Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/mmbazel/aie3-week01-build-custom-gpt


https://github.com/mmbazel/aie3-week01-build-custom-gpt

Last synced: about 5 hours ago
JSON representation

Awesome Lists containing this project

README 

        
# AIE3-week01-build-custom-gpt



Week 1 Assignment 1: Breaking Into Data & MLOps^TM - By MikikoGPT



- I created a custom GPT called "Breaking Into Data & MLOps^TM" that helps people get answers on pivotting to data analytics, science, or MLOps.

- I get a ton of messages on LI from people asking me for advice and help, partially because of my role as a DevRel but also because I spent a couple years creating various forms of content about my journey becoming a data scientist without a masters or PhD. 

- The messages tend to be pretty similar and often times involves re-directing folks to blog posts or videos I've already crated (but that tend to be really long because I consume ins tream but produce in batch). 



Could I outsource answering messages so that not only am I able to focus on long-tail cases but I can also unblock the folks that ask me for help? 



So I created MikikoGPT. 



Other reasons why I liked this direction: 



- I essentially have an easily accessible, reasonable size training data based on content that I personally created and that I can speak to the ownership & provenance;

- I can evaluate the quality of the responses based on actual conversations;

- I can share the GPT as a useful resource as well as customize further as an internal tool.



# Links



- GPT Store Link: https://chatgpt.com/g/g-a1zVL5Jkf-breaking-into-data-mlops-tm-by-mikikogpt 

- Loom: https://www.loom.com/share/ff7fb53122fc4f2ebf0f23c5d18ec804?sid=d9d29f25-971d-4b49-ac25-0c86cd4eab64

- GDoc write up: https://docs.google.com/document/d/1f5Xy7X2pBIwQwBDyUg82IGalMu1YRRnNHixBe8Js2Vk/edit?usp=sharing



# What I learned

## The Pros

- In cases that don’t require a lot of personalization (for example, the message is generic or short), MikikoGPT can give a reasonably comprehensive response in a minute fraction of the time that I can.

- Adding memory (especially the linkedin timeline) is really helpful for some grounding.

- Especially in cases where I’ve written very long posts, using a chat-interface can help deliver the 80/20 of information in a digestible format.



## The Limitations

- Details can get lost, especially with regards to specific links and technology.

- Although MikikoGPT captured the spirit of the advice I’d give in most cases, it didn’t necessarily offer strong opinions, which I think is one of my primary value-propositions.

- There’s additional information I usually leverage as part of my recommendations, including the candidate’s goals, their LinkedIn profile, experiences and skills.

- Hallucinations and links: Although MikikoGPT did reference the uploaded documents at times, links weren’t given even though the blog posts I’ve written included at least 50 links of resources. 

- Maybe because the files were uploaded as pdfs and some of the links didn’t get pulled in?



# Questions I still have 

- How can I ensure that MikikoGPT is updated with recent, relevant information?

- What are the boundaries of MikikoGPT and how do I maintain or influence them?

- I.e. What do I want MikikoGPT to do versus when do I want MikikoGPT to redirect users to contact me directly? 

- The tone and style still feel more like ChatGPT and less like me.

- How can I safely integrate MikikoGPT to take actions on my behalf but also have a review and intervention stage so I can update or influence the response?

- Finally, how can I make sure MikikoGPT is being fine-tuned when needed?