Hacker News: jdub

New comment by jdub in "An OpenAI model has disproved a central conjecture in discrete geometry"

jdub — Thu, 21 May 2026 13:51:16 +0000

Reinforcement learning perturbs the model such that the token prediction process (inference) tends towards the desired result.

New comment by jdub in "Reviving old scanners with an in-browser Linux VM bridged to WebUSB over USB/IP"

jdub — Thu, 21 May 2026 03:52:46 +0000

Ah, it seems like the architecture was designed by a slop machine. OK.

New comment by jdub in "Reviving old scanners with an in-browser Linux VM bridged to WebUSB over USB/IP"

jdub — Thu, 21 May 2026 03:40:23 +0000

Oh, and:

> I must apologise that I haven’t so far open-sourced any part of this that I don’t have to.

With some blather about commercial opportunities. Which is a weird thing to say without linking to the bits that must be shared (under the terms of the various licenses).

New comment by jdub in "Reviving old scanners with an in-browser Linux VM bridged to WebUSB over USB/IP"

jdub — Thu, 21 May 2026 03:35:33 +0000

Hrm, yes-we-scan and printervention are built on SANE and CUPS respectively, which makes sense. But running them in a whole wasm-emulated Linux kernel and userland seems... like a lot.

New comment by jdub in "Goodbye Visa and Mastercard: 130M Europeans switching to sovereign payment"

jdub — Thu, 21 May 2026 01:08:02 +0000

> Giving someone in government the ability to block someone's payments and trusting they won't abuse it might be fine as long as good people remain in power, but do you really want to bet the entire nation's ability to live life on that?

Banking and finance companies honour foreign government sanctions. Ask Francesca Albanese.

Libertarian comparisons of government and non-government behaviour always devolves into angel counting.

New comment by jdub in "The greatest shot in television: James Burke had one chance to nail this scene (2024)"

jdub — Mon, 11 May 2026 09:13:53 +0000

I first saw Connections in the late 2000s... the final scene of the last episode, "inside the British Airways computer" (an entire floor of a large building), had me standing on my couch pointing at the screen.

A year or two before I was born, James Burke wandered between mainframes and reel-to-reel tape machines, speaking with extraordinary prescience about data, communications, decision-making systems, and power:

"This is the future. Because if you tell a computer everything you know about something, it will juggle the mix, and come up with a prediction. Do this, and you'll get that. And if you have information and a computer, you too can look into the future. And that is power. Commercial power, political power, power to change things."

I'm going to watch that scene again, because it's even more important 20 years further along: smart phones, "big data", large language models, Palantir...

New comment by jdub in "Local AI needs to be the norm"

jdub — Mon, 11 May 2026 01:21:55 +0000

> Most people won’t settle for a product if it’s not the very best and incredibly convenient.

... uh?

New comment by jdub in "Over 97% of the 'Linux' Foundation's Budget Goes Not to Linux"

jdub — Sun, 10 May 2026 02:52:57 +0000

Seriously, you'd do better to engage with ideology and philosophy rather than the personal letters of flawed human beings.

New comment by jdub in "A recent experience with ChatGPT 5.5 Pro"

jdub — Sun, 10 May 2026 02:49:20 +0000

It's not a gotcha. It's incredible what these things can do despite being next token predictors from a weird dataset. That's at the heart of the "bitter lesson", and you don't have to believe in magic to see it.

New comment by jdub in "A recent experience with ChatGPT 5.5 Pro"

jdub — Sat, 09 May 2026 12:21:17 +0000

Reinforcement learning for "reasoning" perturbs the model to generate completions in a particular chain of thought / alternative selection structure. It's three next token predictors in a trench coat.

New comment by jdub in "Over 97% of the 'Linux' Foundation's Budget Goes Not to Linux"

jdub — Sat, 09 May 2026 11:25:06 +0000

Ha, if you think Marx is objectionable then you really need to read Das Kapital.

New comment by jdub in "Over 97% of the 'Linux' Foundation's Budget Goes Not to Linux"

jdub — Sat, 09 May 2026 08:51:18 +0000

That terrible analogy does not produce a useful mental model on any level. You probably need to read Das Kapital.

New comment by jdub in "Over 97% of the 'Linux' Foundation's Budget Goes Not to Linux"

jdub — Sat, 09 May 2026 08:47:19 +0000

I've had occasional concerns about the Linux Foundation and how it operates, but there's no question it has been a transformative contribution to Open Source.

A bunch of folks decided to get off their butts and gather donations to support Linux... and then it snowballed. Cool. The creators and members get to decide how they contribute, and projects get to decide if they want to participate. There are alternatives for projects that need to "raise and spend", and some are 501(c)(3).

(Also keep in mind that techrights.org has been an unhinged shit sheet attacking individuals and companies for insufficient purity for decades now.)

New comment by jdub in "You gave me a u32. I gave you root. (io_uring ZCRX freelist LPE)"

jdub — Sat, 09 May 2026 08:07:46 +0000

... there's also a bit of a frequency illusion factor.

New comment by jdub in "Over 97% of the 'Linux' Foundation's Budget Goes Not to Linux"

jdub — Sat, 09 May 2026 07:47:40 +0000

Opponents of obscene wealth/income inequality are typically not motivated by envy – that is your own projection.

New comment by jdub in "A Gopher Meets a Crab"

jdub — Sat, 02 May 2026 12:22:38 +0000

Nah, it's an awful way to learn. Especially to learn to be good or great.

When you start reading, it helps to have some guidance towards good and relevant books, from e.g. school, mentors, criticism, etc. Then, when you encounter a "bad" book, you have some benchmarks from which you can build your capacity for analysis and critique. (Testing your analysis and critique with others helps, too.)

If you start with "bad" books, your concept of quality and what's possible is constrained. (Like when teenage boys read Atlas Shrugged.)

Reading slop code is a terrible way to build a mental benchmark for what's good, what's possible, what's elegant, and writing good code that is respectful to your fellow human beings.

New comment by jdub in "How fast is a macOS VM, and how small could it be?"

jdub — Sat, 02 May 2026 11:39:21 +0000

If guest memory can be reclaimed, it doesn't need to be paged to disk once you hit RAM contention. It's mostly saving accounting overhead, but it'll have some effect on latency, which you're more likely to perceive under contention.

New comment by jdub in "Amateur armed with ChatGPT solves an Erdős problem"

jdub — Sun, 26 Apr 2026 15:16:08 +0000

The dimension of brute force unfamiliar to human brains is "well-read with zero judgement", where connections can be made even if they're not thought through.

Grinding through completions isn't reasoning.

New comment by jdub in "Google plans to invest up to $40B in Anthropic"

jdub — Sun, 26 Apr 2026 05:55:25 +0000

You're hearing "don't invent fire" but what's being said is "for fuck's sake, stop lighting fires in the cave".

New comment by jdub in "Amateur armed with ChatGPT solves an Erdős problem"

jdub — Sun, 26 Apr 2026 05:30:24 +0000

Less reasoning than a dimension of brute force unfamiliar to human brains.