How many AI startups are just OpenAI/Anthropic/etc. API calls with a UI?

Earlier on, they were just wrappers. The models are like an engine, and everyone puts them in a cardboard box and calls it a car.

But we're seeing companies like Cursor which pay a lot of attention to how the models interact with everything. They're not just prompting the AI, they index files, they search and mimic styles. You can @ a certain file to use it as a reference. The composer is autonomous, even extracts commands from the AI to run, or extracts only the code needed. And it double checks that what it's writing is true. There's a whole system in there, it represents more like a modern car with proper driveshaft, gears, pedals and stuff. Copilot still feels like a box on four wheels attached to an engine.

Perplexity was a wrapper earlier on and arguably it might still be. But they've used AI more effectively than Google to figure out what a person is actually trying to search for, and suggest those things.

I think most of these companies will have minimal value added at the start and slowly start to refine it. It's hard to say what percentage.

I spent a good chunk of the pandemic doing startup advisory, and some of that has lasted up until the holiday season, so I can tell you that there was a marked shift one or two years ago where the investment board I had a seat on decided to point blank stop looking at startups that were solely focused on AI, because even the non-technical board members realized that there was little to no added value there -- and we were getting like 70% of new prospects in that category.

I would go into a demo, look at it, ask them how they did RAG on the data, ask to speak to the people doing their AI models, etc. And then sometimes I would spend a couple of hours wiring up a Node-RED flow to show my colleagues how trivial it was.

The stuff we (now they, since I'm taking some time "off" that moonlighting gig to recover from burnout) ended up prioritizing are companies that focus on business processes where triage and "human augmentation" can actually benefit from LLM summarization, some automated decision making, and some data "integration" (not just summarization, but broad correlation of events, etc.)

There's a _lot_ to be done in many fields where, say, you will notice a spike in some piece of data (using conventional ML), gather the data around that event and present it to a human (with pros and cons, including trying to flag if the data is reliable). Think GitHub issues for crop management, and you're halfway there.

_Those_ companies really need to have their use cases sorted out, and not just try to be "the Uber for greenhouse management" because they wrapped weather forecasts into an LLM.

So, in short, the real added value is expediting or improving use of domain expertise captured from (or still held by) humans.

I think as models get commoditized, that doesn't really matter. In fact, the UI is what becomes important. We can already see this with Deepseek and other Open-Source models.

I do wonder after deepseek bombshell how can any one come up with an idea, start building, and not be constantly second guessing themselves on "has the industry changed this week making everything we are doing moot?"

Its almost like you need a full time AI researcher on team, not to help build, but to keep up on all new info and constantly tell the team PIVOT!

Why does this matter? MS CoPilot uses OpenAI. One view is that if a foundational model is your moat, you have no moat. Even using commodity AI, you can still lead providing a service built on it.

I mean, how many SaaS startups are just UIs around Postgres databases?