AI Newsletter 1_16_26
Caleb Bryant
This email is the full newsletter. For a printable copy, click here to download the PDF.
Hello! Welcome to Under the Hood. Last week I talked about Google AI Studio and getting into API Keys. Let's dive right in. I should note, first of all, that while I realized that Google AI Studio is where I could obtain an API Key, I did not really put two and two together to realize that I could really create and manage my Newsletter project right from here. I had basically been "vibe coding" directly with Gemini using the public facing user interface. From there I had created a project that I had uploaded some of my files to. I expected that Gemini (or any other GPT, this is how they all work), would reference my uploaded files and remember the data, every time that I came back to it. This is not necessarily the case. Herein lies the biggest problem with using the public facing LLMs. They are great for short term conversations. However, long term conversations, projects, and memory retention are where the problem lies. This makes sense, right? I mean if the LLMs retained memory for everything every person was entering into the user interface that would be quite a load. Now, some of the GPTs (Generative Pre-trained Transformer) memory retention are better than others. Some do seem to reference uploaded docs better than others. However, in the end, they all run out of context memory. Sure, you can prompt them, "Remember this as part of my profile" or "Retain this in your main memory context", but in the end these LLMs, at least for now, are limited in that way. There are other limitations too, like the fact that they basically have the entire internet trained into them. That's not necessarily bad, except for when you want very fine, or detailed, responses and/or data. They will tend to average out contradictory information that exists on the internet giving you bland, or safe, answers. This is where fine-tuning comes in, but more on that later. Anyway, so what was happening to me as I vibe coded with the Gemini in the public facing Google site, was that any longer sessions would start to get sloppy and Gemini would forget things we already coded, or overwrite some previous existing code while implementing a new method or function. Previous code that I still needed in the program, that is. So, I would eventually learn that starting a project from Google AI Studio is slightly better than just vibe coding on the public facing Gemini site. However, as I'll discuss later, it's not that much better. But, back to the API Key. I had to come to Google AI Studio to obtain an API Key. What would the API Key do for me. Well, at the very minimum, you insert an API key into your code so that the program uses it to "wake up" a model and have it perform a task. In my case, I was using it to essentially go to a couple of websites (called "scraping")to obtain information and insert it into my newsletter (basically meetup.com and Eventbrite.com for events, but that didn't help me with any creative content). All that really did for me was save me the trouble of going to the websites myself and compiling the information manually. However, that didn't really save me anything because it would grab some data and not other data and it would populate my newsletter with bad formatted data. In order to try and solve this, I would go down a rabbit hole of hard coding my program to the most minute detail to try and get the AI model to do what I wanted it to do. At that point, why even have the AI model? I think this is what some of the business world is experiencing with AI model integration in its infancy. Yes, the Gemini LLM, using the API Key, was better than the downloaded Ollama model that I tried using on my laptop. But there was still work to be done. There had to be more value from using AI. I know I said I would talk about the cost as well, but we have run out of time and space. I will get to the cost, I promise. Until next week........
“The best way to predict the future is to invent it.”
— Alan Kay — Alan Kay is an American computer scientist. He is known for his pioneering work on object-oriented programming. Kay has also contributed to the development of personal computing, including the graphical user interface.