to edit text between cells but didn't have that capability. So what it started doing was using the navigation bar, where the website address is, as a mini text editor. It would edit the text there and paste it back. I was like, man, I never would have thought of that, but it creates this creativity.
So I think even with interfaces we see Claude being quite creative. When we first launched the capability for Claude to use computers internally, our first idea of what a good product would be was like a tutorial. So you go to Claude and say, look, I want to do this thing in Photoshop but I don't know how, can you show me? And we thought that would be a good product, but actually Claude could achieve the goal, but it did it in a completely different way - it would explore this part of the menu, do something different, manipulate it this way. So it ended up not being a good tutorial, but it's a good problem solver. This taught us that maybe that wasn't exactly the product we wanted. And it might be better not to look too closely at exactly how Claude does it, because it can sometimes be frustrating - Claude, why didn't you do it in the most straightforward way? I think what's still missing, and I think will be interesting, is what interface would Claude create for itself. And we don't have this investigation very well done yet, but I think it would be good to see, for example, the logs of how Claude is using a product so that the next evolution of the product can be more refined for that. I think it's almost like the classic illustration of a garden in a park where the architect designed exactly what that path would be. But then people say, no, I'm going to cut through here, take this shorter route. I think learning from these - we call them desire lines - for Claude, I think would be a really interesting thing in terms of interface.
I think building on that, Mike, we've been doing a lot of computer use at Veto, and several of these evaluations we end up doing - for example, a clear example here is we asked Claude to make some parts in AutoCAD, and it went there, opened a terminal, wrote the entire XML system, exported it, and the part was perfect. But many times it ends up taking shortcuts that are quite far from human intuition, right? Do you think that's a problem in the long run, that it starts deviating from what we ground in human experience and we lose that intermediate observability? Or do you think it's actually a feature that will allow it to expand further, and despite us not necessarily understanding the way it thinks, it ends up being able to do more things in the future?
Great question. For example, just yesterday - we have a feature in Claude, in the Mac desktop version, where you can teach Claude how to do something. You kind of record a video for Claude and it watches. Then I was testing this because I was interested, I'd never tested this feature before. I showed Claude how to create an event in Google Calendar. Then I went to see what skill it created, and it was using my MCP. Then of course, it was like, I showed you how to do it in the interface because it would be easier that way, right? The version you showed was kind of difficult. I said no, it was funny to see that. So for now, at least, we can have this conversation with Claude. Like, how did you decide to do this? No, I want you to really use the AutoCAD UI for some reason. Generally, I think it's positive that Claude can find these solutions.
What will be more interesting is how we monitor and how we have this visibility into exactly how it decided to do this thing. What's perhaps the scariest part is when the way it designs this solution is so beyond human in a way that we can't even understand exactly how it created the thing. I noticed this even - if you see Claude writing code in an existing codebase, it usually tries to follow the pattern of what already exists, so it stays quite human-readable. But Claude writing code just for itself - you can sometimes see this when it's writing code just to paste two things together - it's not really thinking about whether this code will be easy to read. It ends up being more like a bunch of lines you would never normally write. So this shift is starting to happen, where Claude is writing code for itself and not exactly to be consumed by humans.
A question about innovation, new companies - given that you're seeing things in the Valley, obviously within Anthropic, what kind of AI-native companies would you like to see emerging from Brazil? Where are we uniquely positioned to build at this moment?
One area where I think we're still in the early stages in the United States, but I think this is even more real globally, is in this area of what we call economic diffusion. How Claude can go inside a company and truly transform it. We talk a lot about this enormous gap between what Claude can theoretically do with good prompting or good access to internal tools and how most companies are still using Claude. I think closing this gap isn't just Anthropic's job - I think it depends a lot on the companies themselves. So whether it's in different verticals - I saw a recent example I found interesting of helping construction companies use AI more completely. I think it's more efficient to have solutions for these industries that become a big player than trying to either have us do it - we won't understand every industry - or having each company figure it out for itself, which I think would take a long time. Maybe there will come a moment when the models are so good and there's so much knowledge about how a company works that they can do it themselves, but I think that time will take a while. So I think there's an interesting opportunity in this economic diffusion. I'm not sure if that's the right translation, but how can we help these companies adopt models? I think there are two ways. How can each individual employee improve their own use of AI, but also how can the company transform its processes in terms of what we do every day, how we end up relating with our vendors, how we relate with different consumers, and how AI can accelerate each part of all of this. I think this is still a wide-open area for companies to be built.
Very related to this idea. I see that we started using, first of all, coding agents in a very single-player way, everyone with their own computer accumulating huge chat histories. Similar to what happens in a company, there are people who have to maintain complete context about what teams are doing. There need to be bridges that cross teams to carry knowledge from one team about how something works, how it connects, how a process on the other end works. I feel pain seeing that we're not leveraging the information being generated just from people's conversations with agents. And I feel pain that these loops aren't being closed, aren't being materialized, whether in new instructions, skills, or more things. Within these loops that can be done, do you see within Anthropic some that are clearly a reality, working very well, or do products still need to evolve for this to become mainstream?
Great question. Well, I'll describe a bit how Anthropic works internally, which is kind of the extreme version of this, but by design. Almost all work at Anthropic is done publicly. Of course, public sometimes needs to be within a compartment of people who have access to the model, for example, but relatively public within your group of collaborators. And that's why we do everything in Slack. Anthropic lives in Slack in terms of how we interact with Claude daily. 80-90% of the interactions people have with Claude within Anthropic are in the form of - we even launched a product exactly for this, which is Claude Tag. And when we released Tag, people said, oh, it's just another Slack, maybe I can ask it a few things, like what's our HR policy for vacations? But we actually use it for everything. In terms of - I was using it right before this Zoom - starting projects, resolving questions among colleagues. And what I find super interesting about this is that when a new person joins the company - I was talking with someone who started just two weeks ago at Anthropic. We have this idea of the spin-up, so it's like the colleague who helps you get integrated into the company. I was talking with him, I asked, how has it been for you in a very different way? And he said, for me it hasn't been that hard, because I can ask Claude within Slack how people have been doing things, and since Claude has visibility into everything, how the company is operating, it already has that answer. And it's not the answer from oh, last year we were doing this, but how we're using Claude right now at this moment.
So I agree 100%, but I think this multiplayer area is super important, not just for collaboration, but at least for closing the loop, for this company learning. And what it requires is for the company to kind of change its perspective on AI from keeping everything very individual, super private, and the person needs to change to operate this way. And when you start operating this way, you see the enormous advantage - the company can accelerate. Now, of course, for an enterprise that has been using technology in a very specific way for a long time, it's not a bit harder - no, it's much harder to make this transition. What we've started doing is trying to discuss publicly a bit more how we use Claude to see if we can inspire a bit more of this change. For me, just maybe the last thought here, every time I'm using Claude Code in an individual way, I'm kind of thinking in the back of my mind, man, it's bad to use it this way, because other people won't be able to see what I'm doing in this usage.
