Guillermo Rauch on Product Focus, AI Interfaces and Investing

Guillermo Rauch began programming as a child in Argentina, building websites around his interests and learning HTML through trial and error. He later contributed to the JavaScript project MooTools. At 17, an invitation to work with a Swiss company conflicted with finishing high school; he chose work and eventually moved to San Francisco.

Rauch had started earning online at 12 through Mercado Libre’s referral program. Freelance development and small payments for fixing software problems later helped his family meet expenses in Argentina. The income was useful but irregular, he says, so it brought both opportunity and stress.

His parents supported leaving a demanding educational path for software, despite the uncertainty. By then, years of earning and open-source contribution had shown the family that the internet could change their circumstances. Rauch credits their willingness to take that risk as well as his own work.

What drives him, Rauch says, is using one accomplishment as a starting point for the next. Moving to the US expanded his access to capital, networks and people with similar interests. He still sees advantages in San Francisco but says it is not the only place where people can build companies.

Rauch argues that reliable infrastructure and efficient institutions can change what people are able to accomplish. He believes immigrants may especially appreciate opportunities in the place they move to, and expresses confidence in the US system’s checks and balances compared with his experience in Argentina. These are his personal observations and hypothesis, not a demonstrated rule about all immigrants.

When assessing talent, Rauch looks for tangible work: a product, a link, writing or another demonstration of thought and execution. He wants candidates to connect their technical contribution to a better user or developer experience. This fits the mission of Vercel, the company he founded and leads, which sells tools and infrastructure for building web products.

Stebbings warns that polished storytellers can be much better at selling than doing. Rauch agrees and recommends drilling into implementation choices, alternatives and trade-offs to test conceptual depth. He also acknowledges that visible products are not a perfect test for every role, especially more theoretical work.

A hiring mistake Rauch has made is giving too much weight to prestigious employers on a résumé. Brand names are approximations of merit, not evidence of a person’s contribution. Both speakers say founders can mistake familiar credentials for direct validation of a candidate’s skills.

Stebbings favours repeat founders because they have already learned from mistakes. Rauch agrees that experience is a useful filter, while noting it narrows the pool. He also values early startup employees who observed decisions closely and can explain what they would do differently when building their own company.

Asked whether simple is always better, Rauch says there is no universal answer. He admires deceptively simple experiences, such as a search box or an API that hides extensive infrastructure. He mentions Auth0, an early investment, and Stripe, which he says he admired but did not invest in, as examples of useful simplification.

As a company grows, what made its first product successful may not be enough for the next stage. Rauch, a longtime open-source contributor, recalls a discussion about open source helping market awareness without by itself solving every scaling problem. He argues that funding stages are imperfect labels, but sequencing and adaptation matter.

A multiproduct company may look much more complex than its original wedge. New entrants can simplify a market, but Rauch says leaders must repeatedly reconsider what their customers and stage require. Stebbings warns that adding enterprise features can create feature creep and make a product harder to use.

Rauch’s answer is to pick battles rather than solve every customer problem or enter every market. Sometimes waiting for experiments to reveal a better approach is wiser than reacting immediately. He cites Apple as an example of choosing its moments and says being right can matter more than being first.

At Vercel, he says, teams set a high quality bar and test internally before release. To reconcile speed with reliability, he recommends limiting the blast radius of experiments and gathering evidence before a broader rollout. The description is his account of Vercel’s process, not an independent assessment of its products.

Stebbings asks whether rigorous validation makes shipping too slow. Rauch believes teams can institutionalise fast, careful work by sharing methods and examples rather than requiring the CEO to inspect every project. He uses a factory stress test of an aircraft wing as an analogy for building confidence beyond ordinary operating conditions.

Stebbings suggests AI may make user interfaces less important. Rauch agrees that interfaces will change, but expects information and choices still to need presentation. A natural-language request might replace many dashboard clicks, while new tools allow people to create different kinds of interfaces; Vercel has a commercial interest in those tools.

Rauch expects the tasks people need to accomplish—communicating, deciding and visualising data—to remain, even as methods change. He thinks product designers now have language models as an additional tool and can revisit assumptions embedded in older software. His proposed future designs are hypotheses, not established outcomes.

Stebbings argues that AI may eventually sell completed work instead of tools: a user could specify an advertising outcome and budget while an agent handles the decisions. Rauch says that may happen over a long period, but he thinks near-term products must meet users where they are. He compares the transition to mobile software combining older web and desktop ideas with new capabilities.

Rauch suggests some AI interfaces may need more visible choices, not fewer. Midjourney, for example, returns several images because it cannot know precisely which result the user wants. Unlike a deterministic form, a generative system may produce different or unsuitable outputs, so the interface must help users evaluate and guide it.

Stebbings worries that too many options overwhelm users. Rauch says the answer depends on the stage of the task: in creative exploration, alternatives and rejection can help people discover what they want. He sees good product work as repeated selection and says an assistant should learn from feedback and improve its suggestions over time.

Rauch distinguishes suggestive tools, such as autocomplete or a copilot, from autonomous agents acting without supervision. At the time of the interview he saw fewer successful examples of the latter. He predicts agents will advance but favours a human editor who provides direction and rejects unsuitable work; that is his proposed model, not a guarantee of how agents will develop.

Stebbings worries that an assistant model advantages incumbents with distribution and existing products, making startups harder to fund. Rauch argues that an entrant can instead redesign the workflow so fundamentally that it needs less of the old interface. His hypothetical AI word processor is an illustration, not a claim that such a product already exists.

Stebbings counters that companies such as Notion and Adobe have integrated AI rapidly. Rauch agrees incumbents should not be underestimated and praises Microsoft’s earlier strategic attention to AI. He nevertheless believes a platform shift can move user attention to a different approach, even when an incumbent improves its existing product.

On business models, Rauch describes Vercel as combining seat-based software and consumption-based infrastructure. He predicts more AI products will use hybrids: a baseline workflow plus charges reflecting additional compute usage. This is a commercial thesis from the CEO of a company using such a model, not a universal pricing outcome.

Asked whether open or closed AI systems will win, Rauch says he needs more data. He points to ecosystem standards such as Kubernetes, React and Linux as a possible advantage for open models, while also arguing that an early technological lead can be hard to overcome. His comparison of Llama and GPT-4 reflects his assessment at the 2023 recording.

In a quick-fire round, Rauch rejects the prediction that 99% of cash invested in AI companies will go to zero. He speculates that around 20% of investments might produce a productive outcome, including acquisitions; he explicitly does not mean that 20% will become unicorns. This is a personal forecast, not measured portfolio performance.

Rauch says some angel investments address problems he wanted solved but lacked time to tackle himself. He invested in Scale AI and calls choosing that company partly luck. He believes founders can be well positioned through the willingness to persist and the ability to learn quickly, not only through existing domain credentials.

Market demand matters, Rauch argues, but investors can misunderstand the best solution even when they recognise the problem. He recalls scepticism about outsourcing authentication to Auth0, an investment of his, while also recognising how difficult authentication was to implement well.

He similarly knew teams needed better communication but once set up IRC servers as the solution. Slack later offered a different answer, and he recalls hesitating over entrusting company communication to another provider. The experience taught him to remain open to a solution he might not have built himself.

Looking ten years ahead, Rauch hopes Vercel can help developers create better products and reduce frustrating software experiences. He expects AI to expand the number of people able to participate in creation. These are the ambitions and predictions of Vercel’s CEO, who has a direct interest in the company’s success.