Five confident guesses, a bluff and a reality check
Okay so I was just playing around with my local model (Qwen3.8-Flash-Next) the other day, I mainly wanted to test the vision capabilities of the model. And the chat turned out to be more interesting than I expected, especially once you open the Thinking blocks. The model has no trouble identifying the separate features. It reads the glasses, the clasped hand, the white hair, and it is right about all of it. Naming the man is a different act. It names the wrong person five times anyway, and the fifth one is a bluff. That is the problem with these models, they hallucinate and they sound sure doing it, so if you are not very familiar with a topic and you do not verify your answers you will be led astray. There is a second reason to open the thinking blocks, the model spends part of them working out what I am feeling, once way off, deciding I am frustrated after four bad guesses, and once eerily right, in the block after I say there you go. The conversation ends with me making it reflect on its behavior, and it thanks me for that and encourages me to keep doing it. I tell it that it will not remember this tomorrow.
Reflections on collaboration, burnout, and the need to share responsibility for change
Gary Stevenson’s video about ending regular weekly uploads isn’t just an update—it’s a profound question about creating change without burning out. Stevenson grew up poor, then made a fortune as a Citibank interest-rates trader. His experience gave him unique insight into inequality—both from inside the financial system and from the outside looking in. He later built Gary’s Economics on the relationship between wealth inequality and falling living standards.
A simple guide to RSS readers, without algorithms or accounts
Your phone knows what keeps you scrolling. It has studied how long you pause on each post, what makes you tap, what makes you stop. Every refresh reshuffles the deck to give you just enough that is interesting to keep going. RSS does none of that. It gives you a list, in the order it was published, from sources you chose. No ranking, no reshuffling, no guess about what will keep you there.
Nextcloud, Gitea, Hugo, and a Raspberry Pi — all on machines I control
I wanted to host my own website without giving up the convenience of modern cloud tools. The usual easy version would be to spread the pieces across familiar services: Google Drive or OneDrive for files, GitHub.com for version history, and a hosted platform for the website itself. That can work. GitHub is a very useful hosted Git service, but it is still a service owned and operated by someone else.
Running a small AI chatbot locally with Ollama — no programming required
Most people first meet AI through a website like ChatGPT. That works well, but it can also make AI feel like something that only exists inside someone else’s service. In this post we are going to try something different: running a small chatbot on your own computer.