Interesting article - agree that AI isn't really a moat rather than something that can be expanded on once the core product brings in data. Confused on why many founders don't recognize this - why do you think so?
Also, do you really think data/AI can be a moat for these big companies? Data is becoming increasingly commoditized, and data itself has marginal diminishing returns. After a certain point, the data collected is very similar to prior data collected and doesn't enhance the predicative capabilities of the model. Even if you take a company like Cloudfare/Crowdstrike, as competitors chase them in these fields, they gain access to similar data that supposedly gave those providers a moat. I found this article from a16z to be interesting on this topic: https://a16z.com/2019/05/09/data-network-effects-moats/
On the founders, 1) I think the potential for AI to unlock new use cases is very compelling, but very hard to nail the initial pain point that allows for AI to bring about meaningful change quickly, 2) easy to experiment with something like take in customer support data and run through AI to see what output is delivered, and 3) also lots of capital there to fund founders!
Re data having diminishing returns that's true but the last 1-5% of data is always hard. Why Tesla still spends so much on training since those edge cases are the challenge. For Crowdstrike/Cloudflare, security attacks are constantly changing so the data on those attacks across a larger surface area keeps them in front of others. Perhaps if you have a static field, like maybe soft drink consumer data then it's not a moat but for most areas data having diminishing returns happens pretty far out on the curve.
This resonates a lot actually, thanks for sharing!
A lot of data startups I talk to like to highlight their AI/ML prowess even before serving a dozen customers successfully. In my view, there's a growing tendency amongst founders to make assumptions that are often based on second-hand knowledge (customer discovery etc) and turn those assumptions into algorithms. In fact, technical prowess alone can never help build a real moat I believe, let alone in the early days.
Also, I really don't see how startups relying on publicly available AI (GPT-3, DALL-E) can ever have a moat.
Agreed! Technical prowess alone is not enough to create a moat unless it’s something so hard to do that that team is the only one who can do it.
Startups relying on publicly available AI could create a moat if they can figure out a novel application and then it becomes an execution raise as to how you deliver those insights, get ingrained in customers and expand out from there. But the headstart you get in that case is very short!
largely true but there are some exceptions. In some cases, AI is the product itself(DALL-E?). Also scenarios where the customer onboarding involves ingesting large volume of customer data (CDP with predictive modeling & clustering)
So DALL-E is a good example of a "narrowed" use of AI. Yes it's very broad and took a lot of money and resources to spin up (so another example of it being a hard for a new startup to make AI a core product with limited resources) but also DALL-E creates photos, it doesn't create videos for example. And if you asked DALL-E to give you a recipe for pasta it wouldn't work either.
I witnessed first hand the challenges of acquiring data for AI products in the healthcare sector. Not only is getting the initial data hard (and expensive), but you will also face deployment challenges, at least while your model is still nascent, meaning hasn't generalized on the entire (or a good subset) of the population it encounters. This, in turn reduces the margins of AI companies; they become a blend of software + services. Shameless plug on this here -> https://karimfanous.substack.com/p/ai-software-services
Yes!! So true. And great article, definitely agree that many AI companies at scale will have large professional services arms in order to manage the data complexity of the customers' data siloes.
Interesting article - agree that AI isn't really a moat rather than something that can be expanded on once the core product brings in data. Confused on why many founders don't recognize this - why do you think so?
Also, do you really think data/AI can be a moat for these big companies? Data is becoming increasingly commoditized, and data itself has marginal diminishing returns. After a certain point, the data collected is very similar to prior data collected and doesn't enhance the predicative capabilities of the model. Even if you take a company like Cloudfare/Crowdstrike, as competitors chase them in these fields, they gain access to similar data that supposedly gave those providers a moat. I found this article from a16z to be interesting on this topic: https://a16z.com/2019/05/09/data-network-effects-moats/
On the founders, 1) I think the potential for AI to unlock new use cases is very compelling, but very hard to nail the initial pain point that allows for AI to bring about meaningful change quickly, 2) easy to experiment with something like take in customer support data and run through AI to see what output is delivered, and 3) also lots of capital there to fund founders!
Re data having diminishing returns that's true but the last 1-5% of data is always hard. Why Tesla still spends so much on training since those edge cases are the challenge. For Crowdstrike/Cloudflare, security attacks are constantly changing so the data on those attacks across a larger surface area keeps them in front of others. Perhaps if you have a static field, like maybe soft drink consumer data then it's not a moat but for most areas data having diminishing returns happens pretty far out on the curve.
Hey! Just to let you know I linked this post here:
https://www.libertyrpf.com/i/72573854/why-ai-is-not-a-moat
Cheers! 💚 🥃
Made it into the Liberty Labs section 🤩. Thanks for adding it Liberty!
🤜 🤛
This resonates a lot actually, thanks for sharing!
A lot of data startups I talk to like to highlight their AI/ML prowess even before serving a dozen customers successfully. In my view, there's a growing tendency amongst founders to make assumptions that are often based on second-hand knowledge (customer discovery etc) and turn those assumptions into algorithms. In fact, technical prowess alone can never help build a real moat I believe, let alone in the early days.
Also, I really don't see how startups relying on publicly available AI (GPT-3, DALL-E) can ever have a moat.
Agreed! Technical prowess alone is not enough to create a moat unless it’s something so hard to do that that team is the only one who can do it.
Startups relying on publicly available AI could create a moat if they can figure out a novel application and then it becomes an execution raise as to how you deliver those insights, get ingrained in customers and expand out from there. But the headstart you get in that case is very short!
Well I'm actually glad to be an observer and learn from the mistakes of early movers. And folks like you make it easy too so thank you!
So well said! Thank you for writing this piece.
Thanks for reading!
largely true but there are some exceptions. In some cases, AI is the product itself(DALL-E?). Also scenarios where the customer onboarding involves ingesting large volume of customer data (CDP with predictive modeling & clustering)
So DALL-E is a good example of a "narrowed" use of AI. Yes it's very broad and took a lot of money and resources to spin up (so another example of it being a hard for a new startup to make AI a core product with limited resources) but also DALL-E creates photos, it doesn't create videos for example. And if you asked DALL-E to give you a recipe for pasta it wouldn't work either.
Testify!
Haha just needed to get this one off my chest :)
100% spot on. It's the data that's the moat.
I witnessed first hand the challenges of acquiring data for AI products in the healthcare sector. Not only is getting the initial data hard (and expensive), but you will also face deployment challenges, at least while your model is still nascent, meaning hasn't generalized on the entire (or a good subset) of the population it encounters. This, in turn reduces the margins of AI companies; they become a blend of software + services. Shameless plug on this here -> https://karimfanous.substack.com/p/ai-software-services
Yes!! So true. And great article, definitely agree that many AI companies at scale will have large professional services arms in order to manage the data complexity of the customers' data siloes.