Discover more from Machine Yearning
The Future is Open! Notes from #WoodstockAI | Machine Yearning 004
Why first movers WON'T win it all
Last Friday, my colleagues and I went to the 4/1 Hugging Face meetup (#WoodstockAI) and met dozens of passionate entrepreneurs and devs. We got a ton of questions around whether first-movers would end up winning the whole market. For the GenAI application layer, at least, there’s more to it than just being first! Here are some reasons why.
The Future is Open
The halls were buzzing with enthusiasm. With over 130 open-source demos scattered throughout the Exploratorium (already packed with interesting exhibits), we hardly knew where to start - we saw next-gen literary teaching tools, a dev who converted hand-drawn sketches into richly textured digital assets, literal four-legged LLaMas, and much more.
In many ways, it reminded me of the early days of the App Store once Apple opened it up to third-party developers. At launch, there were only ~500 apps, but one year later there were 35,000, with over 1 billion downloads. Just like back then, we’re in the early innings of this game, and there’s plenty left to be discovered on UX, business models, retention, and so on. Some of the lessons of that era likely apply today; more on this in a little bit.
The Gold Rush
The folks I met aren’t just sitting idly by, waiting for the future to come to them. They’re actively building it! Everyone is an API key away from integrating into their applications state-of-the-art models from OpenAI, Stability, AI21, and other research labs. Daring engineers can pip install into their local file system whatever generative model they need from Hugging Face. This has unlocked a tremendous amount of creativity and potential.
… And yet, the generous availability of cutting-edge generative AI models has a downside.
Enjoying this? Subscribe for more hot takes.
The double-edge of this accessibility is competition. No doubt about it, the competition is fierce. It’s going to be very challenging for AI entrepreneurs to build differentiated solutions, especially in the application layer.1
We saw a similar dynamic in 2008. When the App Store officially opened up to third party developers, entrepreneurs scrambled to get onto the platform. They, too, had to navigate an unprecedented level of consumer accessibility and competition. It was a similar moment of unfettered creativity and opportunity. For anyone who doesn’t remember the top apps in the first year of the App Store, here’s a refresher on the top downloaded apps in 2008:
Source: The Verge, 2018.
It was a weird time in technology. Lots of novelty, but productivity not so much, not at first. Many paid apps that hit the big time were fun expressions of the iPhone’s hardware capabilities. Koi Pond was an elegantly simple pond simulator which took advantage of the responsive touch screen, and iBeer was considered the first killer demo of the accelerometer. It even minted its creators up to $20,000 / day at the top. That didn’t last, but it’s worth pointing out!
Novel usually doesn’t mean sustainable whether you’re talking about GenAI or iPhone apps. Apps like Recorder were quickly washed away by iOS-native apps that targeted the same functionality (Apple released Voice Memos in iOS 3). There’s even a term for this: “Sherlocking,” when a platform releases native functionality that renders a third-party development obsolete. The origin for this was a 2006 conversation between Steve Jobs and developer Dan Wood, whose company Karelia built a tool that extended the local search functionality of Apple’s Spotlight predecessor, Sherlock, to conducting online searches as well:
It seemed that the sky was the limit, until I was called in for a meeting with Apple's Phil Schiller. I listened to him tell me that Apple was going to announce Sherlock 3, and it was very similar to Watson. I watched a demo of their program: all but one of their modules connected to the same service that Watson did and looked almost the same…
Steve Jobs made it clear to Dan that any resemblance was very much his problem, not Apple’s.
"‘Here's how I see it,’ Jobs said… ’You know those handcars, the little machines that people stand on and pump to move along on the train tracks? That's Karelia. Apple is the steam train that owns the tracks.’"
Origins of Sherlocking: “The Long Story Behind Karelia's New Logo,“ Jan 7, 2006
Make no mistake, developers of foundation models own their GenAI tracks. In fact, a similar conversation happened between Jasper CEO Dave Rogenmoser and Sam Altman once ChatGPT was released. Jasper had raised a $125 million war chest at a $1.5 billion valuation, leveraging a head start conferred by its private beta with OpenAI. But almost overnight, that advantage came to a screeching halt.
Jasper’s unique value proposition was perilously diluted by the fact that ChatGPT was fast, powerful, and above all, free. Jasper has plenty of cash to burn as it tries to deepen enterprise relationships and improve stickiness, but developers of less well capitalized text generators (and other first movers) risk getting “Sherlocked” nonetheless.
Features vs. Products
This means it’s a good time to ask yourself: Are you building a feature or a product? These are two very different things. In simple terms, a product is something that customers consider valuable enough to pay for. Products have to solve problems. Features are attributes, functions, helpers that come with a product. They’re useful, but not enough to pay for on their own.
Using the Sherlock example, search is a product. Web search is a feature of search
Microsoft Nuance is a speech recognition product for medical scribes. Spanish speech recognition is a feature of Nuance
This isn’t a constant, since products risk becoming features of larger products overtime. Watson and Sherlock are just one example where the verticalized product (Watson for web search) became a feature of a larger product (Sherlock for web and local search). ChatGPT may do the same to many copywriting apps.
You may be thinking to yourself “well, there’s no way a developer of foundation models has enough resources to go after every vertical.” True, which is why OpenAI effectively launched its own app store with Plugins. Want to make an “LLM for dinner reservations” product? There’s a Plugin for that. What about an “LLM for booking flights”? There’s a Plugin for that.
Now, if you tell me “Travel is about ditching the mundane and discovering the extraordinary. But travel planning? Incredibly stressful. We believe you should be able to plan your entire trip with the push of a button. So we’re building everyone’s private travel agent. Interested?” Sure there are Plugins that can handle pieces of what I may be thinking of, but you’ve got me hooked. Tell me more.
Throughout the Hugging Face event, several entrepreneurs asked me to demo their apps. The tech was consistently impressive, but my first question was usually: “Who is this for?” More often than not, the answer was akin to “It’s for everyone, the possibilities are endless!”
I agree, the possibilities are endless! But if you intend to build a sustainable business, I encourage you to focus on building solutions with a specific customer in mind, a customer who will pay for it in time or money. If you can articulate the following:
The pain they have
The willingness they have to pay to eliminate that pain
How many people out there share that pain (your total addressable market)
then we’re really onto something. In doing so, we can create products that are genuinely valuable and rewarding for both the customer and the entrepreneur. If you need a place to start, I always recommend “The Mom Test,” one of the best resources on practical questions to ask as you’re finding product-market fit. Cut through the BS and figure out if customers will pony up!
Most importantly, have fun!
Most of these notes are aimed at founders intending to build GenAI businesses, but there’s just as much value in diving in and seeing what you can create. Don’t just build a generative AI app because you think it’ll grow into the next unicorn. Build it because you want to, build it because it’s fun, build it because it solves a specific problem and you’re passionate about solving it!
If enough people share that problem and are willing to pay for that product… then who knows! But building and learning are inherently valuable.
There are going to be so many new business models, experiences, and possibilities we discover in the application layer that it isn’t clear first-mover advantages exist. If you’re building an infrastructure, platform, marketplace, or any other product that depends on network effects for its fundamental value, then yes, it’s best to be first. But not every business fits that description. Above all, build something that people love.
The potential of AI is vast, and many exciting opportunities await entrepreneurs who create world-changing solutions. #WoodstockAI was an amazing celebration of all this potential energy. With so much talent and enthusiasm in the AI community, I'm confident that we'll continue to see groundbreaking advancements in the field.
Thanks for reading!
Liked this post? Comment and subscribe for more!
Machine Yearning is a collection of essays and news on the intersection between AI, investing, product, and economics, light on technicals but heavy on relevance.
Ryan Cunningham is a Senior Builder at Andrew Ng’s AI Fund, a venture studio in Palo Alto. As part of the venture studio, he’s co-founded AI startups in the fields of fintech, human capital, swarm intelligence, speech recognition, generative AI, and more. He specializes in deep tech applications (foundation models, drones, autonomous vehicles) and is active in the Stanford AI Alignment community at SERI, where he contributes to research and advocacy efforts focused on AI safety.