I figured out why "dotnet restore" performance is so bad on Windows runners as opposed to Linux runners: GitHub Actions doesn't override NUGET_PACKAGES, NUGET_HTTP_CACHE_PATH to point to the D:\ drive which is where the builds are done with faster disk I/O 👇

Stealing this prompt, we under-appreciate the value of propagandizing ourselves.

>just merge it in and then fix it yourself when they're on vacation truer words have never been spoken

If you remember this, your back probably hurts.

Unpopular opinion: Current code review tools just don’t make much sense for AI-generated code When reviewing code I really want to know: - The prompt made by the dev - What corrections the other dev made to the code - Clear marking of code AI-generated not changed by a human

i struggled a bit but chuck klosterman's Raised in Captivity is good. short story collection from one of the last great american essayists. it's hard to explain but it very much fits the bill. link in next tweet

Should be common sense, but it isn't: the better organized your code, the better results you get from LLM coding tools LLMs like: 1. Static typing 2. TreatWarningsAsErrors = on 3. Tests 4. Consistent application of patterns 5. Consistent code organization (i.e. vertical slice)

I just finished writing "Cheese Wars: Rise of the Vibe Coder", which explores the important question of whether Big Tech giants would mind if most of humanity turned into cheese wheels. The post also offers tips for getting AI on your side, once it's smarter than we are. steve-yegge.medium.com/chees…

Well, that was an interesting experiment. Tested Gemini 3 Pro vs. Claude Sonnet 4.5 vs GPT 5.1 Codex on a real-world task in my harness. Same exact inputs (system prompt, tools, pre-existing session state). Same thinking settings. Here's what I found. Problem: the TUI renderer of my coding agent harness streams out thinking traces generated by the model. It can render markdown to the terminal. Thinking traces are rendered in italic with a dim gray color. If a model's thinking trace contains an inline code block like seen in the screenshot below, then the text color of the text following the code block is not reset to the dim gray. The problem is located in two files, about 700 LOC combined. The solution is to reset the color ANSI codes after an inline block in the markdown renderer to the default color. The full changeset of a working fix is about 50-70 LOC. So not a hugely complex task that requires lots of codebase hunting. Both Sonnet and GPT 5.1 Codex where able to arrive at sensible solutions. Both zeroed in on the relevant files quickly. They did differ in the API choices they made, but the difference was marginal. I liked Codex approach better. Gemini 3 Pro just struggled. It would fully exploit its thinking token budget, going around in circles. It would reread source files already in its context, without having made any modifications to them. And it clearly was not as familiar with the APIs of chalk and marked, both used in the TUI markdown renderer, as Sonnet and Codex were. All of this combined led to a non-working solution. I actually let Gemini try it twice, starting from the same point in the session. Results where the same. I then gave it hints and tried to steer it, but it again went off rails. So, for my specific use case, in my specific coding agent harness, Gemini 3 Pro did not perform better than Sonnet or GPT 5.1 Codex, as oposed to what we saw in Gemini's benchmark results in its model card today. That does not mean it will not work for you. But for me, I see no point giving it more time if it fails a simple task that both Sonnet and GPT 5.1 Codex can do without an issue. Cost wise, it's also not a win, as I don't see Google offering anything like Claude Max or GPT Pro. So even less incentive to spend more time with Gemini and make it work. YMMV, sample of one. Go give it a try yourself. Everything above is fully documented in the live stream video below, if you don't trust my write up. And the code is all on GitHub: github.com/badlogic/pi-mono/…

.NET 10 is here! The latest release brings performance boosts, cloud-native enhancements, and AI integration to supercharge your development experience. With improved tooling for #ASPNET Core, expanded C# features, (cont) msft.it/l/6019tMo7l