That should have been a LEC-RUS-HAM podium. Usual strategy fumble from Ferrari. Should have pitted during VSC. Oh well. On a positive note, @F1 on @AppleTV looks absolutely gorgeous. Probably the first ever instance of a live sporting event in Dolby Vision? #AusGP

Good move. The naming is pretty clear now. Agent tool -> spawns a new agent Task tool -> manages tasks Next, please fix the naming mess wrt “memory” @trq212 @bcherny

@amasad At scale, maintaining not writing code is the hard part. Right now we rely on humans to develop expertise and maintain accountability over code. In this fluid, AI-collaborative world, who maintains the code?

It’s super confusing that Anthropic calls CLAUDE dot md files “memory” files. May be call them “instructions” from now on? That aside, auto memory is awesome!

This is.. not great, and in direct contrast with @trq212’s post from a few weeks ago.

Apple podcasts: https://t.co/eGLOCF7Md2 Overcast:

We no longer have to write the software by hand. We write the test that we expect to pass and the computer generates. Well actually we don't write the test by hand. We write the spec and the computer generates the tests from that. Well actually we don't write the spec by hand. We

Turns out with claude code, my decades long strategy of NOT deeply learning: - regexs - sql - nginx confs - elaborate shell commands - advanced shell scripting - any javascript framework - perf optimization - webpack, cdns, bundlers - 1000 other things ...was entirely correct.

Also, we tried video for first time, give it a watch and let us know if you want this content on youtube: Part 1: https://t.co/BCj8HbPFvV Part 2:

AI tsunami coastline ni first hit chestundi — and that's us. From ChatGPT to Claude Code, @dhruvtv and @gummadi break down what agentic AI means for software engineers. This one's not comfortable. https://t.co/6bKkxaaO2y

Opus 4.6 now takes the crown as the best long-context model. 🫡 As measured on MRCR v2 8-needle at 256k and 1M, and it beats GPT-5.2 on the 256k range. MRCR v2 (8-needle) is a long-context retrieval benchmark that checks whether a model can find and reproduce 8 specific facts

RIP Opus 4.5. You’ll be remembered forever. Like the OG iPhone.

Yet another Karpathy banger. I found myself nodding my head vigorously, as you would watching a rockstar perform. While pausing a few times to hit ‘Yes’ on Claude Code as it ran a deep analysis on 1000+ files, breaking down into Tasks, writing Python and bash scripts as usual.

If an agent can do 95% of a job: - a human only needs to do 5% of it (high-value? yes but still 5%) - frees them up to do 20x more; at worse 5x more - if work to be done is constant, you need 1/5th of current workforce @demishassabis just proved @DarioAmodei’s point.

I love how “tech” has become a euphemism for plastic clothing. I don’t get what tech pants are.

Okay San Ramon, enough of these for the day. Let’s chill now.

> Developer output up 76%, Team output up 89% Interested to see this data normalized by engineering hours put in. Anecdotally, I've seen me and everyone around me put in at least 1.5x more hours this year for various reasons. So, what percent can we attribute to gen AI alone?

live life

context engineering, succinctly put

I don’t like git work trees and also how @cursor_ai handles them. - My `git branch` output is a mess. - I use Cursor’s multi-agent UI for throwaway queries but the resultant work trees hang around for days. - Even after they are deleted, branches still remain.