adtac

LOC is the worst metric for engineering output, except for all the others - Churchill

camcorders have been around for ages but you don't have millions of directors around just because you make a tool more accessible

tiktok alone has 1.5 million directors! it's just that we call them creators now.

the meaning of the word director has changed, that's all. but professional roles shift in meaning all the time. a computer used to be a literal human doing arithmetic. an engineer was someone who designed war machinery. being a doctor used to mean teaching at an university.

human beings are natural tool makers. we always have been. the frontier material to manipulate where the most advanced engineering happens constantly changes: stone -> bronze -> iron -> ink (descartes) -> steel -> silicon -> javascript (YOU ARE HERE).

notice how each step is an improvement/abstraction on top of the steps that came before it. some say english is next in that chain. i honestly have no idea. all i know is there will always be The Next Thing and it'll be much nicer to work with.

i got to try exe a while back and i have to say, the "Login with exe" [1] is probably the most magic thing i've seen since tailscale :)

[1] https://exe.dev/docs/login-with-exe

ed is the standard Unix text editor.

December 2018 was the last version to not have "Development of safe artificial intelligence to benefit humanity"

https://web.archive.org/web/20181223120124/http://www.muskfo...

too many of these vaguely look like what galaxies look like from earth

e.g. https://i.imgur.com/ZjiBF8f.png

just a coincidence?

weekend hackernews best hackernews

if you haven't read the factorio devblog before, please do!

after HN, it's one of my favourite places on the internet because i constantly learn new, random, insane things almost every time. imho it teaches you how to think + shows you what great engineering taste looks like. sorry if i'm overly effusive but each post is so deeply technical and well-written that i can't believe it's free.

you don't need to know anything about factorio or gamedev btw (i don't), just pick a random number between 1 to 438 and start reading :)

or endeavor to

you picked the worst example company to complain about how they're are not trying lol. just in 2025 from anthropic:

Circuit Tracing: Revealing Computational Graphs in Language Models https://transformer-circuits.pub/2025/attribution-graphs/met...

On the Biology of a Large Language Model https://transformer-circuits.pub/2025/attribution-graphs/bio...

Progress on Attention https://transformer-circuits.pub/2025/attention-update/index...

A Toy Model of Interference Weights https://transformer-circuits.pub/2025/interference-weights/i...

Open-sourcing circuit tracing tools https://www.anthropic.com/research/open-source-circuit-traci...

Note that there isn’t the slightest attempt to explain the planet trajectories (specifically, why the planets keep ending up where they do regardless of how many epicycles you bolt on) from a theoretical perspective. My impression is that they have absolutely no idea why the heavens behave the way they do; all they can do is stare at the night sky, record, and see what happens. That is not reassuring to me at least.

- AstronomerNews user, circa 1650 (probably)

Why not something dumb like this: https://chatgpt.com/share/68b60199-b6ac-8009-b50d-3e7cfff1d7... (gpt-4o)

wow not even yin-yang can escape entropy or the heat death of the universe

That's basically your documents lossily encoded.

Vector embeddings are lossy encodings of documents roughly in the same way a SHA256 hash is a lossy encoding. It's virtually impossible to reverse the embedding vector to recover the original document.

Note: when vectors are combined with other components for search and retrieval, it's trivial to end up with a horribly insecure system, but just vector embeddings are useful by themselves and you said "all useful AI retrieval systems are insecure by design", so I felt it necessary to disagree with that part.

if you feel cheated now, I promise you'll feel more cheated if they did this after you rely on it for several months. the least worst option is to change pricing as early as possible.

this is just XMPP and it has been around for years before Slack was even a thing

Slack won anyway for the same reasons most centralised, commercial, closed sourced products often win: less fragmentation, more marketing, stronger network effects, simpler onboarding for normies, richer integrations, and most importantly, an enterprise sales team that actually showers

clicked hoping for an out of distribution argument for why a specific type of Japanese martial arts is perfect for git users :(

good article nonetheless!

1. modern computers are insanely fast

2. they compiled a version of the kernel with as many configurable options disabled as possible

3. modern computers are insanely fast

1. PG is implicitly talking about good essays. And it's very hard to write a good essay if you don't believe it's true.

2. Whether something is true has nothing to do with whether the reader knows it. I know many false things and ideas.

this is like saying "same way cells know how to be anything - quantum electrodynamics"

Our models will be available under an Apache 2.0 license.

^ from the post

https://github.com/SesameAILabs/csm is empty for now, but I imagine they'll be releasing it soon: https://x.com/_apkumar/status/1895492615220707723

Not in the very short term, but we've talked about it internally and we might do it in the future!

If you'd like to use Subtrace on Windows, it would be super helpful to understand your use-case deeply so that we build the right things in the right order. Please reach out to me at adtac@subtrace.dev, I'd love to chat!

Flat pricing with unlimited requests hosted in the cloud (https://subtrace.dev/pricing).

We also have an on-prem version of Subtrace for enterprises. It runs in their own AWS account without ever talking to subtrace.dev so that companies in regulated industries like healthcare can use Subtrace.

What would you like to see?

One of things we're thinking about is automatic method/function call tracing. Something like attaching the entire stack trace of calls done to handle the API request. Ideally using the same UI so that you can see the headers/payload that was sent and the function-level stack trace right next to each other. None of the OpenTelemetry verbosity, all of the observability!

One of my litmus tests for software is "can I use it on a flight?", so you'll be pleased to know Subtrace works great locally without ever talking to subtrace.dev!

    subtrace run -devtools=/subtrace -- python3 -m http.server

re self-hosting: yep! Use the -devtools flag to get a stripped down version of Subtrace running entirely locally:

    subtrace run -devtools=/subtrace -- python3 -m http.server

in a different context, maybe: https://news.ycombinator.com/item?id=38044786 :)

You can use Subtrace entirely locally without talking to subtrace.dev using the -devtools flag. For example:

    subtrace run -devtools=/subtrace -- python3 -m http.server

We have fully on-prem installations for enterprises, of course (e.g. if you're in a highly regulated industry like healthcare).

Thanks!

re the Wireshark analogy: the reason I used that was because: (1) Subtrace operates at roughly the same level in the operating system stack, (2) has similar capabilities, (3) has an overlap in use-cases, and (4) has been the most effective at communicating what Subtrace is in my experience so far. I can see why the analogy is not a perfect 1:1 mapping (obligatory xkcd: https://xkcd.com/624), but naming things is hard and taglines are just names in idea space :)

You can tag each request with arbitrary key-value maps and search over these later. For example, if you add a `x-subtrace-tags: user=foo, project=bar` header on the response, you can apply a `tags.user == "foo"` filter in the dashboard to see all requests across your entire infra from that user and only that user. Each request is pre-populated by default with tags like hostname, pod name, AWS/GCP location, etc.

It's like Honeycomb's wide events but even better because: (1) you can see whole request including the payload alongside the event fields, and (2) it's fully automatic and requires no code changes out of the box (you can incrementally add these tags when you find a need for each one instead of the huge upfront cost from instrumenting the hell out of your codebase).

use subtrace in an ephemeral pod for debugging purposes

That's a great suggestion. It'd be like kubectl exec-ing into a shell inside the pod, but for network activity. I think I'm going to prototype this tonight :)

pod filtering such as namespace and label selectors

Yep, Subtrace already tags each request with a bunch of metadata about the place where it originated so that you can filter on those in the dashboard :) Things like the hostname, pod, cluster, AWS/GCP location are automatically populated, but you can also set custom tags in the config [1].

[1] https://docs.subtrace.dev/tags