At least 27.2 fixed these UI bugs

Just updated to 27.2, and curl is still vulnerable (without custom mitigations applied). Seems like this should be a bigger priority. Be safe out there.

So Apple still hasn't shipped an official update to the macOS built-in CLI CA bundle? If you aren't overriding /etc/ssl/cert.pem, then curl still accepts the bad certs? Seems crazy that they're not on top of this...

Another incredible image, and what a flex – focus stacking when there's motion is hard mode.

I've always loved focus stacked photography, but this is next level. Incredible!

I won't truly feel comfortable delegating work to AI until we have the current frontier-level running 100% locally. Right now, cloud-based AI is the ultimate dependency. When building software, I've always minimized dependencies. You could outsource so many things to libraries, but you gave up control, flexibility, and some security. Sometimes it was worth it, but I could often do better. Now it's easier than ever to build dependency-free systems with AI, but if you go all-in, you have the ultimate dependency: you can no longer work on the project yourself and keep your sanity. I've built some small personal projects with AI and on the outside it's absolutely incredible, but every time I "open the hood", I find a fractal of questionable decisions. The only way to reasonably continue development is to keep using AI. But relying entirely on these companies feels so fragile and I hate it. I get chills every time anthropic refuses a request. Weird times.

Naming things well is still hard. Language models haven't fully alleviated this yet.

You need to be following @vanilagy. Just oozing expertise, craftsmanship, and innovation. You'll also frequently see discussion of low-level technical tradeoffs, which is so refreshing in this era.

I would love to see native-res rendering and MSAA come back by shading in texture space and temporally amortizing the work there. I think this is what @mcnabbd has been experimenting with as well.

Remotion is awesome. I wanted a code-based video editor, and now I know how I'm making video content going forward.

I'm still having fun playing with this. Radiance cascades + lots of physics objects is just awesome. Check it out yourself! cmiller.tech/rc-pegboard/

I have strict rules around the usual AI problems, like avoiding overly defensive logic everywhere and creating premature abstractions for simple things. Failures here can't be caught with deterministic checks, and require good judgement. - I work with a strong model like Fable to develop an aligned plan - then have a swarm of grok 4.6 agents review the plan for decisions that go against the engineering principals I've laid out - we make corrections if needed - fable writes the final plan as a thorough handoff and sanity checks it for completeness - grok 4.6 or gpt-5.6 sol implement the plan - deterministic checks must pass - a swarm of grok 4.6 agents review the implementation against the plan and the engineering principals, and strongly agreed upon improvements are made Then I read the code, and frequently find exactly the kinds of maintainability problems the rules and reviews are already designed to weed out. I just spent 2 hours reviewing and heavily refactoring a feature to fix a pile of fresh technical debt. Fable said the new code was significantly better than the original implementation. I said what can we do to encourage the first pass to be closer to where we ended up and avoid these problems? and it responded: > Every one of your refactors is derivable from artifacts that existed at planning time (your rules, the module's stated purpose, the codebase's idioms). I don't think better rules get you there. It also said custom linters / deterministic analysis can't help, and then suggested adding hyper-specific planning rules based on the exact code we just wrote, which would not generalize well. So, no good advice. Even with the self-healing steps I cannot get agents to consistently write good code.

I've been trying to instill good engineering habits in an agentic workflow for a while now. I keep iterating on it as I find new issues, but I'm becoming convinced that AI generated tech debt is unavoidable.

I polished up the experiment and now it's a toy you can run yourself. If you have a strong GPU or a new phone, I recommend increasing the quality in the settings menu. cmiller.tech/rc-pegboard/