Why all app companies will become research labs

Now I am not at all claiming that everyone should start a frontier research lab. Even if you are starting one, you shouldn’t start from the core foundational approach right away unless you are Thinking Machines, where it is a group of 40 top 0.001% AI researchers in the world.

No Choice But to Become a Research Lab

But I will argue that no matter what level of software you are building, in the coming years you WILL turn into a foundational research lab. Not because that will be what is cool, but you will just have no choice. This especially applies to software startups that are in the vertical AI space, but all current application layer startups in general, hardware or software. I make this argument because at least in the software space, if you are not transitioning into a research lab there will be nothing left to do.

The Case of Cursor

Let's take Cursor for example. There is Base44, Emergent, Rork which all offer AI assisted coding tools just like Cursor. Cursor in fact started as an application layer company: a fork of VS Code with a bunch of features to call and help LLMs with your development.

But they didn’t stop there. I started using Cursor in early 2024 when Michael Truell was reaching out to randos on Twitter to work for him. Back then, having the feature to edit your code in-line and auto-complete all in one IDE was mind-blowing. But they did have their Cursor small model all the way since then. It is not something that they started after raising a fat round. Fine-tuning a foundational model is difficult even now, think about two and a half years ago. But they did start somewhere, and Composer 2.5 is now on par with the best coding models from OpenAI and Anthropic at fractions of the cost.

Intelligence Is All That's Left

All that is to say is, once you build the redundant software infrastructure of an application, all that is left for you to innovate on is the intelligence your software provides. Cursor does not need to innovate more on having agents open in a new chat window, or having positive in-line edits reflect as green lines. They at least do not have to allocate as much resources to it now. Because it is all software you have to build once.

Writing this redundant software, even though it is a one-time job, was a skill and something of value before, but not anymore. With coding agents replacing engineering teams, they are much faster at doing this.

The "Boring" Parts of Code

Writing redundant software is what Clawdfather Peter Steinberger calls boring parts of the code. How he doesn’t have to worry about what the padding in his Tailwind class on his button. All he can do now is make the proprietary Open-Claw system much faster and accurate at doing its job. And that is its own foundational model. All you have to do now is making this model better, and this requires your company to push into ML research.

Vertical AI

I myself am currently involved with working on autonomous physical workers for the services economy at Robinline, where we are building services companies run themselves. Everything from when a lead comes in, the collecting of information, sending the lead a quote, scheduling your crew, carrying out field control and collecting payments is all automated by AI. And we do this to replace legacy CRMs where humans do the job.

The point I am making is, once we build the core software for the whole CRM perfectly (the interface, the basic non-AI functionality, the database CRUD, cron jobs, error logging, etc.) all there is left for us to do is make the agents that handle lead texting and phone calls the best they can be. The most accurate they can be and the most fast they can be.

True Moat

I know a couple guys building vertical integration software in the hotels space right now at Lance, and it is the same thing. Your software is not the moat anymore: it is a one-time job, and can be done by coding agents. However, how good their agent harnesses are at handling guest requests in hotels, passing them and managing housekeeping staff, and keeping hotel operations together: that’s where they build their moat. And it never stops. They can only make that model as better as it can be, and that in turn will push them into foundational research.

The Misconception with Vertical AI and Foundational Research

I claim vertical integration is actually an imperative requirement for foundational research lab companies. People always saw these as two different kinds of startups to build, but they are just two different stages of the same path to growth. One would have argued, why would Cursor even developing Composer be a smart choice, wouldn’t GPT or Claude models just do that in the future? But with Cursor’s enormous data wedge and the fact that they are programming-focused over AI labs’ general purpose models enabled them to build an efficient model like Composer, and people will of course use it now due to how efficient it is.

So we at Robinline will have our foundational model for home services that is more efficient for our industry, Lance will have their own efficient, tailored foundational model for hotels, Legora for law, just like Cursor has one of theirs for coding, and other companies for other industries like banking, legal, audit, etc.

On Data and RL

This essay is also a very important argument for how data and RL (reinforcement learning) is going to be extremely vital the coming couple of years. Almost all of Composer’s technical credit goes to the Cursor research team's focus on RL post-training on data they collect from users as they pass in their codebase and changes. Every action and event is a reward signal, and as much as these events happen on your software, the better and faster you can make your models.

So if we at Robinline are able to get as many service companies on our software as possible, then the more fine-tuning reward signals for us to, (hopefully in the future) train a foundational model that runs a home service company by itself and be on the frontier of AI for the physical economy. Not because that is a cool thing to do, but that is will be left to do.