Most people use these tools like a search engine that talks. Question in, answer out. That works for converting a PDF. It works badly for anything that matters, like choosing between two job offers or working out why a…
I use a chatbot to sanity check my own thinking, which is starting to look like a mistake. I ask whether a plan is sound, it says yes. I say I am having second thoughts, it agrees those are valid concerns. I push again…
I run a small product with a few prompts doing real work. Every time something new is released I try it, it feels better, and I switch. Weeks later I cannot say whether it was an improvement. Sometimes I notice…
A product launched recently whose pitch is defining safety at runtime, for text and for images, rather than baking rules into the prompt or relying on the model's own refusals. I have generally handled this by writing…
The announcement on 30 July cut API prices on the GPT 5.6 line: Luna by roughly 80%, to about $0.20 per million input tokens and $1.20 per million output, and Terra by around 20%, to about $2 and $12. Sol's standard…
Set GPT 5.6 Sol as my default in settings a few days ago. Every new chat still comes back in about two seconds with the shape of answer I used to get from 5.5, and the label on the reply says 5.5 Instant, not Sol.…
Follow on from the routing thread. Now that I know how to keep new chats on Sol, I am not sure I want to. Roughly half my usage is "reformat this", "what is the flag for X", "summarise this paragraph". Instant answers…
My work is debugging across five services. A typical investigation is app logs, traces, a dashboard and code in three repos, and the model's job is to narrow, not to write. Sonnet has been good at it. I also hit the…
I got caught out by how fast eats a 5 hour session window, so instead of just picking the biggest name on the list I am rebuilding my defaults deliberately. The cloud line up I can see on Ollama right now: glm 5.2, kimi…
We call one provider from a document pipeline. Prompts tuned over four months, temperature at zero, output parsed as JSON and written to a database. We currently point at the generic alias because someone argued we…
We got the deprecation notice for the version four of our features are built on, with a suggested replacement and a date. Spot checks look fine, and then it does something subtly different with a long input and our…