On the August 6, 2024, episode of Invest Like the Best with Patrick O’Shaughnessy, the host welcomed Sarah Guo, founder and CEO of the venture capital firm Conviction. Guo founded the company in 2022 as an early-stage VC firm that backs startups focused on AI. In the conversation between O’Shaughnessy and Guo, she shares insights about starting a VC firm after several years at a larger, more established firm (Greylock); having a front-row view on AI product design; the biggest mistakes made by startups developing AI applications; and her views on challenges and opportunities in the AI space.
Here’s a summary of the conversation:
- Guo has found that, in recruiting talent to her firm, evaluating early-stage venture capitalists is difficult due to lack of track records. The Important traits she seeks are core technology understanding, judgment, and business insight, with judgment being the most difficult to assess among early-career talent.
- In her venture career,, her earlier work with the companies Awake (network analysis security) and Figma (a design tool) were instructive. These experiences taught her about company-building and gave her confidence in identifying long-term potential in new markets and teams.
- Her decision to start Conviction was influenced by her passion for AI and its transformative potential. Founding the company was a process; the key steps included leaving Greylock responsibly, building relationships with LPs, and crafting a focused strategy. The first investment was Harvey, an AI-driven legal application company.
- Whether in the AI space or elsewhere, Guo thinks the core framework for evaluating software companies is the same: distribution, quality of the people, and size of opportunity matter. Contrary to the narrative that there’s no value in startups working at the application layer of the stack (as opposed to those interested in developing foundation models), Guo has long thought that this “last mile” technology was important. She doesn’t advise building tech that the foundation models are going to replicate; that’s wasted effort.
- Building useful AI applications requires leveraging domain-specific knowledge (e.g., law, in the case of Harvey, her first investment at Conviction) and avoiding commoditization.
- She’s betting that large incumbents won’t be able to win every market, particularly those markets that are secondary to their primary revenue drivers. Entrepreneurs should therefore focus more on “avoiding the path of incumbent strengths.” The underlying technology might be general, but it’s the applications built on top that matter, because they bridge the gap from the underlying model to the end user. This is where there’s space for entrepreneurs to excel.
- She prefers teams with deep research expertise. She emphasizes minimum viable quality (MVQ) for AI applications to meet a baseline of practical and user-expectation standards. Entrepreneurs need to validate their AI’s quality by engaging directly with customers, and understanding whether the AI’s output is “good enough” is essential for achieving product-market fit. MVQ is a moving target, as new use cases emerge, and quality expectations evolve. It also differs across different domains.
- Some AI applications have already crossed the “uncanny valley” and surpass human capabilities in very specific contexts; writing, however, isn’t there yet. The cost of managing errors and improving AI outputs from 80% quality to acceptable levels must be minimized for broader adoption. Verification and ranking systems that evaluate multiple outputs are critical for improving end-user experiences and building trust.
- Building applications that are quick and easy to replicate is not a sustainable strategy unless there is a defensible distribution mechanism. Initial traction with simple tools may be a good starting point but cannot be relied upon for long-term enterprise value. Many entrepreneurs approach AI applications with this kind of short-term thinking, failing to plan for competition and scalability. Without deeper engagement, these startup teams risk being overtaken.
- For Guo, challenges in the AI space right now include improving multistep reasoning and addressing hallucination in models. AI’s adoption in conservative sectors (e.g., legal, healthcare, government) is progressing slowly, but holds potential for significant impact. Guo is hopeful that the next generation of models will be more able to tell when their output is correct—if the model knows whether it has a good answer or not, it’s much more useful.
- High costs in training and deploying large models remain a challenge as well; efficiency in model development and deployment is critical for future scalability.
- Large-scale AI training workloads expose flaws in alternative chips that aren’t apparent during smaller-scale testing. Companies like Google succeed in chip development because they can test at scale within their own workloads. Entrepreneurs don’t have this luxury. Testing alternative chips requires significant investment and deployment at scale, making it hard for new entrants to compete.
- Nevertheless, Guo foresees a more competitive AI ecosystem in the future, with multiple foundational model providers and hardware providers (e.g., alternatives to NVIDIA GPUs). She cites development approaches like that of Mistral. Its open-source and efficiency-driven approach is gaining traction – efficiency in model training and inference is increasingly important due to constraints like data center size, power limits, and costs.
- Domains like material science and enterprise software configuration hold untapped potential. Bridging gaps between technical innovation and domain expertise is critical.
- In terms of ethical and safety concerns, near-term risks include fraud, misinformation, and misuse of AI technologies. Long-term risks involve runaway systems and biosecurity concerns. AI adoption could lead to significant productivity gains, but also societal challenges like job displacement and regulatory issues. Guo, however, is ultimately an optimist about the transformative potential of the technology.