The slow journey
As I continue learning about AI, I’m facing issues and obstacles. A lot of those can be confusing at first: acronyms and terms alone can be enough to scare you. I don’t like horror movies, and yet, I keep going despite this unknown world. Having the end goal in mind and the potential reward at the end are both reasons to continue on this journey, but there’s more to it.
When I look back at the state of the current code I produced for this project, I’m not proud of it. I mean, I’m genuinely happy for what I achieved so far, and I know I’m still learning, but it doesn’t change the fact that it looks like one of many roads under construction: chunks of the code in the wrong spot or temporary fixes like when they patch the road, but clearly it needs a complete makeover.
For a long time, I’ve been too hard on myself as I thought perfectionism was a good thing.
In fact, without saying a word, it’s an expert gaslighter ready to make you feel inferior as it reminds you you can’t be as good as them.
Keeping track of my progress has been very helpful. When I say progress, I mean adding the “W” (good) and the “L” (bad) as my son would put it: both things that happen during the sessions.
For example, this week, I was trying to retrieve some data that was ingested in a vector database, and I wasn’t sure why the results were not as good as I expected. I thought I did everything right:
Adjusting the temperature (the degree of creativity of the LLM, from none to hyper-sensorial being -- I set it to “0.0/none” for factual answers only)
Getting the system prompt with rules to follow to always give me the result in the format I want
Testing on 1 type of document to avoid confusion because of styles
I thought I was good. Nope! After much digging around, it turns out that I tested so many different things that I forgot a loop in the code that was used twice when I filtered the results. This meant I got the results back, filtered for quality and then re-added them after. I was essentially confusing that poor LLM and it was then showing answers that made no sense or were not aligned with what I asked.
Overall, this meant the quality control was VERY poor, and it was my own fault because of the amount of code I didn’t clean up yet.
After more cleanup and tuning details in the chain of prompts I use, I was able to get much better results, and now I understand a bit more. I was giving instructions to do something, but was not “clear” or was contradicting myself by mixing details in those prompts. This ended up confusing the whole system even more and quality dropped drastically.
The lesson I got from this is no matter where you are in life, you will stumble and make mistakes. The worst you can do is start blaming others. I was literally giving the wrong instructions to the system and I was expecting it to “figure it out”. You have to take the time to figure things out and ask for help when needed.
You have to be ok with you not being perfect and flawless, but once you accept this baseline, you know where to start to get better.
This is a constant battle and it’s worth it. Slow progress is still progress.