A post happened yesterday that is a part of the process of daily writing. I’m getting a little better at using the Gemini 2.5 model for making thumbnails. You have to be really careful about making subsequent requests. Your best shot for success is on the first iteration or just composing the prompt again and making another initial request. It only took a couple of requests to get artifacts and degradation in the image being created for the Substack post thumbnail. In this case, the image being created is really simple. It is not complex at all; you are talking about a title and a background. This should be easy enough for the model to handle. Sadly, that was not the case. My experience with coding models is similar to that degradation after multiple requests. You generally either get what you want on the first run or you are going to start falling toward a downward spiral pretty quickly if things don’t work out.
Today I elected to just use the year, month, and date format for the title with a simple YYYYMMDD for this super exciting missive. Asking the Gemini 2.5 model to just make a simple image with the date actually took two distinct tries. The first one has the title overlapping the artwork which was just problematic. The second attempt however worked out well enough and you can see it if you were on the homepage.
Each one of my Pages documents saved on my MacBook Air is getting stored up to a Pages folder in iCloud, which is interesting. At some point, it will be a huge collection of separate documents. My Google Drive has a similar collection of pages that just sort of sit and wait for future Nels to pull them all together into a writing corpus for training and query purposes. It has been some time since I tried to use that backlog in its entirety as a corpus for training a model. Like the last time, it was during the great ChatGPT 2.0 experience. This upcoming week I’m going to refresh my local LLM efforts with the new OpenAI release on both my Windows machine and this MacBook Air. That should be a fun little adventure.
It’s entirely possible that next week opens the door to streaming again. That was once a part of my daily routine and then those efforts stopped. My plan is to start with Ollama and VS Code and see where that combination ends up taking me. It should be a fun little adventure.