Codex 5.x models are “New Coke”, because the mainline 5.2 models lose against Opus in the Pepsi Challenge. I hope OAI keeps producing Coke Classic & I also wish Anthropic success with their eventual Pepsi Next.

I've written a lot of hand-crafted, beautiful code in my time. But now, I'm basically all-in on AI coding and spending my time and energy on specs and guardrails. With the latest batch of models, this is genuinely viable for substantial programs if you invest in documentation.

Claude: "You're absolutely right!" Codex: "Great idea!" Gemini: "I messed up. Overwrote the file instead of appending. Restoring from git"

Everybody is talking about "AI slop" code as if all code in the workplace wasn't already "legacy slop".

If Apple makes a humanoid home robot someday, I hope they call it the iPed. The pun is just too good.

Sounds like Amazon may be divesting from games in this round of layoffs. If you’re a talented engineering leader and are interested in Apple re-investing in games: DM me!

I generally do very little customization of my development environment & what little I did do was done over a decade ago. ChatGPT helped me modernize my setup, customize a bunch of things, and optimize startup time. Didn't have to do a single web search. Would have given up on step 2 any other way.

1. I too was skeptical of LLMs for coding until < 1 year ago 2. Current tools/models are not a panacea for anything 3. They are wildly useful tools I now use every day 4. This will induce demand for real engineers 5. Like any tool, you must invest in learning it 6. Ignore it and you will fall behind

I've actually got two roles open! Here's the other one, which might be especially relevant to former Microsofties with Windows/C# experience: jobs.apple.com/en-us/detail...

Hello Microsoft/Xbox friends, if you got laid off: Sorry to hear about that! Please reach out via DM if you're interested in working on Game Services and the new Games app at Apple here in Seattle: jobs.apple.com/en-us/detail...

If you were wondering why it seems like smoke & carbon monoxide detectors always false alarm at 3am, that's because _they do_! Ambient temperature drops lead to obnoxious battery warnings at the worst possible time. Grumble grumble.

“Consistency” is the best you can aim for if you haven’t yet discovered an appropriate Abstraction.

Hiring for my team at Apple in Seattle. Come work on some fun stuff :) jobs.apple.com/en-us/detail...

Google Search can no longer find posts on their own BlogSpot service that I know exist and contain unique words coined by that post. Literally zero results. Exact query on every other search engine I tried produced exactly the page I expected...

Letting devs write programs that do ::waves:hands:wildly:: whatever they want, is a mistake. Basically no one can be trusted to write arbitrary code without destabilizing my experience.

One of my directs had to request access to view his own IT request, so that he could forward it to me, so I could request access to view it, so that I could open another request to escalate it... Am I about to enter an infinite loop?

Wikipedia is one of the crowning achievements of humanity... and we all just kinda take it for granted now.

I'm in SF at GDC for one more day. If you'd like to meet up, DM me. (I'm not sure bsky has the critical mass for me yet for this to work...)

TypeScript team choosing Go over Rust is hilarious. Like here's some TypeScript people, who used to be C# people, who used to be Turbo Pascal / Delphi people, and they are like, "Yeah, this Go thing is pretty good." Meanwhile, the type-obsessed folks are like "BUT THE BILLION DOLLAR MISTAKE!"

"Hide Distracting Elements" is probably my favorite new browser feature in a decade.

Above ground power lines in any major metropolitan area are kinda inexcusable at this point.

I’ve always found the “control plane” and “data plane” API split to be a helpful framing, but struggle with the duality in the same way as “code is data, data is code” infects my thinking. However, I’m now convinced: Control Plane APIs should not exist…

I think it’s safe to say that “Semantic Import Versioning” and “Minimum Version Selection” have proven to have dramatically better ergonomics and ecosystem outcomes than any other approach out there…

Said it before, will say it again: I'll take an ID over an object any day of the week.

Snapshot tests – with proper tooling – are easy to write and maintain. Yes, they can be brittle, but they are also a good way to ferret out non-determinism and unintended side effects. A while ago, I created the CLI snapshot tool I've always wanted & you should use it :) github.com/deref/transc...

Found another excuse to use cuelang.org and it's really, really good and getting better.

Big UX opportunity for LLMs: Re-contextualize user-provided strings. Example: I should be able to write a lazy commit message like "fix index out of bounds" and then have log report that as is if given a path, but rewrite it to "fix index out of bounds in parser" at the top level.

The killer use case for `copilot: fix` is to catch bugs where my logic is absolutely correct, but I just used the wrong damn variable of the same type and no amount of staring at it overcomes my dyslexia.

Is it just me, or is it just plain silly that we all just live with the fact that our terminals interpret control sequences by default without any negotiation or default escaping policy? Like whoops, I accidentally cat out a binary file! I guess RANDOM STUFF IS GOING TO HAPPEN.

I'm kinda shocked at how awkward and primitive the workflow is for most folks developing against LLMs APIs. I'm realizing that it seems exactly parallel to the average dev's basic interactive shells vs a lisper's deep REPL usage with live code loading capabilities.

Kim operates Hudson’s account because he can’t type. But we don’t hold it against him because he’s so cute.

Despite their many flaws, I love the Xcom game series. But “Dark Descent” is the best Xcom-like I’ve ever played & it’s not just because I also love the Alien universe.

Can we all collectively agree that subdomain.domain.tld is the wrong order and just agree to "reverse" dns? Without remotely considering any of the certain negative consequences, I propose `w:` scheme for to replace https: with corrected DNS ordering. We can drop the // silliness while we're at it

Me: blah blah blah I'm confused Claude: "[...]. Those page breaks in RFC format can really throw you off when they split up grammar definitions. It's a common source of confusion when reading RFCs - the page headers/footers splitting up what should be a single continuous grammar rule." O_o

Because package.json _still_ doesn't support comments: node ./package.js | sponge package.json

Declarative code is more information dense, which means it takes more effort to read and write per unit. You need less of it, so the total effort required is less, but the increased intensity of effort makes it less likely that anybody is going to push through.

The biggest problem with side projects is that every weekend I spend at least 10 minutes debugging only to uncover code like this: func doThing(args Whatever) *Result { return nil // TODO: actually do the thing. }

If you want to go fast, go alone. If you want to go far, go together. If you want to go far fast, go as one.

Engineer on my team typo'd "git diff" as "git diss", and so of course had to have Claude create this for me: gist.github.com/brandonbloom...

I want an AI agent that turns any code comment in "machine checked documentation".

A Tcl implementation where all the builtins are implemented exclusively by invoking an LLM.

Interacting with AI is a lot more like managing than it is like doing the task yourself. Achieving execution through management it's all about processes and artifacts. Interestingly, relevant examples of processes and non-code artifacts are in short supply in the training set!

I've tried a bunch of AI coding assistant tools, and so far the only one that seems to function at all with non-trivial codebases is codeium.com/windsurf -- However, in the "edit" mode it really can make a mess of things pretty quickly.

I got into Twitter pretty late. I discovered the Clojure community and grew a lot as a developer during that period. I've occasionally missed that experience, but haven't seen a compelling revival... though I'm thinking _maybe_ BlueSky has a chance. And now I'm excited about AI...

Felt social, might delete later.