About this book and the project
This project is based on a book I wrote (with contributions from my esteemed colleagues) to help people new to startups to find out more about what are they are getting into. It's also a good primer for anyone wanting to learn more about founding a startup, mostly from an (East) European point of view. The book is primarily targeted to those without extensive experience in startups, or even business in general, and covers topics such as finding the right idea, how should a great MVP look like, how to assemble a team, how to approach and present the idea to potential investors, and what to do when first investments start coming it. The book covers concepts, skill and attitudes for a seed-level startup. It's ideal for students.
This project (AskTheBook) was created because of the my fascination with GPT, and what's currently possible with natural language processing models. It kind of looked magical so I wanted to explore how easy it is to train or fine-tune a model, and it just so happened that I had a book to train it on. During this project I've come to appreciate both the good sides and the bad sides of doing this with GPT, and I've learned a lot. For example:
- I have a big respect towards the quality of training data, having seen what the algorithm can do on my own book.
- Also, it's VERY expensive to run. I might turn this into a business, but it won't be cheap.
- I have a great appreciation for the models offered by OpenAI, and I kind of got the hang of the positive and negative sides of them.
- The way of interacting with natural language engines is completely mind-blowing for me, as I had extensive and really deep experience with "normal" computing. Once I was a kernel developer, and I still can whip out a couple of assembler lines if I must. The fact that I don't work with this interface by giving exact parameters in exact formats and don't get exact data in return, but a free-formatted text, is literally psychodelic. I like it.
This project was done completely with OpenAI's APIs, and that's with their second-best model,
curie. I would have used the best one,
davinci, which is the one probably used by ChatGPT, but it's too expensive to expose to the general Internet for free.
OpenAI has released their previous algorithm, GPT2, for public use, but have not yet done so with GPT3. The main reason I've used OpenAI's APIs is because I wanted the GPT3 experience, but from my experiments with self-hosted GPT2, if trained well, it still might be good enough for something like a book. We'll see.
There's supposed to be a blog series on how I did this, but it's falling behind the actual project. Stay tuned.
Do you want people to ask your book questions? Contact me at email@example.com.