We Pinned Our Model Version. The Provider Deprecated It Anyway.
Why re-qualification, not inference, is the recurring cost of production AI
About We Pinned Our Model Version. The Provider Deprecated It Anyway.
Pinning a model version does not protect you from deprecation. This Towards Data Science piece argues that the recurring cost of production AI is not inference but re-qualification: the eval reruns, prompt retuning and regression testing you owe every time a model changes under you.
It breaks down what that tax covers and how to budget for it before a deprecation notice forces the work on you.
Read the guide on Towards Data Science.
More Guides
AI Overviews Cut CTR by 23.1% in France
Google launched AI Overviews in France on July 22, 2026. Ahrefs tracked 963 domains in Google Search Console for 28 days before and 9 days after, and the most exposed sites lost 23.1% of their click-through rate. If search is your main acquisition channel, this is the data to read before you plan next quarter: it shows which kinds of pages get answered without the click. Read the study on the Ahrefs blog.
An honest side-by-side of two coding agents from someone who actually used both
Pi Agent vs OpenCode after 100+ Hours of Real Use
A comparison written after real sustained use rather than a weekend trial, covering where each agent holds up and where each one falls over. The useful kind of tool comparison: specific about the failure modes, and willing to say when the answer is "it depends on what you are building".
Turn Codex from an interactive assistant into something you can script
Running Codex as a Headless Agent
A practical walkthrough of driving Codex headlessly, so it becomes a programmable step in a pipeline rather than a chat window you sit in front of. The shift from interactive to scripted is where coding agents stop being a productivity toy and start being infrastructure. This covers the mechanics of getting there.
Time to First Token
Time to First Token is a structured 10 week curriculum covering LLM inference serving and optimization, designed around 30 minutes of study a day. It walks through batching, KV caching, quantization, throughput versus latency tradeoffs and the serving stack choices that decide your infrastructure bill. Worth the hour a week the moment self-hosting or inference cost stops being a rounding error.
Standing Rules Instead of Permission Prompts
Ashley Childress spent four and a half months rebuilding how she works with AI coding agents, and the core move is a good one: stop treating permission prompts as your safety system. Clicking "approve" forty times per task is not oversight — it is fatigue theater. She replaced it with 134 standing rules that define what the agent may do up front. What the guide covers Why click-to-approve stops being a real guardrail once volume goes up. How to write standing rules that constrain behavior instead of interrupting it. Treating the agent like a direct report: context, expectations, and review at the right altitude. Nine concrete practice changes, each with the reasoning and evidence behind it. Why it is worth your time This is the rare "here is how I do things differently" post that shows its work. It is especially relevant now that agents are being handed more autonomy by default — the guardrails have to move from per-action prompts to written policy, and this is a working example of what that looks like. Read it on dev.to .