It's a strange feeling when the AI coding assistant you're building starts helping build itself.
It's the ultimate dogfooding test. Every weakness becomes obvious. Every useful feature proves its value immediately.
What a disaster day with the change in Copilot pricing model to usage based. With my nomal usage I exhausted by Pro credits on day 1. Now if I keep using it I will be spending easy $500 - 600 per month. #copilot#microsoft#github#disaster
@acolombiadev Today is day 1 and by afternoon I have already consumed 35% of copilot usage. I am using pro plan. I think extra usage will cost me up to $300 per month if I keep using after this. Have to move to Claude code 200 plan which is way way cheaper and can do lots with that
AI is a cheat code for life.
i know literally NOTHING about data centers. ZERO.
i just asked the AI to "save me money on electricity" and it told me to "stop using the AI."
efficiency: 100%.
we really living in the future.
@KaiXCreator I use Ollama which give access to many models and easy run via claude code, open code and codex. Claude Code with Kimi 2.6 and sometime with deepseek 4-pro also works very well if you want to save cost
Love spec driven development. Replaced 40 CRUD modules with one engine. A module is ~4K tokens now.. single shot, no back-and-forth patching. Took 3 rewrites before the abstractions held. Spec-driven isn't for 3 screens. It's for when you can't bear writing the 41st.
I love spec driven development. AI app builders shouldnβt jump from idea to code. First turn the idea into specs: modules, fields, workflows, UI. Blueprint before build.
It should help you think through the product clearly before the code exists.
The things that humbled me:
q4_0 KV cache quantization made the model loop the word "wise" 200 times. Took me embarrassingly long to trace that to missing repetition penalty in the sampler.
#buildinpublic
Built a local AI assistant with zero dependencies beyond .NET. No Ollama. No Python. No llama.cpp CLI. Just dotnet run, drop in a GGUF model, and it works.
https://t.co/WrDL98Xyc0
Confessional narrative about how Blazorly optimised for speed but kept shipping the wrong thing. Live preview reframed as a trust mechanism, not a feature. Closes with a reframe: "We stopped trying to be fast. We started trying to be believable. That turned out to be faster.