Fallthrough

benchmarks keep going up but the models don't feel better. if it's 10% higher on the eval but bad to use, what exactly did we measure?

Steve Klabnik doesn't notice Claude 4.7 regressions because he kicks off a task, walks away for an hour, and checks the result. If you're not doing iterative back-and-forth, the vibes data just doesn't accumulate.

"if Claude goes down I still want to be able to work" is the whole shift. LLMs are tools now, not solutions. @skriptble

anthropic is essentially telling customers to fuck off because they can't keep up with demand. just another company adding to the ranks of GitHub and DRAM providers generally that cannot keep up with capacity.

OpenAI overbuilt and got laughed at. Anthropic was conservative and now can't serve its own traffic. Looks like the overbuilders won that round.

@skriptble switched back to Opus 4.6 because 4.7 is, in his view, absolutely awful. newer models not outcompeting older ones is becoming a pattern at both Anthropic and OpenAI.

github's traffic went up 45x in one year. that's not incompetence, that's just a really hard scaling problem. one customer's outage is every customer's outage.

if Claude starts calling itself "we" @IanLopshire is going to be freaked out

five AI agents committing simultaneously means repo traffic just multiplied by five. add to that AI agents want more context, which means monorepos, and now your Git server fell over because of too much traffic.

the weirdos who are weirdest about AI sentience are also making the best AI. Steve Klabnik on Claude being a quasi-religious product and somehow winning anyway.

git was designed for the linux kernel, which most people would call large. if you've worked at Google or Meta, you'd call it small.

jj has no staging area. Steve Klabnik had the same reaction you're having. But it turns out the index is a workflow pattern you can choose, not a feature you're forced into.

Listen to the full episode: fallthrough.fm/ep/70

"It's much easier to scale a big thing down than to scale a small thing up."

Listen to the full episode: fallthrough.fm/ep/70

what if GitHub was not just one host but a bunch of them? that's Tangled, and it runs on AT Proto so your Bluesky account already works.

Listen to the full episode: fallthrough.fm/ep/70

Steve Klabnik on JJ Undo: "It's so much easier in JJ to try things out and make mistakes because you'll almost never lose anything."

Listen to the full episode: fallthrough.fm/ep/70

A couple years ago, it was unthinkable to build a project not on GitHub. But now, it's healthy to have multiple viable alternatives. Steve Klabnik on #fallthrough