DataScienceGPT

Instructions

To login, go to the DataScienceGPT tab and use the code you received by email after buying the Chatgpt course.

The example tab on the left has some Q&A examples. Also, you might want to check out the DataScienceGPT Highlights section of the course. It has a list of questions/answers that I found particularly interesting or related to novel concepts, with some additional comments/context from me.

Finally, make sure to actually go through the course before using this tool, especially the sections Chatgpt - Foundational Knowledge and Chatgpt - Practical Projects. Understanding how the model is built and knowing how to ask a question has a significant impact on the quality of the answers.

This model has been built almost exactly the same way as the one showed in this lesson. The main difference is that in that lesson I am just storing in the database the 40 DS product questions. On the other hand, here the database is massive and consists of almost a decade of DS-related personal documents, notes, interactions with people with inside knowledge, and online material.

The other difference is that this model uses GPT4 instead of GPT3.5. GPT4 is much slower and more expensive, but it is worth it. The difference is incredible, especially when the model has to 'think' (e.g. What kind of data could I use as a data scientist to show that users might be interested in a new feature X?).

Currently, the model doesn't implement conversational memory. It is mostly meant for asking specific job interview questions vs having chat-like interactions. If asking follow up questions, make sure the follow up includes all the relevant information, similarly to the Airbnb A/B test example in the DataScienceGPT Highlights section of the course.

Re: how to ask a question, the main concepts in the course apply, such as what's described in the How to ask a question to Chatgpt lesson and the two Product Data Science via Chatgpt challenges. When I played with this tool, I mostly used/tested it for two use cases:

1) Standard product data science job interview questions. My impression is that the model is suprisingly good in giving the right answer for a variety of questions. I mostly tested it on insights, feature development (should we test feature X), metrics, and A/B testing.

If you ask the model to guess something or hypotethical scenarios, at times the model returns 'I don't have this info'. If that happens, try forcing it to reply by adding to the question something like 'If you had to guess', 'It's fine if you are unsure. I just want a guess', etc. When it finally gives up and answers, the answers are often shockingly good.

2) Inside knowledge (e.g. can you list the key metrics at company X, can you describe how company X performs Y, how does company X defines Z, etc.). Most inside knowledge is related to the classic large tech companies that have hired a lot of DSs in the last decade: FB, Airbnb, Uber, Twitter, etc.

In this case, results are mostly hit or miss, depending on whether the information is there. If the model says it has no knowledge about it, try rephrasing the question using the company product type instead of the company name, e.g. messaging app instead of whatsapp (btw this is a good advice in general)

For any feedback, feature requests, comments, etc., just email me. Please, keep in mind that this tool is meant to (a) complement the Chatgpt course and (b) help people learn about Product Data Science and, as a consequence, also get ready for interviews. Using it in real time to cheat during video interviews is not why I created it and I definitely don't support that use case. Not to mention that that strategy is unlikely to succeed anyway. The model might take too long to return the results for a live interview AND any decent interviewer realizes right away if the candidate understands what they are talking about.