The audio and video are in Spanish. These notes are a translation.
This episode is not an interview. It is a talk I gave at an innovation event about what I have learned evaluating artificial intelligence companies over the last three years as an investor in San Francisco.
The quote that best sums up that experience is from Elad Gil: the more I learn and the more time I spend on this, the less I understand. In most markets the opposite happens. Here the rules are being rewritten.
That is why the talk is called questions and not answers. They are seven questions I do not have an answer for, and an underlying thesis about why Latin America is being left out of the agents wave.
Structure and Key Highlights
From Lima to Mexico City to San Francisco
- I graduated in 2018 and I joined one of the first venture capital funds in Peru
- In 2019 I moved to Mexico, to Dalus Capital. My first week coincided with Finnosummit, the largest fintech event in the region
- David Vélez of Nubank was there, the founder of Creditas, Adolfo from Clip, people from Kavak. It felt literally like traveling to the future
- In 2021 I joined Reach Capital in San Francisco, expecting the same leap. It was the opposite
- What I found was FTX, Clubhouse, BeReal, and Quibi dominating the headlines. Interesting technologies with superficial use cases and little impact on the real economy
- A year later ChatGPT arrived, and that really was a magical moment: the first chatbot that did not feel like a robot
Two Years of Disappointment
- Between 2023 and 2024, most products built on top of AI were the opposite of wow: copywriting tools, chatbots, notetakers
- My reaction: it does not make sense that the most innovative technology in history gets used to write LinkedIn copy
- What changed starting in the second half of last year were two things
- First, OpenAI’s reasoning models, which unlocked a new layer of possibilities
- Second, two years of iterating and of understanding how this tool was actually being used
- The result is companies like Cursor, which does not make developers 20 or 30% faster the way a typical SaaS promises, but several times faster, because the tool does the work
- Harvey in legal, Decagon in customer support, and a wave of agents where the promise is no longer to make you more efficient, but to do the work for you
The Gold Rush You Can Actually Feel
- I had always heard older investors talk about the internet, the cloud, and mobile as gold rushes. I understood it in theory, and I had never lived it
- Since the middle of last year it feels like that: a hundred companies building tutors with AI, a hundred building customer support agents, a hundred verticalized in real estate, insurance, and financial services
- With value propositions that blow your mind when you hear them, not like the superficial uses from before
The Gap with Latin America Is Accelerating
- With one foot in Silicon Valley through the fund and the other in Latin America through the podcast, I started to feel that something did not add up
- In Silicon Valley there is no startup that does not use AI. In Latin America there were very few
- The data point that illustrates it most: Y Combinator made a brutal pivot, and practically every company in its latest batches is an AI company. Latin American teams are very few. In the last batch, one Colombian and one Brazilian
- All of this is anecdotal, but the gap accelerates because of the speed at which everything is moving
- Bill Gates’s line: we overestimate the power of technology in the short term and we underestimate it in the long term
- What I add to that: AI is redefining what long term means. What used to be a decade is now the medium term, or even the short term
Experts Do Not Exist
- The most counterintuitive thing about all of this is that almost everyone is just starting, even in Silicon Valley
- Unlike other waves, where the United States or Europe were years ahead of us, here we all start almost equal if we get on board
- The proof that not even the best know: Siri is still a disaster at the most valuable company in the world, Google had its problems launching Gemini, and OpenAI suddenly looked slow when DeepSeek released a model comparable to theirs at a fraction of the cost
- The three habits I try to keep: be skeptical of anyone who presents themselves as an expert, follow curious people on Twitter who try tools every day, and do not be afraid to say that I do not know
- The third one is beginner’s mindset, one of Nubank’s culture principles. David Vélez talks a lot about which questions you do not have an answer for
- That was exactly the question I used to build this talk
1. What Does an AI First Organization Look Like?
- In Silicon Valley today, teams of two to ten engineers build products as complex as the ones that used to require 100 or 200 people
- People always talked about the cost of adding people: communication, bureaucracy, internal politics. AI changes that equation, because the marginal benefit of one more person now competes with implementing tools
- The frame I liked, taken from one of the founders of HubSpot: before, organizations were structured by rules and by functions, and in a world of agents they are going to be fluid
- People interacting with agents, in dynamic teams assembled by task and not by area
- What is interesting is the consequence: if the best salesperson in the company is in another area, the organization will have to allow that fluidity of talent
2. What Skills Are Missing?
- Allie K. Miller’s tweet sums up well what I see when I invest. There are two extremes
- On one side, founders who are very strong technically, who understand the infrastructure and the frontier models, but who do not understand the business and do not talk obsessively with customers
- On the other side, the people who understand the business and not the technology
- What is missing is the middle: someone technical who understands the potential and also the limits, and where it does not make sense to use AI
- More of a growth mindset, because everything is being figured out on the fly, and MVPs have to be more fluid so they can adapt to the feedback
3. Who Is Going to Capture the Value?
- If the economy, or a market, grows 10 or 20% a year thanks to this technology, who keeps that? I do not have the answer, but I have three ways of thinking about it
- The model layer versus the application layer. When DeepSeek came out, people discussed whether models become a commodity the way it happened with the cloud, and there are arguments on both sides
- Horizontal versus vertical. I see generic voice agents for support, and specific agents for insurance brokers or real estate brokers. I think there will be winners in both, depending on the industry and the function
- Startups versus corporates. Corporates have brand and distribution. Startups have speed and a blank slate
- The nuance that seems important to me: even technology companies founded two or three years ago are carrying code written before AI, and that becomes a backpack
4. What Markets and Business Models Are Going to Emerge?
- The analogy is the iPhone. It did not only bring the internet. It brought a GPS good enough for Uber to locate you and the driver. Without that, Uber and Airbnb did not exist
- The question is which are the Ubers that AI is going to enable. Three lenses we use at the fund
- Creating new markets: services that did not make economic sense before. We are about to invest in a company that lowers the price of a service by 30 or 40%, and with that the market becomes much larger
- Unlocking blocked markets: removing barriers to adoption. Being able to use technology with your voice, or with a simple chat, brings in a lot of people that the interfaces used to leave out
- Selling the outcome instead of the tool. Instead of a CRM, they sell you a salesperson. Instead of a system for doing taxes, they do the taxes for you
- The commercial consequence of that is enormous: you no longer compete for a company’s software spend. You compete for the spend on salaries and on services
5. How Do We Compete When Everyone Has the Same Technology?
- When you look at company number 99 building a tutor with AI, they all have the same voice, the same latency, and a model that is equally intelligent
- The idea I cite, from a Sequoia investor: differentiation is not going to be the AI itself, but how creatively you apply it
- Data: it remains valuable only in specific contexts. The model already has the whole internet, so the advantage sits in data that is truly proprietary to a particular process
- Distribution and brand: if everyone can build very good software, what is left is who knows you and whom you trust. If there are 20 Microsofts, it matters which one is the known one
- Speed and culture: it sounds like a cliché, but today it is literal. I see companies in our portfolio from before ChatGPT that find it extremely hard to make the change, and they are small teams
6. What Is the New Interface Going to Be?
- The obvious thing would be to say that everything is going to be chat, because ChatGPT left us with that image. I think that is one extreme
- A founder in my portfolio was telling me about generative UI: mixing chat with traditional interfaces into something hybrid
- The other extreme is to stop interacting. Last week we invested in an autonomous ERP: instead of administration filling it in, the agent collects the information and updates it so that you review it later
- Three directions I see: multimodal (voice, video, even touch), immersive internet, and more natural interfaces
- The third one seems the most interesting to me for frontline workers, in factories and at points of sale. Traditional software never penetrated there, because it did not fit their workflow. Voice and video do
7. Which Behaviors Are New, Which Get Reinforced, and Which Disappear?
- A creative explosion: just as the cloud made more people build software and become content creators, AI raises that level. Today you do not need to know how to write in order to write well
- Humans and agents collaborating: it is something we are going to have to get used to
- And the repetitive, manual tasks disappear, the ones that add little value and that have little technical difficulty to automate
The Final Message
- The person who knows the most is not going to win
- San Francisco is a bubble where you run into people at the forefront, but the people doing the most interesting things are the most humble, the most curious, the ones who experiment the fastest
- Because we all start from zero, there are no written rules, and many of the ways of doing business, product, and design from the past no longer apply
- The work is to tell which of those rules still hold and which ones go
Silicon Valley Glossary
| Term | Meaning | Context in the episode |
|---|---|---|
| AI first | An organization designed around AI, not one that adds it on top | The first question: smaller teams, more fluid functions |
| Model layer and application layer | Who builds the model and who builds on top of it | The discussion about where the value gets captured |
| Commodity | An undifferentiated product that competes on price | What people discussed would happen to the models after DeepSeek |
| Horizontal and vertical | A generic product, or one specialized in an industry | Two strategies that, in his view, are going to have different winners |
| Batch | Each cohort of an accelerator | Y Combinator’s, today almost all of them AI and almost no Latin American teams |
| Generative UI | An interface generated by the model, mixed with the traditional one | The alternative to chat as the only interface |
| Frontline | Frontline workers, in a factory or at a point of sale | The people traditional software never reached |
| Selling the outcome | Charging for the work done, not for the tool | What moves the budget from software to salaries |
| Beginner’s mindset | Assuming you do not know, and asking | A Nubank culture principle he uses as a method |
| MVP | The minimum version you use to test an idea | Which now has to be more fluid, because the ground keeps changing |
People, Companies and References
People Mentioned
- Enzo Cavalie: host of the podcast and an investor, the person who gives this talk
- Elad Gil: a Silicon Valley investor, author of the line that opens the talk
- David Vélez: founder of Nubank, cited for beginner’s mindset
- Adolfo Babatz: founder of Clip, part of his first week in Mexico
- Allie K. Miller: a former AI executive at Amazon, author of the tweet about profiles
- Bill Gates: for the line about overestimating the short term and underestimating the long term
Companies and References
- Reach Capital: the fund in San Francisco where he invests
- Dalus Capital: the Mexican fund where he worked starting in 2019
- Y Combinator: the accelerator whose pivot to AI he uses as a thermometer of the gap
- Cursor, Harvey, and Decagon: the examples of products that do the work, rather than products that assist
- DeepSeek: the model that reopened the discussion about commoditization
- Nubank, Mercado Libre, Creditas, Clip, and Kavak: the Latin American references in the talk
- FTX, Clubhouse, BeReal, and Quibi: the Silicon Valley he found when he arrived in 2021
- Uber and Airbnb: the example of what gets unlocked when a new platform appears