My twitter feed was once a great source for everything new in AI. Until recently it’s divided into two groups — the threadbois & doomers.
The threadbois are usually declaring products like MS Excel (245M paid users, $44B / ARR) dead because someone on Twitter has successfully built the 97th clone of a UI wrapper on ChatGPT.
Whereas the Doomers sound smart by throwing in complex words but it’s always some version of “what if chatbots reach AGI and kill us” or whatever Sci-Fi concept they’re calling as “research” this week.
If you share my disappointment with these overly generic & extreme takes. I’ve tirelessly scoured the internet & compiled a few fresh perspectives.
Let’s dive in.
Fast Growth, Faster Churn.
Earlier last week, Andrew Chen raised an important point about how we’re blinded by the rapid growth of these AI apps & we’re ignoring the high churn / low retention problem.
Before we dive into why this is happening. Let me make this problem more real for you.
Lensa is a great example. The AI portrait app was doing an impressive revenue of $2M a day. However this was short-lived as revenues quickly plummeted to a low thousand figures.
Don’t get me wrong. $37M in 60 days, after app store fees is one hell of a feat. But it does highlight that the retention problem in AI still needs to be solved.
The main reasons behind the churn problem are:
- UX & Usability of AI apps sucks: Most of them are poorly stitched clunky MVPs that are front-end wrappers on APIs. Delivering a frustrating experience.
- Missing Last Mile Features: Generative features aren’t enough, as they are merely hooks. For example: Users of a pitch deck AI app now want same features as Google Slides.
- No Network Effects: Almost all AI apps are currently single player, which doesn’t have any viral coefficients. Apps must integrate collaborative & social features to retain users.
- Vitamins vs Painkillers: Vitamins are “nice to have” whereas Painkillers are “need to have” and non-negotiable. Most AI apps don’t even fall under Vitamins, as they are merely candy solutions.
- Frequency of Use: Stickyness of an app is also a direct function of how recurring the problem is. For most AI apps it’s a one time thing.
The TL:DR; If you want to build a sustainable AI app, don’t get blinded in growth metrics, focus on what fundamental value are you solving for someone & think of how often can you keep solving for them.
The catch-up game is reversed.
The fast growing AI tools on the market are facing a new challenge.
Turns out that the “Generative AI” features merely serve as hooks to attract users. However once they are onboard, they demand features & an experience comparable to the legacy tools.
Consider an AI writing or pitch deck app. Users now demand:
- The features already present in Google Slides / Docs
- The performance / reliability & snappy-ness of these features.
It’s amusing to note the reversal of this race as, not a long time ago the traditional companies were trying to catch up with the new AI players.
While the gap in the old race is reducing as most generative AI features are wrappers on top of the ChatGPT API. But the reverse catchup of building features that legacy tools spent years building & refining — might spell doom for many companies.
This explains why there has been an increase of listings of these rapidly growing AI companies on MicroAcquire.
The bottom line is — after being through many AI winters, we are now headed towards the “Plateau of Productivity” and while that’s generally a good thing for AI but the reality is many companies won’t survive the feature race. Brace yourselves for an upcoming season of acquisitions & deaths.
Infrastructure never dies
“In a gold rush, always sell shovels”
It might sound cliche, but it’s true even in the AI gold rush, the most value is currently being captured in the infrastructure layer and no, I’m not talking about NVIDIA.
Remember how earlier I mentioned that the traditional tools were catching up to their AI alternatives? It’s not like they suddenly became more productive or their operational inefficiencies disappeared.
The answer for what’s driving this transformation lies in — AI SDKs. They’re pieces of boilerplate code that you can simply plug and play into your app & quickly add AI.
Vercel was able to see this & they quickly launched their own AI SDK.
You either create vast SDKs like Vercel or go ultra niche on a single UI component like Steven Tey’s Novel.sh — It’s notion like “Help me write with AI” but open source.
The reason why this is powerful is consider these two gentlemen.
Jim Raptis who was building an image editing app & Tim Bennetto who was building a thread writing app. Both were able to quickly just add AI to their products without the heavy lifting.
MOATs wise, SDKs are very defensible as the cost of switching them is very high and unlike web3, AI is fully embraced & encouraged by big tech, which is driving the confidence & fomo amongst the medium & small tech players.
The gist is — Everyone and their mom is going to integrate AI into their apps, thanks to SDKs.
(P.S: If you’re looking for more SDK ideas, highly recommend reading this article by A16 which highlights the emerging tech architectures for LLM apps)
That’s it for today.
I’m super pumped about starting this Substack
I’ve always wanted to have my corner on the Internet where I share my thoughts & obsess about things with strangers from all over <3
Whoever is the most subscribed substack, I’m coming for you.
Just kidding lol we have 3 subscribers (Thanks mom!) , please consider subscribing & sharing with your friends.