Sarah Guo on AI investing: capabilities are still underhyped, but most companies are 'wrongly hyped'

Summary

Sarah Guo, founder of Conviction, argues that AI is simultaneously underhyped and wrongly hyped — the capabilities that matter most are still underappreciated, while the things generating the most excitement aren't necessarily the right ones.

The GPT-4o image generation moment is her evidence for how early we still are. People assumed image generation was a solved problem after diffusion models produced realistic outputs. GPT-4o's autoregressive approach — building images token by token rather than denoising from noise — delivers far more controllability and semantic understanding, and Guo expects the same compression curve that has played out across every other AI capability: expensive and slow today, cheap and fast within months.

The app layer question

For founders building at the application layer, Guo's framework is straightforward: watch what the labs actually care about, and don't build what they're already building. Anthropic and OpenAI are explicitly optimizing for AGI, which means image generation and voice were almost incidental products — but the labs still pursued them because consumer engagement justifies training costs. The safer ground is the last mile. Base models get companies to "pretty great"; there's durable value in getting to "excellent," and the labs neither can nor want to serve every vertical workflow in law, finance, healthcare, or video editing.

On the creative tools side specifically, she thinks the ceiling is still high. Controllability — giving artists, advertisers, and communicators the specificity they actually need — remains an open problem. Video is harder than images and requires different architectural changes, so workflow and collaboration tools on top of generation infrastructure still have room.

AI rollups

Guo is skeptical of rollup strategies that lead with "we're better at finance." Private equity is a mature, optimized asset class, and engineering credibility doesn't transfer cleanly into underwriting alpha. Her threshold for a rollup thesis that works has three components: genuine edge in asset selection, a target industry that would benefit substantially from AI (reflected in growth rate or EBITDA) but is too fragmented or culturally distant from Silicon Valley to adopt it organically, and a technology wedge that is fundamental rather than incremental to the business. Conviction has something in the portfolio along these lines, still in stealth.

Revenue durability

Guo draws a sharp line between what she calls "chicken nugget revenue" — trial usage with 30% monthly churn, driven by novelty — and deep, multi-year enterprise contracts with integrations. The distinction matters now because fast revenue growth has become the visible benchmark in AI, and founders are under pressure to match the curves posted by Wiz, Cursor, and Ramp. Her advice is to identify the actual bottleneck to durable revenue and treat everything else as distribution. To investors evaluating AI companies: buyer beware.

Healthcare is her example of the pace shift. For a decade she avoided it — too slow, too conservative, too hard to sell. Now companies like Abridge are posting enterprise revenue numbers that would have been implausible two years ago, because they're building things people are actually willing to buy quickly.

Investing outside domain expertise

Guo cites Harvey as her clearest example of investing into an industry she knew almost nothing about. She doesn't know the legal industry; she knows the opportunity is large enough and the technology capable enough that she's willing to spend 20 years learning. Her process is to map what every expert in the status quo believes, understand why they think something won't work, and then decide whether she still has conviction. Experts see today clearly — they just don't see the future. The best theses tend to generate mixed opinions, not unanimous skepticism or enthusiasm.

Firm building

Conviction launched in October 2022, timed to what Guo saw as an accelerating window. She left Greylock with a clear instinct for early-stage, concentrated investing, but no assumption that the right partner would be immediately available. It took a year to bring on Mike. She has no interest in scaling to a large firm or selling equity at Andreessen scale — her benchmark is hit rate and multiple, not AUM.