Skating towards the puck and Generative AI startups
This article was originally posted to LinkedIn
On Wednesday evening at a xmas party in Sydney I met Yash Varma and we had a fascinating conversation about early stage investing in AI startups, in particular the new wave of generative AI startups built on top of Stable Diffusion and GPT-3.
Early stage investing requires an element of predicting (or betting on a prediction) of the future. Of placing bets on founders and companies that are “skating to where the puck will be”. But when AI, and in particular generative AI, is moving at such a blinding pace, everyone is left furiously chasing the puck down the ice. Making start-up investment decisions is harder than ever.
For example. Dreambooth for Stable Diffusion was released in late September 2022. After setting it up on a server it allows for uploading of your own images (e.g. photos of yourself) and generating new images that incorporate these in new ways (e.g. fake photos of yourself in various styles). Almost immediately developers were experimenting and posting on Twitter pics of themselves as Viking warriors, cyberpunks and the like.
AI generated self-portrats using Stable Diffusion Dreambooth
Only a week or two later, a platform called Strmr (now Astria) launched, which allowed Dreambooth training (for a fee) via an easy Web interface, and also via an API. A week or so later, in late October, indiehackers started launching “AI Avatar” makers - paid services that allowed non-technical people to upload photos and generate hundreds of AI generated images of themselves in a range of scenes and styles.
Two of these services quickly went viral on Twitter; Pieter Levels with Avatar AI and Danny Postma with ProfilePicture AI. Being indiehackers of the “build in public” mindset, both have been quite transparent with the popularity, and profitability of their sites. Avatar AI made $10,000 in the first day, and is now making more than $100,000 per month.
Danny and Pieter hitting product market fit resulted in many copycats, but the one this week getting all the attention is the mobile app Lensa, which hit #1 in the App Store and rumoured to be generating over $1m a day!
Downloads and revenue figures for Lensa AI
A venture capital backed startup, Lensa is growing quickly by offering the same service for a much lower price, and throwing stacks of money at digital marketing. Importantly, although it meant they were slower to launch (a month after Avatar AI and ProfilePicture AI), they have taken the time to build their own infrastructure rather than rely on the now commoditised Astra API.
The next wave of avatar makers might bump Lensa from the mantle, with Danny Postma predicting “Snapchat & Instagram will unleash DreamBooth avatars as their 2.0 filters” and going on to say “Joining the avatar race to the bottom is a waste of time for me”.
Twitter post from Danny Postma
This example demonstrates not only how quickly generative AI is developing, but also the difficulty in placing bets in this environment. As a generative AI startup, is it possible to dig a moat in a month?
We have seen the same, on steroids, happen in the past 10 days since the launch of ChatGPT by OpenAI. Hitting over a million users in under a week, ChatGPT has already inspired and enabled thousands of indiehacked projects. There are already more than twenty ChatGPT Chrome Extensions available!
But history repeats itself. If everyone has the same ideas and the same access to the underlying technology (as they did with Astria for image generation), how can they have a genuine moat?
Yash and I talked about how a potential differentiator for AI products is proprietary data, and the ability to uniquely train an AI model with that data. An alternative to Software-as-a-Service, now startups can offer and monetise Models-as-a-Service (MaaS). This will be an interesting space to watch.
The AI puck is quickly accelerating into the distance. Maybe we can’t skate to where the puck will be, but with luck we can keep it in sight.