A warm welcome

Hello and welcome to the inaugural edition of my newsletter. I'm glad you're here.

Let’s establish a few ground rules before moving on: 

1. Much like the startup to which this newsletter is attached, you should expect this column to change over time. This is tech, folks – the only guarantee is that if you don't constantly make your old self obsolete, someone else will. 

2. A core belief of mine is that the best way to find out what is right is to put something wrong in front of others. That doesn’t mean I’m not attempting to be insightful and correct with what I write here—quite the contrary—but what it does mean is that I'm eager to hear your thoughts and disagreements with anything that I write. In short, please think of each post as a long-winded debate topic instead of a lecture.

3. Thoughts and opinions expressed in here are mine alone.

AI for Programming

Let’s kick off Automato with an automation topic that we’ll likely return to regularly: programming. 

In my completely anecdotal but almost certainly correct opinion, as of today, the leading industrial use case for large language models (AI) is in software development (by a wide margin). The Stack Overflow 2024 annual software developer survey supports this claim: 76% of survey respondents said that they’re either using or planning to use AI as part of their software development process in 2025.   

AI’s early success in the software engineering industry makes sense for a few reasons. 

First of all, AI as we know it today is being built by teams of software developers. As one might expect, software developers understand software quite well, especially the elite devs and researchers that get jobs at companies like OpenAI. Given that software is so front-of-mind for these companies and their employees (along with the ecosystem of startups that are building new products on top of their AI models), going after coders is as natural a first step towards capturing real economic value as any other. To think that a group of AI software startups would first try to disrupt an industry other than software development is like asking a typical high school student what AI is good for and expecting them to say something other than “Homework.” 

Second, the software development industry is, quite understandably, very tech forward! Many individuals are drawn to the industry because they genuinely love programming, which manifests itself in contributions to open-source (read: unpaid) projects, time spent learning new programming languages and paradigms outside of working hours, and in the hundreds of thousands of views that code influencers like The Primagen and Theo regularly receive. Given their general openness to experimentation, it’s not surprising that developers have been eager to get their hands dirty with AI tools.

Finally, the modern software ecosystem is steeped in the “free and open-source" tradition, which, if you’re not familiar, is the practice of exposing your project’s code to the world (using online code repository hosts like GitHub) so that anyone can use it or request to modify it. This open and collaborative practice has inadvertently created a massive online database of code that AI companies like Anthropic can use to train their models. As a result, years of developers’ code contributions to out-in-the-open projects have been snatched up and are now being used to threaten the livelihood of these same developers. Anyways, this newsletter isn’t about moralizing, so let’s get back to business. Sucks to suck!

Cursor

Due to the reasons mentioned above, a relatively new class of AI-centric programming tools is already seeing high rates of adoption. At present, the most beloved of such tools is Cursor, the “AI Code Editor.”

For those who don’t know, many programmers write code in what’s called an “IDE” (integrated development environment) which is like Microsoft Word or Google Docs but with bells and whistles for programming instead of plain old writing. 

Cursor is a new flavor of IDE (it’s actually a modified version of an open-source Microsoft product) that makes AI a front-and-center part of the development experience. People enjoy using Cursor for features like advanced autocomplete, file editing and fairly frictionless Q & A: 

Because the AI assistant is embedded right in the IDE and can therefore edit a program directly, Cursor is seeing a somewhat surprising level of adoption not just from software engineers, but also from people who don’t know how to code at all, leading to a new English-first coding style called “Prompt & Pray” programming (I’m just kidding). 

Overall, the product's first year(ish) was nothing short of an explosive, viral success, which resulted in Anysphere (Cursor’s parent company) raising a huge $105 million Series B investment from prominent VCs such as Andreesen Horowitz. Here’s a quote from their Series B announcement: 

Now—just to be clear—unlike the firms that participated in the aforementioned financing round, I don’t have a laundry list of LPs begging me to speculate on startups with their capital, so I’m probably missing something here. That said, I'm unclear about Anysphere / Cursor’s long-term edge, for a few reasons that I’ll detail below. 

Risky Business:

Before moving on, I think it’s worth acknowledging that the Cursor team seems like nice people. I’ve listened to them on the Lex Fridman podcast and emailed them personally for support and don’t have anything bad to say. However, they’re a great example of the complications of building AI-native products, so let’s get on with the analysis of their tricky road ahead.

Competition is for losers

There is an obvious threat that I want to get out of the way first: intense competition from Microsoft and other startups. Microsoft clearly still wants people using its Visual Studio Code IDE, which, you’ll remember, is Cursor’s estranged parent. A recent announcement from Microsoft about a new, generous free tier for GitHub copilot (another AI programming assistant) for all Visual Studio Code users shows that Microsoft isn’t going down without a fight (to add to the strangeness of the situation, Microsoft is a massive investor in OpenAI, which is an investor in Anysphere). Anysphere certainly has a big enough piggy bank to stay in the game for a while, but I’m not positive that getting to $100 million in revenue in under a year like they did is as much a sign of genius business acumen as it is a sign of a low-hanging-fruit product that went viral. Cursor did not create an industry or spend years selling people on their vision of the future. They just took a mature, beloved product and gave it a youthful haircut (a very nice haircut indeed!). As such, Microsoft or other startup competitors like Windsurf pose legitimate threats to any current or future profitability that influenced Cursor’s most recent valuation. 

Programmer Skill Atrophy 

Another, albeit subtle, threat is that too much help from AI actually turns out to be something that programmers sour on. You need to look no further than the author of this post to find an example of this trend. I paid for and used Cursor daily for several months, but I ended up cancelling the $20/month subscription when I realized the way in which it was dulling my skillset.

I don’t expect too many people to make this choice, especially those who have never programmed without AI, but to me, the energy required to pause, think and write well-formed questions to ChatGPT (or, dare I say, read documentation!) is well worth it in order to help with skill maintenance. One of my business hot takes is that the increased productivity that companies feel when adopting Slack is an illusion caused by the friction it removes from email, but in reality it’s that same friction that causes critical thought and efficiency in the first place. Same goes for AI IDEs (either that or AI will replace me first). What's that story about the tortoise and the hare again?

RAG and the DoDo Bird 

Here’s where things get really interesting. I recently came across a new database vendor called TurboPuffer that has a big quote from Cursor in the testimonials section of its landing page: 

So unless this is outdated, I’m going to assume that Cursor is using Turbopuffer as a database vendor. 

To understand why this is interesting, we must first visit a concept that has emerged from the large language model/generative AI movement called “Retrieval Augmented Generation” (RAG). RAG is a fancy term for a simple concept: because each language model (think: ChatGPT) only accepts a finite number of words (tokens) as input, you need to retrieve the subset of information (out of all possible info that you could theoretically grab) that you’re providing a model with carefully. For example, try pasting this entire essay along with the question “What’s this guy’s deal?” into ChatGPT – it'll probably tell you that your question is too long (phew!). Now try again, but only with one paragraph. Congratulations, you’ve just done RAG.

Given Cursor’s stated use of Turbopuffer, my speculation is that part of Cursor’s magic is that they are very good at selecting the relevant snippets from your codebase to send to AI (i.e. they’re good at RAG), which is extremely important for getting high-quality answers (the ultimate test of their product’s usefulness) when the amount of information that you can provide to a model is limited. 

But...what if...the amount of information that you can provide...becomes (essentially) unlimited? 

A brand new model from Google, Gemini 2.0 Flash is certainly making moves in this direction. The model takes up to 1 million tokens as input (here’s a good video from Theo describing the model). 

For reference, here are the token counts for the OpenAI models that Cursor was likely designed for (all of these are state of the art models, this is not a diss in any way):

That’s right, we’ve gone from 128k tokens of maximum input length (8k if you look at the bottom of the picture) in April 2024 to 1 million today! The Google model is also very inexpensive in terms of the cost per token. 

All of this is to say that Cursor’s elite ability to generate quality code could be in jeopardy if entire codebases can just be dumped into a model prompt with ease. Even the fastest horse isn’t going to outrun my ‘24 Subaru (hot dang!). 

Also, the possibility of the single-file web app is realistically already here. A lightweight Python framework like FastAPI or Starlette could easily support this with a very low token counts. This makes the ability to archive and traverse a nested file system less important, which would make a Cursor copycat designed for the non-programmer demographic easier to build.

The AI Software Development Tool Paradox 

The final sticking point that comes to mind is what I call the “AI Software Development Tool Paradox” (I’ll come up with a better name).  

I tweeted about this when I saw that Devin, a new product being touted as the first AI software engineer (what’d I say earlier about developers going after developers?) was charging $500 per month for a subscription. 

As the frugal startup founder that I am, I can’t imagine a better first question for my $500 per month AI software developer than “Hey Devin – here's how to call OpenAI’s API – can you please build me an AI software engineer?” 

This is a problem for the industry as a whole, but I can’t help but scratch my head whenever I see these businesses that sell software that builds software. Like the snake eating its tail, when these products eventually get good, they’ll devour themselves. Cursor might have the advantage right now of reduced per-token pricing from OpenAI and Anthropic (I don’t know if this is true, just saying it’s possible) but if we are to believe all of the rhetoric around everyone becoming a programmer, then it follows that everyone will be able to build their own software (note that open-source, highly-customizable IDEs such as Neovim already prove the appetite for such products among experienced developers). I would not be surprised if we see Anysphere move into the infrastructure space (like competitor Replit) to mitigate this risk.

In closing

There’s a big piece that I’ve left out of the above discussion: marketing. In 2025, there’s nothing harder than getting people's attention, and it’s obvious that Cursor has found a way to do just that. Do not underestimate this as a serious competitive moat. 

I think that accelerating a developer’s productivity or giving someone the ability to produce working code for the fairly low cost of $20 per month is a wonderful thing. I’m not trying to wish ill will on Cursor—I just think that they are a perfect example of the tough road ahead for anyone operating in this space. It’s easy to see the huge valuations and think that high-growth tech companies have it all figured out, but if I had to bet on one winner in AI, it would be the consumer (hopefully the human one).

Time will tell. See you next time!