Sarah Guo on Conviction, AI Investing and the Future of Venture

Harry Stebbings welcomes Sarah Guo, founder of venture firm Conviction, back to 20VC. In this April 2023 conversation they discuss AI investing, venture-firm design, inequality and risk. Both are investors, and Guo's views on Conviction and portfolio companies reflect her financial and professional interests.

Guo says she left Greylock after about a decade with gratitude for its people and mentors. She wanted to return to entrepreneurship, focus on the earliest stages of company building and test a different model for a small investment team. Her conviction that AI represented a major technological change was central to the decision.

Building a fund also means running a business: payroll, offices, recruiting and serving both founders and limited partners. Guo says the additional responsibilities and uncertainty differ from being a general partner at an established firm, though she finds the work rewarding.

Her thesis is that AI could enable small teams to build very valuable companies and drive change for decades. Conviction wants to identify such companies early and support founders through a specialized community. These are the fund's ambitions and her predictions in 2023, not established returns.

Stebbings worries that AI could concentrate wealth among a small number of people. Guo agrees distribution could become more unequal, while arguing that productivity and abundance are valuable and that society should address distribution through democratic policy. She does not claim progress automatically benefits everyone equally.

On regulation, Guo says the gap between policymakers and technical practitioners is a concern, and the speed of change makes the adjustment harder. She believes people developing and enabling technology should educate and work with policymakers and the wider public. This is a policy preference, not a conclusion about any particular law.

Stebbings asks why an enabling technology warrants a dedicated fund. Guo says Conviction could eventually look like a broader software fund if AI becomes pervasive, but specialization now helps build relationships with researchers and provide access to models, data, compute and design partners.

She thinks investors can help in different ways, and distinguishes genuine interest from fear of missing out or pretension. In her view, deep understanding of this fast-moving field was still limited. Stebbings says he is transparent about his own technical limits and focuses on other company-building support.

Guo says highly publicized large AI funding rounds represent a small number of cases rather than the experience of most founders. She acknowledges both unusual traction at some companies and reduced price sensitivity among investors, while arguing that constraints often encourage discipline and creativity.

Training a very large model from scratch can require specialist researchers, thousands of GPUs and months of work. Guo believes only a small number of companies would need to take that route; most can apply existing models, fine-tune them or build elsewhere in the stack. The numbers and allocation are her estimates, not a cost model for every project.

She says founders should ask whether a model will be useful and wanted before committing to expensive training. Some technically sophisticated founders, in her experience, have changed their sequencing after examining that question. Customer demand is a separate question from the ability to build the model.

Asked whether a roughly $100 million fund is large enough, Guo says she deliberately chose a size that forces focus on early-stage investments rather than growth rounds. She describes initial checks ranging from about $1 million to $8 million or $10 million and a more concentrated portfolio than many seed funds. These are her fund plans at the time.

Very large rounds at high initial valuations may not fit the fund's economics, she says. She does not want to invest a token amount purely for a prestigious name; in her view, long-term reputation comes from returns, important companies and how founders experience the partnership. Her fund has a direct interest in this positioning.

Guo identifies legal AI company Harvey as one of her investments and an example of applying models to a text-heavy profession. She believes some work done by junior legal associates could be assisted or performed with such tools. Her assessment is an investor's view, not an independent evaluation of legal accuracy.

Other opportunities she names include code generation, workflow automation, model tooling, retrieval, feedback and multimodal creative work. These are areas Conviction is interested in backing, not a guarantee that any specific application will prove reliable or commercially successful.

In the Harvey example, Guo values the combination of AI research experience and legal-domain knowledge. She expects many future AI products to be built by product-oriented software engineers as tools become more accessible. Her view is that customer understanding and technical capability can complement one another.

She expects coding tools to move beyond autocomplete toward more iterative, end-to-end work in which a person describes a system and requests changes. She does not conclude that software engineering disappears; she expects engineers to use the tools. The scope and timing are predictions made in 2023.

Stebbings asks whether personalization will fragment products into many tiny offerings. Guo thinks general systems can accommodate many preferences through natural communication, just as a generalist engineer can build different applications. She sees powerful general technology and customization as compatible.

On startups versus incumbents, Guo says a startup's classic advantage is speed, potentially more valuable in a rapidly changing environment. Large organizations have distribution and resources, but she is skeptical that data alone guarantees a moat because founders can collect or generate data creatively.

She praises Microsoft leaders for backing AI and using the OpenAI partnership to enter other markets. She considers Amazon and Apple structurally advantaged but says, from her 2023 vantage point, they may need further investment. These assessments and judgments about competitive position are her opinions at the time.

For idea generation, Guo recommends deep, high-resolution conversations with customers rather than searching generically for a startup concept. A founder might solve a problem they know personally or investigate a hard research question tied to a valuable market, such as usable 3D generation.

Polite interest is not the same as demand. She suggests asking how important a problem is, presenting a concrete proposal and requesting a follow-up. Whether a customer will commit time or money gives more useful feedback than a flattering reaction to a broad idea.

Guo calls herself a founder-first investor, but does not ignore difficult markets. She asks whether a founder understands structural obstacles and has a plausible way to change distribution or navigate toward a better opportunity. A difficult history is a warning to investigate, not necessarily a reason to reject every company.

She names Benchling and Rippling as investments she regrets not making, citing founders whose abilities could alter the shape of a market. She says she has become more willing to take risks on exceptional founders, while recognizing that a founder's prior history can raise questions. These are her retrospective judgments, not a general return analysis.

At the seed stage, Guo says defensibility is often not yet present because the company is barely formed. She looks at the team's trajectory and ability to develop a thesis as the market evolves. Her argument is against demanding a finished moat on day one, not against eventually building durable value.

Large multi-stage firms and small specialist funds operate under different incentives. Guo says a very large fund must decide whether to deploy larger checks, hire more people or devote less time to each small investment. She recognizes advantages in scale and reach while identifying possible risks from group dynamics and internal politics.

She advises founders to learn how prospective funds work, consider who can help over the next 18 months and whom they trust for the long term, and speak to founders who have worked with those investors. She is surprised how few founders, in her experience, seek such references.

Asked whether generalist seed investing is dying, Guo says Conviction is deliberately specialized but rejects a universal conclusion. She knows successful investors with narrow expertise and others who operate across many technologies. Venture is dynamic, and she is skeptical of definitive claims that only one model can work.

Her podcast No Priors began as an idea to talk with friends working in AI, but she learned that high-quality media is a business and requires substantial effort. Stebbings adds stories about building an audience early. Both hosts have podcasts and an interest in their success.

In a quick-fire question about the end of 2023, Guo predicts continued difficulty for technology startups as companies digest capital raised in earlier years and work toward durable economics. She explicitly frames this as a narrow view of startups, not a forecast of every part of the economy.

She argues that AI's opportunity extends beyond the software budget into work currently purchased as services. Legal work is her example: automating or assisting parts of a service could address a larger spending pool than software sold to law firms. She also recognizes the associated questions about labor displacement and wealth distribution.

Asked which firms she would back, Guo praises Sequoia's execution and culture and chooses her own fund for the seed category. She predicts undifferentiated small seed funds could face a harder fundraising environment, and questions some multi-stage firms' approaches. These are competitive opinions from a fund manager.

Guo highlights near-term misuse: tools that generate code can also help generate malicious code, including for state actors and hackers. She calls for discussion and investment in defenses while saying longer-term AI safety also matters. The point is a risk assessment, not a claim about a specific incident.

She says she has become more comfortable admitting that early company outcomes and market structures are unknowable. A compelling story about whether a startup or incumbent will win is still a story; founders and other actors can change the market through their decisions.

On reserves, Guo says Conviction primarily invests in early rounds and accepts that it may leave later opportunities to others. The choice focuses the firm rather than constantly re-underwriting its own portfolio. Stebbings questions reserve strategies more broadly; their exchange reflects their fund models, not a universal rule.

She is grateful for her limited partners, whom she describes as long-term oriented and including founder friends. She also finds the broader LP landscape clubby and would welcome more independent conviction rather than simply following well-known brands. These are her fundraising experiences and opinions.

Asked in April 2023 about the 2024 US election, Guo speculates about Donald Trump's prospects and legal exposure, then says she does not know who will win. The remarks are historical political speculation, not a statement of later legal or electoral outcomes.

Guo defines success for Conviction first in terms of venture returns measured as multiples rather than absolute dollars. She also wants to back important companies, build a small partnership that makes better decisions together and earn the trust of founders. Those goals are aspirations, not reported fund performance.

She adds a broader hope that AI can be used productively and in ways aligned with people. Stebbings thanks her and closes the interview.