KolenCh

That’s active vs passive transformation and AFAIK mathematicians often think in passive ones but physicists would use both. And the one you quoted is passive.

The most interesting bit to me is how they focus on addressing the scale and reliability issue in the agentic era by adding more resources and rewrite the "primitives on how they built the foundation" as they call it. The numbers are impressive! Let's hope we're going to get a more reliable GitHub soon (they say early next year I think.) The number: 180x faster against old architecture. More at https://blog.kolen.dev/link/github-universe-2026

Interesting that the physics result is not released yet. From the limited wordings, “unified completion of the Standard Model of particle physics” is probably SU(5), which is experimentally disproved as one describing our universe. So the wording sounds a bit confusing, ie they are not proving or discovering Grand Unified Theory (GUT), and there’s no Higgs. So it is not really the standard model, but models that’s closely related to that. It is still an impressive result. IMO at the level of the recent quasi Riemann hypothesis level result, shy of getting the millennium prize (no mass gap) but is actually a very meaningful result, both of which arguably more impactful that the one they solved a millennium prize problem (Navier–Stokes).

More comments in https://blog.kolen.dev/link/openai-shepherds

> refuse to engage

You speak of this as if the default is to engage.

To make an analogy: math is like OSS. You are free to fork and create your own feature, maintainer has no obligation to look at your PR and improve it to a state that’s worth merging. Eg This happened to Zig where the Bun team maintained a fork for a while. And I’m sure there’re countless examples. An instance of a maintainer finding a PR so interesting and keep digesting it and eventually even give a talk about it, is irrelevant. It is entirely their free will to do so.

Somehow I’m expecting to see Bousso’s work here. Eg

https://en.wikipedia.org/wiki/Bousso%27s_holographic_bound

Software engineers have learnt to read the reaction to coding agents as a split between people who love the craft and people who care about the product. Seen through that lens, mathematicians upset about OpenAI’s Navier–Stokes result look like craft-lovers in denial, moving the goalposts. I think that misreads them. “After Math”, by Silvia De Toffoli and Eamon Duede, argues that mathematics was never after the answer: the understanding is the product. Then I extrapolate to research software, where I’d argue the same holds for code that establishes scientific results.

My previous submission covers the Navier–Stokes “incident” itself. Here I want to look at a parallel: AI agents are finding security bugs and producing mathematical proofs faster than the people working in these fields can absorb them.

Can mathematics borrow from coordinated vulnerability disclosure? I’d propose a registry that records work as it develops, with timestamps that others can verify, together with circles that gradually widen before public release. This would give people a record of their contribution without having to rush an unfinished proof out. But it only helps if we also give credit to ideas and partial results.

Have you read the article? It has an unbalanced emdash: `word— another word`. A novice mistake that an LLM likely will not commit nowadays. In standard typography—depending on which guide you need to follow — it can only be padded by space on both side or neither.

I used to distro-hop a lot and in recent years settled with RHEL. The key change over the years is how fragile the supply chain becomes, so while distro like Arch Linux is a great learning experience, I just want to put a layer of maintainers who have higher stake than others between me and the world. (RHEL has a free tier for developers limited to 16 instances.)

How would this compare to shpool: https://github.com/shell-pool/shpool

I think it should be “Claude Fable is relentlessly protective until it isn’t” and pull more on the thread that it “hits a hidden guardrail” and drop into Opus. Both the fact that it knows and deployed such a workaround on a CSS problem and the fact that it is nowhere near cybersecurity/biology/frontier AI dev and triggered the guardrail terrifies me.

rsync regression since 3.4.3

This link (edit: I link to https://github.com/NixOS/nixpkgs/pull/522245#issuecomment-45... but probably the anchor is stripped automatically?) is the clearest that I can find to point out what's the problem in a neutral tone with links to issues you can track yourself.

There is other less civic discussions around the issue that I decide not to cite.

One comment describe the situation aptly: 3.4.1 contains security bugs, later versions fixes it but contains regressions. So choose wisely.

*Apparently* both are AI related: the security bugs are probably discovered by AI. Rushing to fix lots of these results in the use of AI, which leads to regressions.

That’s not what he said. And that part is from the GTD principle, much older than the “second brain” wave. That GTD principle, backed by psychological research, is basically saying that you should extract what you have in your mind immediately when you have it / remember it, so that you don’t need to keep it in your mind all the time. That free your mind to perform at optimal capacity. In psychological lab test, they basically test people doing “remember X and perform this next task Y. I’m going to ask you about X after you finish Y.” The mere fact that you need to remember X degrades your performance.

:)

I experienced both kind of sessions, one kind is like very complicated thing and it finish thousands of lines on its own for quite a while (long horizon problem), another kind is like a seemingly simple task that the agent done in a minute but then I need a few back and forth to get it right easily taking me 30 minutes of my time.

The former kind of experience can make us misjudge how much time we think a task would take us (with agents) to do. And then when the second kind happens, it would be quite disrupting as now we felt like it is delaying our progress.

So tracking the time taken when the second kind happens can help us calibrating what we can expect. I mean if we’re lucky it might take us no time but then we can’t expect being lucky all the time.

Have you tracked your time and confirmed this?

The reason I an ask is, it would felt like a 5 minutes task, but I track my time and found out often time I thought I’d just quickly check the progress made by the agents and it would easily becomes a 10, 15, or even a 30 minutes task.

“Claude” is more specific than “you”. Why rely on attention to figure out who’s the subject? Also it is in their (people from Anthropic) believe that rule based alignment won’t work and that’s why they wrote the soul document as “something like you’d write to your child to show them how they should behave in the world” (I paraphrase). I guess system prompt should be similar in this aspect.

It eliminated some ambiguity. It should be quite self evident that even without an example it is quite impossible to eliminate all ambiguity (it’s a feature of human language.) The more important property is that it never introduces more ambiguity. Ie at worst it doesn’t help, but not making it worse.

This is uniquely interesting to me because humongous amount of data are produced in HEP so automatic research (data analysis) in that domain is I think a huge enabler. It makes trying out ideas much cheaper (in human time) and you could flag interesting results for humans to investigate further. One deal breaker is because big data is so big in HEP, current operations are already very storage and compute bound. If only we can raise trillions of dollars in HEP and perhaps join force with ESA and build a fleet of data centers in space.

My naïve answer to this is, one should never be interested in things because of how useful it might be, but because of the thing itself. Otherwise, AI won’t be the first to make one losing interest. With interest, AI may even make it more addicting. Another distinction to make is, when you use AI, are you taking a shortcut, or channeling it to automate the boring stuffs so that you can explore things you otherwise don’t have time to explore?

I learnt this the hard way: if anyone is sending multiple emails, with seemingly very important titles and messages, and they get no reply at all, the receiver likely haven’t received your email rather than completely ghosting you. Everyone should know this, and at least try a different channel of communication before further actions, especially from those disclosing vulnerability.

Small bugs? May be. But there’s a lot of lack of functionality and stability. I’d recommend InFuse if anyone is hitting those problems. If it has been running fine for you then there’s no need to switch. The problem is related to source codec. Depending on that you’ll have difference experience. So that’s why the experience varies because there’s vast differences in source formats. A good client not only handles well on some sources, but many if not all.

A.k.a. How do you deal with a time dependent network?

Nice illustration.

To be a bit picky, there’s no unprocessed photo. They start with a minimally processed photo and take it from there.

The reason I clicked is that when I saw the title, I’m tempted to think they might be referring to analog photo (ie film). In that case I think there’s a well defined concept of “unprocessed” as it is a physical object.

For digital photo, you require at least a rescaling to turn it to grayscale as the author did. But even that, the values your monitor shows already is not linear. And I’m not sure pedagogically it should be started with that, as the authors mention later about the Bayer pattern. Shouldn’t “unprocessed” come with the color information? Because if you start from gray scale, the color information seems to be added from the processing itself (ie you’re not gradually adding only processing to your “unprocessed” photo).

To be fair, representing “unprocessed” Bayer pattern is much harder as the color filter does not nicely maps to RGB. If I were to do it I might just map the sensor RGB to just RGB (with default color space sRGB) and make a footnote there.

I think there’s a spectrum and you said it as if there’s only two sides.

For me personally, I built my “data centre” as cheap as possible, but there’s a few requirements that the computers you’re using would not cut it: storage server must be using ZFS with ECC. I started this around a decade ago and I only spent ~$300 at the time (reusing old PSU and case I think).

There are many requirements of a data centre that can be relaxed in a home lab settings, up time, performance, etc. but I would never trade data integrity for tiny bit of savings. Sadly this is a criteria that many, including some of those building very sophisticated home cluster, didn’t set as a priority.

It only works on Mac and windows as a VM.

Nix is for reproducibility. Nix and docker are orthogonal. You can create reproducible docker image via nix. You can run nix inside docker on systems that doesn’t allow you to create the nix store.

Some people take Moore’s law in a strong sense: doubling rate is a constant. That is long dead.

But if we relax it to be a slowly varying constant, then it is not dead. That constant has been changed (by consensus) for a few times already.

Your mistake is to (1) take that constant literally (ie using the strong law) and (2) uses the boundary points to find the “average” effect. The latter is a really flawed argument as it cannot prove it hasn’t been dead (a recent effect) because you haven’t considered it’s change over time.

Original title is “What Are Lie Groups?”.

This article is the shallowest I have read from quanta magazine. I expected more, give there articles in mathematics.

https://freakonomics.com/podcast/can-you-change-your-persona... People have different modes of personality so to speak. People behaves differently with a different crowds all the time.

This is equivalent to inverse variance weighting. For independent random variable, this is the optimal method to combine multiple measurements. He just used a different way to write the formula and connect that to other kinds of functions.

He also frames it as a different goal too: normally when we (as a physicist) talks about the random variables to combine, we think of it as different measurements of the same thing. But he didn’t even assume that: he’s saying if you want to have a weighted sum of random variables, not necessarily expected to be a measurement of the same thing (eg share same mean), this is still the optimal solution if all care is minimal variance. His example is stock, where if all you care is your “index” being less volatile, inverse variance weighting is also optimal.

As I’m not a finance person, this is new to me (the math is exactly the same, just different conceptually in what you think the X_i s are).

I wish he mention inverse variance weighting just to draw the connection though. Many comments here would be unnecessary if he did.

I’ve been using AirPods in macOS in Microsoft Teams call all the time.