Every trade moves the market. But by how much, and for how long? Fitted 4 microstructure impact models on BTC data to compare. The decay of past volumes turns out to matter more than the functional form
Results probably meaningless β 1 day of BTC data vs months in the paper. But fitting price impact models is fiddly enough that an end-to-end instructional is useful.
Bought a microphone for virtual conferences in 2021. Conferences went back in-person. Microphone still here. My wife lives under a constant low-level podcast threat.
Technical debt used to accumulate slowly. Now you can take out a payday loan with an AI agent. The APR is paid when the tokens are gone and its just you and the code
Started drafting a blog post when when Opus 4.6 could do no wrong. By the time you read it, someone's telling you Codex is superior. All fun and games until VSCode takes it all away from me
Opus 4.6 scores best on minimal editing. GPT 5.4 edits the most. Don't forget the magic "minimise changes" prompt! Says a lot about where we are and how far we have to go
Rewrites scored poorly; single precise tweaks scored best. If a semicolon is missing, add it, not a forest of null checks. Adding "minimize edits" helped, so prompts still matters. Still feels like you need to say the right incantations though