Software 3.0 companies build systems around foundation models to create end-user products, much as SaaS companies built products around databases.
2
The strongest AI application companies start from a specific customer problem and often create markets that did not exist before.
3
AI companies need to plan around model choice, inference cost, pricing, quality, safety, and non-deterministic outputs as part of one product system.
Summary
Sarah Guo presents Conviction's framework for evaluating AI-native companies. She describes Software 3.0 as systems that manipulate foundation models to create useful products. The opportunity is often in creating a new category rather than improving an existing one, but that requires teaching customers what the product does, finding a use case, and developing a pricing model. Guo argues that successful teams work backwards from customer needs instead of starting with an impressive model capability. Products may begin as simple GPT wrappers, then add prompting, post-processing, evaluations, routing, smaller models, memory, tools, and other system components as the problem demands. AI businesses also have unusual economics. Seat pricing may not fit products that replace services, while inference costs can change product design and margins. The talk is candid about uncertainty, especially around pricing, model costs, and where value will accumulate across the AI stack.
AI applications are moving from infrastructure toward customer products
Guo says 2023 brought foundation models, vector databases, new modalities, and tools for training, hosting, evaluation, inference, and labeling. In a survey of 70 engineering and product leaders, inference and foundation model platforms were consistently among the most useful tools. Public API usage does not necessarily mean lasting dependence. Companies are increasingly fine-tuning smaller models and running their own systems. Guo expects this tooling consolidation to make 2024 a year focused more on AI applications. She also says many enterprises were still deciding what to build internally and what to buy, although willingness to spend was increasing.
Application companies should work backwards from a customer problem
Guo contrasts capability-first thinking with customer-back thinking. Engineers may begin with a question such as whether a model can perform a particular task, but Conviction looks for a defined customer, a real problem, and a product that solves it. She says there is a scarcity of this orientation in AI. For application companies, working backwards from the customer is likely to become the dominant approach. The product may use advanced models, but the starting point is the user's need and the quality floor the product must reach.
The largest opportunities may come from creating new categories
Guo says traditional venture investing often involves entering an existing category, such as CRM, security, or log management, and improving price, performance, or features. AI creates opportunities in markets that did not previously exist. She gives foundation model APIs, legal copilots, and video avatars as examples. These markets can begin very small and grow quickly. A company that creates a category has to define the use cases, create awareness, develop a budget with customers, and decide how to price something buyers may not yet know they need.
Minimum viable quality depends on the whole product system
Guo and her partner use the idea of minimum viable quality, or MVQ, to ask what combination of product, models, and supporting systems reaches the quality floor for a function. Some valuable use cases work with a model out of the box. Others need better prompting, post-processing, and evaluations. More demanding products may need small and large models, orchestration, memory, and tool use. The system must also handle failure. For a search query about a recent Warriors game, a model can lack current information, choose the wrong search result, or return a page the user still has to parse.
AI products add quality dimensions beyond latency and accuracy
Traditional software often focuses on latency and quality, but Guo says AI products introduce additional concerns. Teams need to consider safety for what users enter into a model, the non-deterministic nature of outputs, consistency, and the range of ways people may use the product. Her argument is that the best teams are customer-driven and use technology to fit the customer's needs. Model capability alone does not define a good product. The surrounding system and its behavior under real use matter just as much.
Domain knowledge and fast product cycles shape strong teams
Guo uses Harvey, Haen, Pika Labs, and Sierra to illustrate different combinations of research experience, domain knowledge, product understanding, and execution speed. Harvey combines research experience with legal expertise. Haen's founders understood advertisers, creators, image generation, and computer vision. Pika Labs came from researchers who had struggled to make a short film themselves. Sierra combined experience with large brands and customer support systems. Guo says the products were young but already had traction and user love, which she connects to rapid iteration and the ability to learn from research and other builders.
AI pricing and margins may not follow SaaS patterns
Guo compares AI economics with the familiar SaaS model, where businesses may target high gross margins and predictable operating margins. AI companies face uncertainty on both revenue and cost. Seat pricing may not fit a product that replaces work, and charging per generation ignores the fact that some outputs are much more valuable than others. She sees a possible model in replacing a service, such as video editing, while charging less than the human service. On the cost side, teams must account for training, data acquisition, inference, partnerships, and human review. Falling model costs could change product and sampling decisions.
A GPT wrapper can become valuable through workflow and distribution
Guo compares AI applications with SaaS companies that initially operated on databases. Salesforce became valuable through its data model, business logic, standardized workflows, and distribution, rather than because it used a database. She applies the same logic to AI products. A foundation model may be an important component, but the company can build value through the system around it, including workflow design, a specific user experience, customer understanding, and distribution. She also says classical machine learning experience alone may be less useful in a period when model capabilities and system design choices change quickly.
Model routing and smaller models can improve AI business economics
In the discussion after the keynote, Guo describes a common pattern in more mature AI-native companies. They may begin with a powerful model to learn what the product can do, then route simple requests to cheaper and smaller models. An intent classifier can decide which model should handle each query. Teams may also use an off-the-shelf model, fine-tune one, or distill one to improve latency and performance. Guo says smaller models have been more effective than she expected, which changes what is possible for user experience and how widely companies should consider fine-tuning.
