Nowhere to H(ai)de
2022 was the year of AI. From coding to content creation, we are witnessing a Cambrian explosion of new AI-powered tools, businesses, and consumer products.
To get a sense of AI's blistering pace this year, consider this: ChatGPT reached 1 million users in just 5 days, smashing records set by juggernauts like Spotfiy, Instagram, and Facebook.
It's not all good news for the consumer AI ecosystem, though.
Paradoxically, the success of consumer AI products is already threatening many of the companies in the ecosystem. In a world where everyone has access to the same models, APIs, and ever-cheaper compute, it just became harder than ever to build an effective moat.?
The breakneck pace of innovation has led to crowded niches, with dozens of companies in every vertical jockeying for their piece of the AI pie. There are dozens of AI writing tools, image generators, and coding assistants.
Competitive markets are nothing new, so why is this notable in the context of the AI boom? Because many of these companies are built on top of the same models and APIs, which are now available to nearly everyone.
This is the catch with the majority of consumer AI products: they are all built using the same “secret sauce,” which means that it isn’t really a secret at all.
You might be wondering: what is to stop Big Tech companies from integrating these tools directly into their products?
The answer: nothing.
Established tech companies want to keep users within their ecosystems, and have both the knowledge and capital resources to add AI at every level of their tech stack.
This leaves companies building on APIs and open source models at risk of being disrupted from all sides; by competitors in their vertical, by Big Tech looming above them, and by new AI models coming to market.
Does this mean the latest crop of AI businesses are to be short lived? For some of them, absolutely. However, savvy operators who know how to ride the AI wave stand to see massive upside as AI penetrates every part of the consumer experience.
Here are my major takeaways for the consumer AI ecosystem:
Execution matters: AvatarAI vs Lensa
While consumer AI products represent a paradigm shift, business fundamentals and GTM strategies still matter.
The wildly successful indie hacker Pieter Levels (@levelsio) was very early to the AI boom, first building This House Does Not Exist (AI-rendered homes) and then InteriorAI (virtual home staging) on top of a Stable Diffusion model.?
He then ventured into custom AI avatars with AvatarAI, which allowed users to upload their photo and receive dozens of AI versions of themselves in return.
AvatarAI proved to be his most popular consumer product, so Levels doubled down and scaled the offering, which is now doing ~$100k/month in revenue.?
While Levels was building in public, Lensa was building an AI avatar feature into their existing mobile app, but with a much different strategy. They stayed under the radar until launching the feature, at which point they burst onto the scene with a few key differences: natively mobile app, low price point, huge marketing push, and their own GPUs.
Because they ran their own GPUs (rather than relying on an API), their cost to train the model on each users’ photos was a fraction of what Levels was paying:
领英推荐
Lensa’s focus on marketing and influencer outreach allowed them to get their product in front of a massive audience shortly after launch, while their low price point and natively mobile experience created a low barrier to entry for customers.
Lensa’s avatar product took off on social media, sweeping across Twitter and Instagram and reportedly producing $1M/day in revenue in early December.
Amidst cries of “AI killed X!” and sweeping claims that jobs will now be automated away, this serves as an example that AI alone does not win the day. The divergence in outcomes between these products demonstrates that how you build, launch, and scale consumer AI product is as important as the product itself.
Niches (Still) Make Riches
One of the remaining moats in the AI landscape is the ability to finetune models for specific functionality. While APIs and open source models are somewhat “general purpose,” they can be tweaked to produce outputs specific to different use cases.
Lexica Aperture is one such example: it is a text-to-image generator built on a finetuned Stable Diffusion model which produces striking, photorealistic images of people.
On a recent episode of Where It Happens, Ben Tossell (author of AI newsletter Ben’s Bites) observed that there is still a risk that these niche players get disrupted by newer models (e.g. GPT4 is due to be released next year).?
If your moat is how well you can fine tune a model, there is no reason that those tweaks could not be included in the next major upgrade to the underlying model.
Greg Isenberg (host of Where It Happens) responded that the risk of disruption by Big Tech companies can actually create an opportunity for smaller, nimble players.
Those who can quickly spin up high quality products in a specific niche become prime targets for acquisition by larger companies that would rather buy vs build. Of course, this will require more than just a great product; aggressive business development and product positioning will determine the big winners in the race to get acquired.
Building within specific niches provides a path forward for both standalone products and companies targeting acquisition. It also offers a glimpse into what the future of fine tuned, consumer AI products could look like. What if we extend the idea of fine tuning these models all the way down to the individual consumer?
Enter: Model as a Service (MaaS)
We are accustomed to curation in every part of our digital lives. Netflix recommends movies, Instagram serves us content based on our scrolling habits, and Spotify suggests playlists.
Now, imagine models that understand your individual tastes, and then generate content based on your preferences. Instead of serving you existing content, these models would be creating brand new content based on your specific inputs.
This leads us to what Tim O'Neill calls “model-as-a-service;” an AI model trained specifically for the individual consumer.
So, instead of prompting your AI writing assistant to “write this article in the style of Shakespeare,” you can train the model to sound exactly like you (or, your ideal version of you).?
This is the type of functionality I feel is lacking in current AI writing assistants. They are limited in how you can modify the tone and “train” the model on existing text. I expect to see a 10x improvement in this area within the next 6-12 months.
With personalized models, you wouldn’t have to iterate a dozen times with Stable Diffusion to find an image in the style you want; the model would produce images based on your aesthetic tastes.
You can extend this idea a step further and imagine that Netflix uses your viewing habits to generate an entirely new series for you to binge. AI may enable a paradigm shift away from a “one-to-many” mode of content distribution to a “one-to-one” mode of content creation.
Companies that find a way to make fine tuning and training models user-friendly to the average consumer will have a massive advantage in the consumer AI marketplace. It stands to reason that companies specializing in “model-as-a-service” will crop up, offering to translate your data into a series of customized AI models.
In my opinion, this will represent the "true" paradigm shift in consumer AI products. Currently, we have products that supercharge our capabilities in writing, design, coding, etc. Once we can each fine tune our own models (or pay someone to tune them for us) to generate entirely customized content, the world will start to look very different, very fast.
In the short term, we have seen a “race to the bottom” as dozens of undifferentiated companies race to spin up lookalike products based on the same underlying models and APIs. The space is incredibly noisy as everyone scrambles to integrate AI features and jump on the bandwagon.?
Despite all of this noise, there is massive upside for those who build innovative products with a long term vision. The winners will be the ones who build durable moats through smart execution, savvy business development, and always keeping an eye on the horizon.