Todd Stumpf @stumpf

Well. Fuck... Hope you're having a better day than I am. On the plus side, I am tweeting from a pool. So there's that. 🐬

Oh, dear! This turns out to be a terrible idea if you pay for your own tokens, as it turns out. w/o tables: - HTML5 +40% token count over MD - LaTeX +33% token count over MD w/ tables: - HTML5 +35% token count over MD - LaTeX +51% token count over MD Markdown for the win!

I mean. I had this thought first, basically. A few hours ago. Making jokes about the need to reintroduce xhtml. Not really as well considered. Or documented. Or rationalized. Or proven. But it was my own thought, dammit. But read this:

Thariq @trq212

3 months ago

x.com/i/article/2052…

1K 2K 18K 14.5M 34K

Called a single threaded sqlite server "a toy" and Claude took exception. We got into it. It got ugly. Claude eventually called me "wrecklessly categorical". I've never been so proud of an LLM before. 🥲 #claude47 #sqlite #llmshade #aitherapy #aitwca

You can lead a programmer to water, but... #A2A #MCP #geminiart

The most recent Tahoe version made my #Warp and #Tailscale VPNs hate each other. Fortunately I found this blog post: mxcao.me/posts/tailscal… I wasn't clever enough to tweak the split-tunnel via warp-cli, but I was able to define a #Cloudflare #ZeroTrust device policy that did.

I really can't tell at this point: who is the rubber duck in my technical discussions? Is it the AI? Or me? 😅

Re: #antifa: Release the #EpsteinFiles

I am stunned I need to point this out, to both sides, but here we are: Gerrymandering disenfranchises voters, and is as evil as poll taxes or voter intimidation. Do not support someone who is so misguided they feel denying others their vote is okay if their own side wins.

There is a new 🐐👑! #ChatGPT and #Claude chased their tails for two days on this e2e test for a teleporting menu -- #GeminiPro 2.5 solved it in fifteen minutes! 💪 Thank you #Roo Code -- you made my day!

Another late-night Claude Code post. First, if you've just arrived here at the party, Claude Code is NOT the same thing as Claude 3.7 Sonnet, Claude.ai, nor any other Claudey thing. It is its own new experimental thing, also from Anthropic, makers of Claude and Other Claudey Things. Claude Code (CC) is a new coding assistant, one which, strangely enough, only runs in a terminal. Like an xterm, or a bash shell. Or any of six WSL shells that don't quite work. As a result, CC looks comically retro-futuristic: A late-1980s vision of what AI might become. And here we are. CC is what we all thought Devin was going to be last year. When Devin came out in December, people Muntz-laughed and moved on. But CC is a bona-fide AI software engineer. It deserves the title. And the funny thing is, they appear to have eschewed RAG completely, and just told Claude to go figure stuff out on its own. I am here to say that I am addicted to Claude Code. I can't put it down. I don't mean that figuratively. I mean I literally do not know how to put my computer down and go to sleep. Because Claude Code keeps doing stuff. It keeps solving massive problems, one after another. I throw larger and larger things at it, and it is unfazed. Chomp. Chomp. Chomp. It's like that old Assassin's Creed game, Rome maybe, when you had that big network of spies working for you towards the end of the game, and you just sent them out on missions while you sat on your fat ass, and it was absolutely just as much fun as "running" around the game world? Well I remember. This, is that. You know what? We can't be more than 2-3 months away from being able to say, "Yo, CC, just... go make tests. For everything. All the stuff I failed to test over the past 2 decades, go redeem me. Write tests for it all, and make sure they are clever and meaningful, and follow our testing patterns." And then you deposit like, I dunno, $5000 into its gaping maw. It just goes off for a week or two, doing its thang on a branch somewhere, mostly I/O bound waiting on your builds. And one day you get the email you've been waiting for. It says: "Send Money". After a few more weeks of this, it finally takes you from 10% to 90% test coverage, so that when you die you will be admitted into Good Engineer Heaven. All other coding assistants will follow CC's approach, in some form factor. They are all falling over themselves right now, as we speak. Because yes, to answer all your exact same FAQs: Claude Code is that much better. The race is on!

I have no idea what to call our book, so I’m appealing to all you MFs who are making this so hard. Right now I'm thinking of changing the working title to "No Vibe Coding While I'm On Call!" @RealGeneKim and I are nearly done co-authoring a book on chat coding. We were gonna call it The CHOP Handbook, for chat-oriented programming, which is like regular engineering except the LLM writes the code and you just chat with it. Then comes along vibe coding from the famous Dr. @karpathy, promising to waft all your pain away, and it takes the world by storm. But the thing is, almost nobody read his entire wall of text tweet. Gene and I were chatting with @jessieay from GitLab about AI metrics, and at one point she said in passing, “So I told my team, no vibe coding while I’m on call!” Wow. I was really curious and asked her later if she had meant that her team was going to be checking in code they didn’t understand, and that she would have to debug it while she’s on call. Without missing a beat, she just laughed and said, “Yeah, did you read the tweet?” And funnily enough, just like you, I hadn’t either. Everyone stops after like 3 sentences. Give in to the vibes… embrace exponentials… barely touch the keyboard… ask for the dumbest things… the code grows beyond my regular comprehension… And everyone’s like, YEAAAAAAH, hell yeah, I love turning my brain off! We can code without thinking? Sign me up! They go full Milhouse: “I’m not only not learning anything, I’m forgetting stuff I used to know!” After the first half Dr Karpathy’s tweet recommending that you turn your brain off, they stop reading the rest of what he said: …not too bad for throwaway weekend projects… it’s not really coding… it mostly works… hey, where is everybody? Not real coding? It only mostly works? It’s for throwaway projects? Dude he’s saying don’t use vibe coding at work. It’s a lot of fun, sure, use it when you can, but not for real work! In my mind, CHOP = Vibe Coding + Engineering. You still do all the vibey stuff, but you leave your brain on, and you use chat to engineer real production systems. Vibe coding by itself is just the world’s biggest game of pin the tail on the donkey, with you wearing the blindfold. So you tell me. What do we call our book?

I've been using Claude Code for a couple of days, and it has been absolutely ruthless in chewing through legacy bugs in my gnarly old code base. It's like a wood chipper fueled by dollars. It can power through shockingly impressive tasks, using nothing but chat. You don't even select context. You just open your heart and your wallet, and Claude Code takes the wheel. It even helps keep you in the loop by prompting you every eight seconds to ask if it can use basic read-only commands that you would allow _anyone_, even North Korean hackers, to run on your machine. But you learn to watch it carefully, because it pushes *hard*. As long as the bank authorizations keep coming through, it will push on bug fixes until they're deployed in production, and then start scanning through the user logs to see how well it's doing. Claude Code's form factor is clunky as hell, it has no multimodal support, and it's hard to juggle with other tools. But it doesn't matter. It might look antiquated but it makes Cursor, Windsurf, Augment and the rest of the lot (yeah, ours too, and Copilot, let's be honest) FEEL antiquated. I know it's experimental, and we don't know all its limits yet. But from my experience so far, it feels like a bigger step into the future than any we've seen since coding assistants came out. So. Not only does Anthropic have the best model out there for coding in the real world, they also seem to know how to use it better than anyone else. Between them always having the best models, the best chatbot interfaces, the best predictions from their CEO, and now Claude Code -- I've begun to suspect that Anthropic is literally the only company on the planet who knows what the fuck is going on.

I finally tried out Claude Code today. I used it to track down a bug that's been plaguing me. The session, which was like 5 minutes long, went like this: Me: Hey Claude, help me track down this bug. Claude: Here are a bunch of things you already tried. Me: I already tried all those things. Claude: Oh, well, heck, just use a debugger then. What happens on line 1657 during startup? Me: Use the debugger and immediately find the bug. Claude: Me: Bah! What good are you? I solved this the old-fashioned way! Claude: So yeah, it's pretty good. I guess. I'm still hurt.

#NoKings #NoKingsDay

... Endure. And in enduring, grow strong. Just a line from a (great) video game, but good advice.

I have talked to old men who were on battlefields during that minute.They have told me in one way or another that the sudden silence was the Voice of God. So we still have among us some men who can remember when God spoke clearly to mankind.

It was during that minute in nineteen hundred and eighteen, that millions upon millions of human beings stopped butchering one another.

Vonnegut: When I was a boy, and when Dwayne Hoover was a boy, all the people of all the nations which had fought in the First World War were silent during the eleventh minute of the eleventh hour of Armistice Day, which was the eleventh day of the eleventh month.

Shit.