The Collisons are great writers

Stripe just published their annual letter.

The main point I wanted to discuss is their defense of LLM wrappers. But I couldn't help but admire the writing abilities of co-founders Patrick and John Collison:

For too long the crypto economy was an isolated atoll, with vibrant native customs but few exports to the rest of the world.

Why care about stablecoins? Improvements to the basic usability of money make economies more prosperous. Consider the transitions from coins to banknotes, from the gold standard to fiat currency, and from paper instruments to electronic payments. Stablecoins are a new branch of the money tree.

GDPR alone is estimated to have reduced profits for small tech firms in Europe by up to 12%. Those cookie banners hurt, whether you accept them or not.

They can even make credit card fraud sound interesting.

Fraudulent actors today operate on an industrial scale, with teams of engineers, managers, and data analysts. (We are yet to verify whether they have HR departments. If you know, please tell us so we can give them some peer feedback.) Fraudulent actors generally target times when fraud teams are offline—we see more fraud on Saturdays, Sundays, and Mondays—but we see subtler patterns, too, like the fraudsters' own work schedules. Fraudsters are particular about their lunch breaks.

Clearly the Collisons write without LLMs. They raw dog the page like Mark Twain. I would buy an ETF of well-written annual reports by CEOs. The Collisons, Bezos, and Buffet all have this in common.

Anyhow, the main point is addressing the Collisons' defense of so-called "LLM wrappers." These are applications whose main capability stems from the LLM that powers it.

Jasper was an early example of an LLM wrapper. Jasper's whole appeal stemmed from its use of language models, and not from any proprietary technology, underlying dataset, or breakthrough capability. Jasper is probably now in trouble and I would bet it gets sold for parts in the near future.

Here's what the Collisons say:

Much as SaaS started horizontal and then went vertical (first Salesforce and then Toast), we're seeing a similar dynamic playing out in AI: we started with ChatGPT, but are now seeing a proliferation of industry specific tools. Some people have called these startups "LLM wrappers"; those people are missing the point. The O ring model in economics shows that in a process with interdependent tasks, the overall output or productivity is limited by the least effective component, not just in terms of cost but in the success of the entire system. In a similar vein, we see these new industry specific AI tools as ensuring that individual industries can properly realize the economic impact of LLMs, and that the contextual, data, and workflow integration will prove enduringly valuable.

AI applications may verticalize as SaaS apps did. There's reason to think, however, that the pattern won't apply again. (These are rough thoughts so feel free to weigh in on Github.)

Software-as-a-service (SaaS) applications were made possible by cloud computing:

• Before, people had to buy software CDs, install software on their computers, and connect to on-premise databases.
• After, people could access software over the internet, and connect to cloud databases.

This change made it a lot easier for people to develop, deploy, and distribute software, leading to a huge increase in the supply of software, including lots of specialty (i.e. vertical) apps.

Despite this, there were two significant constraints on SaaS products that left vertical opportunities available:

• Development cost: It cost a lot of money and time to build out secure, reliable, and usable software. Building just the Salesforce app has cost them hundreds of millions of dollars if not billions over many years. They had to focus on their core product to be successful.
• User expectations: Salesforce's app was used by a lot of people every day. There had to be basic consistency to the product experience over time, otherwise users would be confused and switch to something else.

Both of these constraints go away with AI applications:

• Development costs are going to zero with tools like Cursor.
• User expectations won't matter because agents will be doing the work on people's behalf.

Taken together, this means that LLM-powered agents from companies like OpenAI, Anthropic, Google, and more may not leave as many vertical opportunities as SaaS horizontal winners did.