As a Human, I’m Excited About AI. As an Investor, I’m Still Wrestling With a Few Questions
Riffing Off VC Charles Hudson’s Blog Post, Here’s What I’m Trying to Answer
If startup founders sometimes ‘Build in Public,’ is the analogou sventure capitalist motto to ‘Think in Public?’ Anyway, there’s no doubt that the story of the trailing months has been Artificial Intelligence. Over Homebrew’s first decade we’ve always been interested in what we’ve called ‘Applied AI’ (along with Applied CV, Applied ML)— opportunities where the technology itself was being extended and commercialized for a specific purpose (contrasted with core R&D or base model development). Companies such as Shield.ai, Kettle, and MasterfulAI, among many others, were Homebrew investments which fit this definition. But it’s also clear we’re at a new inflection point where our previous hypotheses needed to be updated. So like a stone in a polishing tumbler, ‘what are our principles here’ had been tossing around my head for a handful of quarters. And then I read Charles Hudson’s post, which prompted me [AI PUN] to just write this down.
In “Honest and Naive Questions from a Generalist Seed VC Grappling with the Generative AI Revolution,” Charles (whom I love) touches on similarish topics to what Satya and I have been chatting about.
I. Base Models
- Given team, data, and compute costs, will the ‘price of entry’ and ‘price of innovation’ on base models increase or decrease over time
- Will different data types produce/require their own base models, and under what conditions are these base models likely to be produced by different companies/sources vs under a single corporate umbrella
- How does one measure ‘quality’ and what characteristics will base model owners compete on besides ‘quality’ [price, latency, privacy, etc]
II. AI ‘Middleware’
- In a multi-base model world, won’t there be some value created by dynamically switching between models depending on the use case? Won’t most application owners who seek to integrate ‘AI’ be interested in “best results” more so than having to choose a model upfront
- Will this middleware layer have access to enough model…