David Crawshaw

building exe.dev
co-founder of tailscale
forever a programmer

blog: https://crawshaw.io
work: https://exe.dev
github: https://github.com/crawshaw
location: Berkeley, CA

Public Key

npub1zlv43zfyah2arh89x0d67gh7xlep4lm9u6ggvw2sje6c95aravtq0hlytv

NIP-05 Address

[email protected]

Profile Code

nprofile1qqsp0k2c3yjwm4w3mnjn8ka0ytlr0us6laj7dyyx89gfvavz6w37k9sprpmhxue69uhhyetvv9ujumt0d4hhxarj9ecxjmntqy28wumn8ghj7un9d3shjtnyv9kh2uewd9hscyhmr6

Publishing to

relay.momostr.pink

relay.damus.io

offchain.pub

Author Public Key

npub1zlv43zfyah2arh89x0d67gh7xlep4lm9u6ggvw2sje6c95aravtq0hlytv

Show more details

Last Notes

2026-06-02 19:36:29 GMT

exe.dev VMs are served off local NVMe but by default regularly replicate their disk off machine. We are thinking of introducing a VM variant that doesn’t replicate (cheaper, also maybe faster start for odd reasons). What do you call VMs like that? About as durable as a classic VPS.

2026-04-12 00:57:59 GMT

Growing up in Australia, I spent a lot of time using a shell on US computers. Watching your keystrokes commute hundreds of milliseconds is leaves a lasting impression. So I am very happy to be rolling out exe.dev regions: keep your compute close.
https://cdn.masto.host/inuhnet/media_attachments/files/116/389/016/900/922/119/original/d1fddb809961deba.png

2026-03-18 20:53:56 GMT

- reply

getting this right is key to rolling out region support. right now we have users testing out nyc, tokyo, london, frankfurt, and sydney computers.

2026-03-18 20:53:05 GMT

exe.dev is behind a new global load balancer (of our own making) and anycast network. qualification testing shows good routing from lots of locations. if you have a computer with us I would very much like to hear if your connection is good!

2026-03-06 20:23:21 GMT

There is an HTTPS API for exe.dev now, and I am really happy with how it turned out. Instead of dozens of endpoints and a mountain of docs, you send the SSH lobby commands. https://blog.exe.dev/apis-for-the-restless
https://cdn.masto.host/inuhnet/media_attachments/files/116/184/092/065/044/633/original/1008e2fbf1067c8c.png

2026-02-07 19:39:16 GMT

https://crawshaw.io/blog/eight-more-months-of-agents

2026-02-03 17:44:37 GMT

Some careful thinking about LLM API pricing by Philip: https://blog.exe.dev/expensively-quadratic

2026-01-31 20:16:51 GMT

- reply

yeah I understand how any feature can be useful. It can also cause an outage, as it did yesterday.

2026-01-31 17:37:44 GMT

- reply

(The only init system I liked was the old BSDs, where /etc/rc was a single shell script that executed at boot. Yes that is very limiting, but it was also enough for me. I can reboot.)

2026-01-31 17:37:00 GMT

I have zero interest in hating systemd, what came before it was awful. But I do wish systemd did not keep biting me with features.
I just learned that you can have both a foo.service file *and* a foo.service.d directory, with extra parameters in the directory that breaks everything. (Hidden from tab complete by the file.)
You are apparently meant to change the way you open your text editor and use `systemctl edit service` to combine the files.

2026-01-26 14:26:01 GMT

- reply

I find this is one of the best uses of the technology, because the complexity of libraries for APIs comes from the intersection of all the lessor used features of the API. So if I describe what features I want, I get a much smaller library than one that was written generically against the API. The smallness makes it easy to read, a huge win.

2026-01-23 00:15:30 GMT

SSH has no Host header equivalent. Here’s how we made one, with some help from DNS: https://blog.exe.dev/ssh-host-header

2026-01-02 22:47:29 GMT

- reply

software engineer salaries are higher than ever, how does that fit with losing negotiating power?

2025-12-25 16:15:43 GMT

- reply

I’m referencing https://en.wikipedia.org/wiki/Copenhagen_interpretation#
What I am getting at is discussions around words like intelligence do not seem productive. We can try stuff, and some of it works. That is all we really seem to know.
https://cdn.masto.host/inuhnet/media_attachments/files/115/781/084/025/245/859/original/0bc1e8595241755a.jpeg

2025-12-25 14:22:08 GMT

- reply

I find many articles on the topic end up in a wishy washy space that reminds me of people discussing interpretations of quantum mechanics. The physicist school of “shut up and calculate” is what I have found the most useful. (I too get a lot of value out of these things.)

2025-09-30 00:30:19 GMT

- reply

and a modern 4K display has 25MB of pixels. Feeding it at 60Hz needs ~1.5GB/s. Just incredible how far hardware has come.

2025-09-21 22:44:43 GMT

- reply

solving a smaller set of problems is a powerful tradeoff, it lets you focus on other things.

2025-09-21 22:43:15 GMT

- reply

I suspect I could replace many of my remaining uses of C/C++ with Rust, but old dog, etc. I will the day the big dependency I need is written in rust.

2025-09-21 22:42:07 GMT

- reply

I think a lot of people programming in Go broadly agree. I use it because it’s the best tool to solve many problems quickly. Definitely not all problems. And Go libraries set out to solve many, but not all problems. I still reach for C (and C++) sometimes.

2025-06-10 23:38:41 GMT

I suspect this is not interesting to the crowd here, but if you are, here are some of my experiences using agents over the past six months: https://crawshaw.io/blog/programming-with-agents

2025-06-07 19:45:36 GMT

May we all strive to bring wonderous change like Bill Atkinson.

2025-02-14 22:31:45 GMT

systemd question: I have a service that does a lot of cgroups things. I would like systemd to stay far away from its settings, so using iptables and other subprocess namespacing is the same as running the service as root under screen.
I found delegate=yes, which helps. But there’s still some iptables oddity under systemd. What else should I look at?

2025-01-29 17:21:36 GMT

We set up a Sketch blog for our research into having LLMs write programs. Our latest project is a merge conflict resolver. https://sketch.dev/blog/merde

2024-10-16 17:31:11 GMT

- reply

my concern is that use of unsafe says the core features of the language are not good enough to write a logging package.
This used to be a principle of standard library development: https://groups.google.com/g/golang-dev/c/nslTL7R3xjI/m/V6t-HOu4ieQJ

2024-10-15 22:46:12 GMT

I kind of want to buy an ipad and apple pencil just for math notes. Why is this a bad idea?

2024-10-15 16:51:26 GMT

Engineers having fun
https://cdn.masto.host/inuhnet/media_attachments/files/113/312/470/086/970/001/original/7d1b1b0bc3e5ec6a.png

2024-10-15 16:39:23 GMT

I think it’s weird that the log/slog package imports unsafe.

2024-10-07 22:04:32 GMT

- reply

agreed, it needs some commercial space to bring it to life. The tiny ferry from sf is fast and fun, maybe there’s an opportunity to build things that are too expensive for SF proper.

2024-10-07 20:09:50 GMT

- reply

about 1/3rd of this now exists.
https://cdn.masto.host/inuhnet/media_attachments/files/113/267/949/505/647/406/original/a74ca6ab7922f564.jpeg

2024-10-07 20:06:38 GMT

- reply

have you been out there recently? It’s now mostly a construction site for ~10 story apartment buildings and is trying to go up market. (It actually looks pretty good.)

2024-10-04 17:54:13 GMT

- reply

The avo package is wonderful because it makes it easy to decompose gnarly AVX instruction runs into testable blocks. https://github.com/mmcloughlin/avo/

2024-10-04 17:37:38 GMT

Hacking on one of those fun projects where yesterday I was trying to post JSON from JavaScript, and today I'm summing float32 values in AVX-512 assembly.

2024-09-16 23:21:57 GMT

- reply

the ferry from Treasure Island to SF takes about 5 min and is small and friendly.
I was by the Ferry Building a few weeks ago with my bike, saw it and found myself on an impromptu Bay Bridge ride back to Berkeley.

2024-09-16 19:26:01 GMT

Following where httpts leads me, here's the classic python one-liner:
python -m http.server 8000
but for your tailnet:
go install github.com/crawshaw/httpts/cmd/tailweb@latest
tailweb
instead of your LAN it serves to your tailnet with your colleagues.
https://github.com/crawshaw/httpts/blob/main/cmd/tailweb/tailweb.go

2024-09-16 18:36:32 GMT

- reply

in the go.mod or in the library itself? I agree the go.mod is heavy but the library is a bit better.

2024-09-16 17:49:20 GMT

I use tsnet all the time to serve out internal dashboards/tools/etc on a tailnet. It's a bit of work hooking it up to an HTTP server, so I'm experimenting with a new API. https://github.com/crawshaw/httpts
https://cdn.masto.host/inuhnet/media_attachments/files/113/148/490/116/137/117/original/1767498c4da4c374.jpg

2024-09-15 16:37:17 GMT

- reply

Then I get on here, or HN, and read long, often angry comments about how these tools are useless. Meanwhile, I get a lot of value out of them every day.
I have put more time into writing these two little message than getting an LLM to build me an entire phone UI, responding to modern touch events. The strong difference of opinion on these tools among programmers is unfathomable to me.

2024-09-15 16:34:51 GMT

This morning I set out to build a web UI for a tiny internal tool to be used on a phone. I thought Claude could help, as I don't know how to detect touch/swipe events in the browser.
250 words later, Claude had built all the UI for me. (200 of those in the first instruction, then two followup sentences for UI polish) I gave it two dummy placeholder JS functions that load/store the data, it did the rest.

2024-09-11 17:44:54 GMT

- reply

https://cdn.masto.host/inuhnet/media_attachments/files/113/120/162/019/435/576/original/57769fcf8f596639.png

2024-09-05 15:40:05 GMT

- reply

I really enjoy these made up universes. They are never as simple as nonsense, but are legitimate alternate realities. In some system the heap allocator really is swappable.
The name probably comes from lua_setallocf.
And unixnanos is monotonically increasing, so you *could* imagine a VMM fulfilling those pages. It's like reading alternate-past fiction.

2024-09-05 15:32:56 GMT

Gemini taking hallucinations to the next level. There most definitely isn't a debug.SetAllocf. But even if there was, imagine the fun times from a heap implementation that picked memory locations using cur_time_ns.
(Other LLMs suggest the boring testing.AllocsPerRun.)
https://cdn.masto.host/inuhnet/media_attachments/files/113/085/667/958/838/252/original/4dc8089cab7c6ed3.png

2024-09-04 18:16:43 GMT

I like it when others maintain code and don’t usually rewrite things. But surprisingly, I don’t think I’ve ever used a cgo-based wrapper without rewriting it. Not sure what this means.

2024-09-03 16:18:39 GMT

- reply

*build new terminal*

2024-08-22 21:50:44 GMT

Hard to beat the feeling: waking up with a simple heuristic in mind that replaces pages of messy heuristics, and on trying it all the painstaking tests written against the old code pass.

2024-08-12 14:29:00 GMT

- reply

oss supply chain is like a game of go fish

2024-08-02 01:35:50 GMT

There are surprisingly different code paths in Go for
b, err := os.ReadFile(path)
s := string(b)
and
var sb strings.Builder
f, err := os.Open(path)
_, err := f.WriteTo(&sb)
s := sb.String()
Fun to read through. Guess which is more efficient?

2024-07-29 07:02:10 GMT

- reply

oh implicitly in a compiler! I've done that! It's not type theory there. Look for "dataflow analysis" using "intervals" or "value ranges"

2024-07-29 06:38:31 GMT

- reply

subtyping ranges can be done with refinement types, which are not easy but not Very Hard. Nigel does it in Wuffs: https://github.com/google/wuffs/blob/main/doc/glossary.md#refinement-type

2024-06-10 14:20:45 GMT

Monitor recommendations for Ubuntu? I tried using my Apple Studio Display, which worked gloriously for about a week then broke after an Ubuntu software update. :/ (both DP->TB3 cable and USB4 pass through doesn’t work).
I have some forgettable 4K monitor that works but it’s nowhere near as nice as 5120x2880 at 200% scale. Anything else nice out there that works with Ubuntu?