Tech, life and everything else
I previously shared my transition from Komoot to OsmAnd, and after some time, I’ve grown accustomed to its comprehensive capabilities. Whether for cycling, hiking, or general navigation, OsmAnd truly functions as a versatile “Swiss Army knife” for offline mobile navigation and tracking.
If you’re reading this, it seems that my GoBlog Telegram Bot (just another embedded plugin) is working!
This new bot probably has no real use, but it was a good experimental project to try and get a bit more familiar with the current state of GitHub Copilot, especially as we can now use it at work.
Probably the biggest use is for uploading photos from my phone. I can just send them as files and get a link back. But I can also send text and it publishes it. Even with frontmatter to configure parameters for the post.
Regarding Copilot: For whatever reason (maybe my adblocking DNS?) the VSCode extension keeps asking me to log in again, but still works for the non-chat features.
However, I am still a little worried that AI will make my work obsolete.
I tried GitHub Copilot (Free) in Visual Studio Code again for some small GoBlog changes. Copilot can now generate tests (although it doesn’t feel intelligent, as you need to correct quite a few things), it can do code reviews before committing and it can generate commit messages. Of course, it can also do code completions and write complete code, if you want it to do so.
To reduce my dependence on USA-based products, I switched from using the OpenAI API to Scaleway’s Generative API for my blog. Not only is it cheaper, but it’s based on open-source models, hosted in Europe. 🇪🇺
It was straightforward to just add an option to change the REST endpoint in my GoBlog plugins, and then use Scaleway’s API for creating the post summaries and the image descriptions. For the post summaries, I am now using llama-3.3-70b-instruct, for the image descriptions pixtral-12b-2409.
Let’s hope that there will also be a model that supports audio transcriptions until April 2026, so I don’t have to renew the credits on the OpenAI API for my Telegram Voice-To-Text bot. 🤞
Although I also use AI for some features on this blog and sometimes chat with some AI agent (whether it’s ChatGPT, Claude, Microsoft Copilot or GitHub Copilot), I have mixed feelings about its mandated use at work (Shopify is just one company doing it).
My blog still has a long way to go when it comes to accessibility, but I’m excited about two new plugins I’ve just added to make things better.
The first plugin uses AI to generate image descriptions for me – far better than I could write myself. It even works in multiple languages, adapting to the blog’s language settings. The second plugin improves how these descriptions can be accessed on mobile devices. Now, you can simply tap on an image to read its description (on desktop, the descriptions are still available as tooltips when you hover over the image).
Since image descriptions aren’t created automatically for older posts and I haven’t manually updated all of them yet, you can check out this post as an example of how it works.
Every few months, I inevitably catch a cold or some other virus and end up stuck at home for a few days. Right now, I’m in one of those phases again. Thankfully, this time hasn’t been too bad, and I’m already on the mend. After a lot of Netflix marathons, I realized I couldn’t just sit around doing nothing – I’m the kind of person who needs to feel productive, even when I’m under the weather.
It seems like I finally fixed a memory leak in GoBlog yesterday, that sometimes made my blog crashing. How? I used Anthropic’s new Claude 3.5 Sonnet to write me a new HTTP compression middleware that compresses HTTP responses using zstd or gzip when possible. I needed to instruct a few changes and modify some code lines as the initial implementation was wrong, but thereafter, it finally seems to work better than my original implementation that probably leaked some objects anywhere. Claude also helped me to write unit tests (which I adapted a bit, though) and increase code coverage.
Even though I wonder how they finance it (reading their post suggests, there will be a paid option later), DuckDuckGo’s AI Chat is probably a good privacy-friendly way to try out the different LLM models for chats. You don’t have to register, you don’t have to pay, you don’t have to change settings that your chats aren’t saved or used for training.
Due to OpenAI’s move to prepaid billing, I (finally) had to pay for using the OpenAI API. But this also allows me to finally use the gpt-4-turbo-preview model as I have reached tier 1. Now the AI-generated summaries on this blog will be created using this new model. 😄