I tried out a new AI tool today, Cursor AI. I highly recommend it, especially if you have a coding background. At first, I tried making a toy app that takes a list of colleges and gets public web information and makes sense of it to gather all the relevant dates and deadlines for applying to those colleges. It used a locally running LLM (ollama with deepseek R1 quantized model) and within about an hour it was a working web app! Most of the time was spent taking error messages and feeding it back into the prompt so it would correct certain aspects of the app. It was really impressive that it could make something that actually worked and if I had a more powerful local LLM or if I wasn’t too cheap and used my anthropic/openai api credits, it would have probably run a lot faster (but i’m sure anthropic would have rate limited me :p I’m still sour about my experience with that…) It was really cool to be able to use an LLM in my coding adventure.

Next I thought, ok let’s take this a step further and since I work with a lot of images in marketing, let’s create an app where I submit images into a vector database and search on it via text and retrieve results. So this isn’t a keyword search, it vectorizes the images and it vectorizes my search query text and then does the retrieval that way. It was even able to take an image and search via the image. and I was able to create this in about two hours. part of why it took so long was that I did it at a starbucks that I swear was limiting my wifi, and it kept on redownloading various dependencies. but anyway, it eventually worked. I am astonished. I’ll include a video of the app in action. It is far from refined, but this was 2 hours of work from someone who hasn’t coded in many years.

youtu.be/tIayo40TX…

I heard that a lot of would-be computer science majors are turning away from CompSci because of the fear that AI will take their job. After going through this, I think it’s more important than ever to have comp sci engineers. What I created was a toy, just to prove a point. If I wanted to scale this to make it a real workable application, it would need real expertise. And someone with that expertise would be able to use tools like this to be so much more powerful.

I think we’ll see future startups where there is a full stack architect and a tech-savvy business person who knows the tech capabilities and the customer needs. There are some unicorns who can be this all-in-one package, so, maybe one-person startups might be possible. It will be really interesting to see.

This was really eye-opening for me and it reminded me of the joy of creating something. If I had to do this without cursor.ai, I would never be able to find the time to do it and balance work and family and sleep.

update to last blog article: since I last posted that deepseek wideseek … blog post, I’ve researched a bit more on what makes R1 different, and it was much more than the mixture of experts (MoE) in training. They incorporated a lot of different techniques in pre-training, RL, and post, as well as engineering down to assembly level code to get the most control and efficiency out of the hardware they had to work with. So that last blog article is a gross misunderstanding of why R1 is different.