catlifeonmars

Probably not for lack of trying though. The incentives for corporate espionage are strong.

Are these novel coding tasks, or could they have ended up in the training set?

Does the number/frequency of safeguard stops affect the measured speed?

> Also I think there's huge differences to the kinds of tasks and workloads people are throwing at it, leading to wildly different perspectives, but everyone's talking about the same "Fable".

This right here. Humans are really prone to over indexing on a handful of personal experiences vs a representative sample. And once someone has formed an opinion, confirmation bias helps cement it, despite evidence to the contrary. Reality is probably somewhere in the middle.

> it’s quite rational that demand increases as price decreases

This is not necessarily wrong, but neither is it right; it is underspecified. Even if we were ignoring other factors and taking into account only the assumptions of a classical economic model, the shape of the demand curve still depends on the size of the market and the price elasticity of demand.

There’s still a cost to underestimating asset depreciation from an accounting perspective. Imagine being left holding the balance sheet when you budgeted for X amount of depreciation but the demand drops significantly more than you expect during the next fiscal year. I guess my point is that artificially constrained supply is not a free lunch.

> Do you really think banks would be any better?

Absolutely, yes.

If you’re talking about OpenAI and Anthropic that’s an extremely low bar.

May I ask what in particular gives you confidence that these two companies know what they are doing wrt cybersecurity.

This feels very suspiciously like a finite field problem.

Laser printers were a side project that almost got the inventor fired.

That’s fascinating. I’ve seen this sort of thing on protocol implementations in layer 4+ but it’s rare to find it in layer 3.

GP said

> They ran it without a jailer

And you said

> getting code execution behind the jailer is still effectively game-over

So I’m really confused, which is it?

Additional reply bc I’ve passed the window for my other comment.

I like the design choices you’ve made with the syntax.

My advice on the README: when it comes to prose, more is not better. You should say what you mean in exactly as many words as is necessary and not more. LLMs like to add extra grammatical components that detract from the meaning, which translates into the reader needing to do extra work to determine which phrases or choice of words are significant.

Most people just want JSON with comments and the ability to edit long strings, which I think is legitimate. You can tell by how many “json with comments” variants are out there in use: JSONC, JSON5, and HuJSON come to mind. Someone should probably standardize it, so I can add a fourth variant to the list :D

Edit: obligatory XKCD comic: https://xkcd.com/927/

Respectfully, the fact that you copypasted LLM output here to respond does not instill much confidence that you know what you are doing.

TOML is probably my go-to these days.

I think where TOML shines is when your configuration schema is relatively flat and simple. If you want to heavily nest configuration it can quickly get difficult to read. I’d love a more limited version of TOML that restricts the nesting depth.

There are numerous inaccuracies in the “Why not TOML” section of the README. Here are a few:

> CVE-2026-44244 injected a [core] section into .git/config for code execution.

Git config is not TOML. It’s an ini-based bespoke syntax

> SOML needs quotes for a literal dot.

So does TOML, the dot notation is for nested keys. You use quotes for a literal dot.

——

FWIW I like the decision to have a first class duration type.

Welcome to VoLTE :)

This is a good way to injure your snake, it definitely depends on the species and the individual temperament of the snake, but there are ways to entice a snake to eat dead mice (you can warm them, and wiggle them).

It helps to understand how snakes detect and catch prey.

> The mouse survived!

Without more context, I’m not sure if that’s better or worse

That’s not quite equivalent. For one, a responsible owner is not feeding live mice to a snake. There’s at least an opportunity to ethically euthanize the mice prior to the act.

> At least the human has the mouse’s wellbeing at heart.

You don’t know this to be true

I do a similar thing with ping payload sizes to be able to fingerprint the traffic: I’ll use prime numbered packet sizes so that they can be detected even when the packet sizes are summed.

Btw, one of those numbers is not like the others (has more prime factors that evenly divide 60).

> and a little bump every minute if your metrics have enough resolution

Amusingly, you can also observe periodic bumps if your metrics don’t have enough resolution, due to aliasing effects dominating signal (guestures vaguely at Nyquist frequency).

> I actually don’t know how to do that and stay under a bounded edit distance.

I’m sure there’s an interesting result there, but I am not the person to find it :D

Good episode btw

You are not wrong about the complexity of modern stacks.

However, VoLTE predates k8s adoption in most telcos and the physical network appliance == NF model is still alive and well, so it’s not necessarily a fair comparison. You’re also right about the added complexity of SIP/RTP (let’s agree not to talk about MMS lol).

How would one craft an input that maximizes the runtime of your diff while staying below some (reasonable) bound of edit distance apart? Basically suppose I wanted to provide a malicious diff designed to hang or timeout the diff program, how would I construct it?

Not OP, but actually I do because I have been forced to live in places where it’s dangerous to walk and where all the scenic landscape has been replaced with asphalt.

I get your point in general, but I think vehicles are a poor analogy because in many ways they have changed the places we live for the worse (and yes I’ll concede occasionally for the better). It turns out we need places to park all of these vehicles and they compete for space with humans.

This is a silly take. I think you underestimate how talented some programmers can be. I think you also overestimate how far being good at coding will get you. Pre-LLM, the job was maybe 30% coding. So at worst someone who codes by hand is only disadvantaged by a manageable amount.

I get that they’re not equivalent, price wise, but have you tried searching for “55 inch computer monitor”? They sell large ones for use in video conferencing rooms for example.

What benefit is a feudal lord to a human soul?

I was going to point out VoLTE as an example of a system which runs over IP but still retains a lot of the concepts from circuit switched days. In theory (if you’re a network operator), you could expose dedicated bearer provisioning with QoS guarantees for applications to build on top of. I guess my point is that you can have tighter QoS control if you can get off the public internet and go through carrier IPX instead.

How do you handle degenerate trees? I’m curious what heuristics are used in practice to eg backtrack or whatever. I’m not very well versed in TED but superficially the problem seems to resemble wave function collapse.