Want to build an AI infra company? Here's something I'd happily pay for.
Manage embeddings for me, *including the ability to refine off-the-shelf embeddings for my use case.*
Customized embeddings, as a service.
Buried in the OpenAI cookbook is an absolute gem of a guide by @sandersted on how to do this.
Given pairs of items that should be close together in embedding space, you can train a simple map on top of the outputs of a baseline embedding model. github.com/openai/openai-…
This problem is ubiquitous and high value.
Typical LLM use case: take a corpus of content (e.g. docs, code), a user asks questions, use embeddings to try to pull relevant content into a prompt, answer.
Any app like this is very sensitive to how well the embedding lookup works!
I stepped away from Heap last month, after nine fantastic years.
Here are some things I wish I could tell my 2013 self.
I was excited to write some code, and I had no idea what was in store on the path to 350+ people and a transformative product.
1/ What the company needs from a CTO is going to change every 18 months, often dramatically. You'll pretty much never get comfortable.
Being in a leadership role means your job isn't "technology". Your job is to make the company win. Get used to adapting around that.
2/ Making the company win will often mean letting go of something you think you're good at because we need someone who is better at it, or someone who can give it more time and attention.
This is hard if you let your self worth come from doing that one thing. (So don't!)