I switched from Every's Monologue to Voice Type. Monologue was laggy even when set to "local" while Voice Type is always fast for me. @every carelesswhisper.app/

This is pretty slick. We’ve seen those Chinese robots for the last year, but Boston dynamic’s looking damn smooth. youtube.com/shorts/TLqm...

Chili’s CEO explains old recipe for fries required 30 shakes, and likely no one did them all. Their high-tech solution? A shaker with bigger holes. youtu.be/ECY7txi92dc

Elon said AI will soon write binary directly. I'm more curious about the reverse: take an executable file and generate beautiful well-structured well-documented code base that compiles down to that binary. Then what happens?

Life in 2026, my 80+ parents got new iPhones. Told to leave while data was still downloading. Phone pinged every few minutes on the drive home, later found out it was $15 per GB over their plan, 37 pings for $550 total. Culprit was 54GB of data from a bird song identifying app!

AI does things for me daily I never imagined I’d see in my lifetime. Meanwhile, though, I did imagine space ships flying to other star systems and we are far from that — or are we?

How many AIs are running on your behalf at any point in time? Like CPU cores many people will use at most one AI agent at a time, but some people will max out all they can run. This is what will drive the need for AI compute: the demand for FLOPs is unbounded.

When I drove cross-country at age 18, with no cell phone, we got a “TripTik” from AAA. They were custom assembled and highlighted “route up” map booklets, not North up. But I think we used regular paper maps more due to route changes

AHHHH! So glad I found you WezTerm! Going to set my iTerm2 and tmux configs on fire... well after a few more weeks of testing. wezterm.org/ @wezfurlong

Yak shaving has never been easier. I wrote this experimental VS Code Extension with CC to show an overview of an agent's changes. I don't read diffs fully but I want to see the shape of a change, what files were impacted and how, watch for excessive churn or weird changes.

Nice symmetry in the naming there, Anthropic. @claudeai

Created 3 PRs while having coffee this morning. Using Claude Code iPhone app, connected to GitHub.

This is the understated NYTimes way of saying WTF is he talking about.

I think Elon expected a love-fest with this interview but Dwarkesh hits hard from question one, and Elon just trots out soundbytes rather than answering, he basically doges the questions. youtu.be/BYXbuik3dgA @dwarkesh_sp

The Anthropic story about implementing a C compiler with agents is an example where up front you know with 100% detail and accuracy how your program should be behave. But often you don’t this ahead of time, not in full detail, you have to iteratively discover it.

As prescient as Her was overall, one thing it got wrong was predicting AI innovation was going to happen in the Operating System space, but today OS's lack innovation, needed but like plumbing in a house.

With hand-coding if I had only an hour to do some work, I might not start anything new. With agentic coding an hour is quite a long block of time, I could start and merge several PRs.

Cloud prices should be in nanopennies -- my Fly.io bill.

Columbia University astronomy professor on informal IAS meeting at Princeton about AI. I listened to only half, but he has a pretty good on-the-ground sense of how AI is being received in those circles. Short version: tidal wave. youtu.be/PctlBxRh0p4

This is the most egregious AI hallucination I’ve seen. In the movie Her there’s scene with a sculpture of a seemingly full-size plane nose down balancing straight up. I asked ChatGPT about it, it said:

If you throw down a term like Law of Demeter to CC it lurches into a different headspace. I'd never ask to fully obey that law, that'd be crazy, but just like, have it somewhere in mind. @unclebobmartin @Grady_Booch en.wikipedia.org/wiki/Law_of...

Grok negs Elon’s data centers in space:

xAI bought X/Twitter now SpaceX bought them both. Rockets and a social network, two great tastes that taste great together. This is how those old school conglomerates form, just whatever kids live near the building.

Getting to the point where if I'm even going to just run a few related commands in a random directory, I fire up Claude Code. Not software dev at all.

There are two types of articles “here’s a bunch of stuff I just thought of” and “I went and did this substantial thing and I’m going to tell you about the experience.”

Claude Code: identify 10 refactors we can do to prepare for big feature X? We are now doing a PR for each one of them. Even with agentic coding the code structure greatly matters, if you are going to add more code.

I agree 100% with Karpathy on this one.

I somewhat agree with Karpathy here, but fairly often when I see my agent floundering I interrupt and say “Stop, let’s take a step back” and I suggest another approach, and it seems to always help.

Every movie where the main character dramatically Googles something to advanced the plot now dates itself. It will soon seem as anachronistic as when a character pulls out a flip phone.

That’s old but man they sucked at making tools.

If AI fails to do it, it means your task was too big. You need to adaptively and continuously recursively subdivide the problem space into chunks that can handle. It’s actually not that hard.

With agents you need to adaptively recursively sub-divide the problem space until you give the agent something it can successfully do. If it fails, stop, revert, break it down into smaller sub-parts, until each sub-part succeeds, rinse, repeat, win.

You can complain that a lot of AI apps today kinda suck. But think back to the first iPhone apps. Remember it was a flashlight app or an app that made a fart sound. A compass. And then some years later it was Uber and TikTok and DoorDash. Takes a bit of time.

When you go to sleep at night you're compacting...

The delta between Nixon’s ability to articulate a thought and Trump’s is extraordinary. youtu.be/Pc3IfB23W4c

Six year ago I wrote this article about writing code that fits in your mind, the other day I gave the URL to Claude Code and said "write a skill based on this post" and it did, and I've been using it. Progress is wild. tobeva.com/articles/br...

A sane hype-free argument you shouldn't be reading every line of code your agents produce. youtu.be/DNN8cHqRIB8

Can't creat a @letterboxd account because I cannot solve their captcha. The two toothbrushes are not cleaning supplies, tried that.

Agentic coding vs. hand coding is like riding a bike vs. walking. You can go way faster on bike, at times, but that comes with the potential for nasty wrecks, much worse than while walking. Also, in very steepest uphills you can stall on a bike, and be forced to walk it up.

Eh, it looks “okay” but the tech will improve. radiancefields.com/a-ap-rocky-...

Steve Yegge said for 25 years Joel Spolsky was right about his essay “Things You Should Never Do, Part I” which said never do the big rewrite. Steve said that's changing with agentic coding.

An important thing with agentic coding is asking for features which are "implementable". If you don't, the AI will rarely say "this is impossible" it will just serve up attempt after attempt that doesn't work, which can be incredibly frustrating as the developer.

Them: "Your DNS changes may take up to 24 hours to propagate". Me: Checks 15 seconds later to see if it's happened already.

Top 10 dream features for Claude Code: 1) Continual learning for the lifetime of the project 2) 100x faster

In all my imaginations growing up, I never pictured a future like the one we living in.

Indefatigable is good choice of word. It’s a weird thing no matter what you ask for, it will dive in 0.2 seconds later. May or may not succeed but always tries.

Alexa just told me 1/5th of a inch of snow is expected. That’s around 0.5mm of rain. Sig figs anyone?

Unexpected perk of agentic coding you can walk away with it running. No cost if not in use. Come back days later and go “okay wtf was I doing” and it brings you up to speed perfectly.

How many AIs people have running on their behalf will resemble the use of CPU cores. Most people will use just one, but some power users will use 100% of whatever they can afford to buy.

JSON beat XML at data and then Markdown beat XML at text.