bscphil

Are you using the vanilla Debian kernel? The advantage of Ubuntu is that you don't need to depend on DKMS for something as vital as a file system.

I agree. I think for this kind of research it would be much better to at least attempt it first with a weaker model where the whole process is transparent. I was working on a project a while back, and Deepseek's reasoning said something to the effect of "I remember (from leaked copies of the source code)..." Of course that could be hallucinated, but it would be good to know whether Astra is bringing to bear knowledge on how to post-train exactly this model.

The thing I'd be worried about when doing this is spending $1000+ on a TV, then realizing it has some sort of pre-downloaded ads that it shows on the main screen ("sign up for your free six months of HBO Max!") or otherwise has a nag prompt that appears to tell you to connect to WiFi. I think that would be basically intolerable.

I think the places the average person starts out with AI do look like this. Most people are not paying a $20 a month Anthropic subscription just to try it out. They're using free chat apps, reading the garbage llm summary Google puts at the top of every search page, dealing with the low quality tools foisted on them in their email app or word processor.

Have all these gotten a lot better in the last year? Sure! But there's still a several-orders-of-magnitude difference in quality between whatever garbage tier of GPT-6 Luna you can use for free on `duck.ai` (just for one example) and Opus 5.5. Or even GLM 5.3 Flash. It's a way bigger difference than the improvement in ~free models has made up for in the last year, even though those improvements have been substantial.

I think a certain sort of person, if they do try to use something better along with a coding harness, as also immediately inclined to read along with the reasoning traces - and understandably gets freaked out by the nonsense it contains. That's a mistake, I think; not just because models behaving "reasonably" doesn't depend on the rationality of their reasoning traces, but also because I think if you transcribed my brain activity into a serial stream of text, I'm not convinced it would be more coherent.

Heard of them, but I think it's fair to say their home page doesn't inspire a ton of confidence?

Mullvad doesn't let you forward ports.

Maybe I'm misunderstanding you, but a better attack on a reduced-round AES has been achieved by LLMs. Anthropic did it and published it already, a chosen plaintext attack that improves on state of the art against reduced-round by "200-800x". https://anthropic.com/document/aes_mobius_bridge.pdf

The only reply to this criticizes the hedging, but I think it's actually true and it does make it harder to tell if you're vulnerable to this. And on top of that, quite a few systems have brotli installed by necessity. On Debian 12, for example, I have the following dependency chain on a VPS I run:

grub-pc -> grub-common -> libfreetype6 -> libbrotli1

which is really nasty.

Right, I have to think P(AGI by century's end | LLMs turn out to be an AGI dead end) is still at least 50%.

(tongue mostly in cheek) sounds like I should be renting out these platforms and selling inference on them to OpenRouter

My experience has been that the "cost per task" estimate wildly underestimates the cost of top models on most real work, because the tasks (as I understand it) tend to be isolated bits of problem solving or code generation, which rewards writing code that works the first time and not generating too many reasoning tokens. Sol is great for both. However, a lot of real work involves tracking down bugs, working through tests, examining many parts of a codebase, refactoring... even with subagents and good tool use this tends to bloat context before the problem gets solved. Deepseek's absurdly low cache price really shines in these instances, but even GLM 5.3 Flash easily beats the top models on price in real world usage. Maybe that's different if you factor in subscriptions.

Yep, any time you're asked to set up a passkey it feels like Russian Roulette with future access to that account. A year or so ago, a site I relied on started enforcing biometric passkey enrollment. Without thinking about it I logged into the site on my phone and enrolled there, but unfortunately (a) that meant it didn't get registered as an "other device" in the internal system, and (b) the site didn't work correctly on mobile devices (legacy system). So logging in on a desktop would then demand that I activate the passkey, but of course the passkey was on the phone. No way to access it, and the site didn't offer me the QR code / bluetooth method. I don't remember exactly why, but it just didn't work - probably had something to do with it not liking Firefox on Linux or something.

And to add insult to injury, it wasn't even a single factor login the way passkey advocates promise. The site still demanded a normal password to log in! It was actually 3 factor authentication; something I know (the password), something I have (the passkey), something I am (the biometric).

This wasn't some random website that didn't matter made by nobodies. It was a US government website for managing grant money.

Adobe almost certainly has patents covering aspects of Photoshops design and tools?

Microsoft famously has patents covering aspects of the ribbon interface in Microsoft office, which of course this suite must implement. https://en.wikipedia.org/wiki/Ribbon_(user_interface)#Patent...

Meanwhile, Wikipedia has a policy that says screenshots of applications have to be "as small a version as possible" in order to meet the fair use exception for presentation of copyright work. I can only imagine an exact clone of the interface could be similarly ruled to infringe on Adobe's intellectual property.

I was trialing MiMo-V2.6-Pro recently due to its high benchmark scores, and it argued that substring matching the names of audio codecs in a search field was a mistake because "a user searching for 'aac' would get unwanted results for 'alac'." Which isn't exactly counting letters per se, but there are still weird issues with understanding words as strings rather than as tokens.

> A surprising number of people in the state would not consider 1 mile 'walkable'.

In a surprising number of American cities, 1 mile is not walkable, simply because no one thought to put in a sidewalk or have adequate protection for pedestrians. That goes double for places like Lansing, where the sidewalks that are there are often icy during the winter.

Oh cool, someone talking about the really weird bug that I've never seen reported elsewhere. Seems like an issue where certain versions of Unity (to name one engine) doesn't like it when your actual game controller is the second joystick it detects. Not really Keychron's fault or Linux's fault, just a weird engine quirk.

Anyway, if someone finds this by searching, at least on Arch it's as simple as adding a custom hwdb entry at /etc/udev/hwdb.d/61-input-id-local.hwdb:

  id-input:modalias:input:b0003v3434p0283*
   ID_INPUT_JOYSTICK=

and then run systemd-hwdb update.

The modalias value is found at /sys/class/input/<yourdevice>/device/modalias

Why did Anthropic settle with authors for 1.5 billion then? Surely their lawyers must have decided there's a pretty good chance of judges ultimately deciding that it is copyright infringement?

> It's implicitly assumed when you read something that a person write it.

I basically agree, but it's more specific than that, I think. The basic promise is that the writer believes the thing you are writing. This is a feature of human speech that is so central that Grice made it part of one of the maxims of his pragmatic theory of conversation: "do not say what you believe to be false" or "that for which you lack adequate evidence."

In writing, using generative AI destroys the pragmatic basis for this conversational maxim, because it means that a consequential number of "people" you and I talk to online are not really people at all. They're entirely artificial, and the text generated was never even read by a human, certainly not a statement of a human's actual beliefs about anything. (This isn't even a human-preferring standard, the problem with LLM text is that even the LLM doesn't believe it.) Even in cases where a human does interact with the text in some capacity before it is published, it's likely that a plurality to a majority of the ideas reflected in the text come from a language model that does not actually believe them.

Writing is something that aids in the development of human ideas. It's not just that in the ideal case, a human being will have certain ideas and be able to communicate them to other people. Rather, we often have inchoate, unclear, or even contradictory ideas and opinions, and the process of writing itself leads us to figure out what it is, exactly, that we believe. So even in the "ideal" case of someone like OP who engages with LLMs more critically, and claims responsibility for every bit of text they publish, the process is still detrimental to the production and communication of true beliefs because their ideas were formed in dialogue with an agent that did not actually believe them. It's a process that is dangerous to bypass because you skip most of the thinking involved in writing without even realizing it.

Because the point of the test is to see what the agentic swarm approach is capable of. If your test assumes that a version of the thing you are trying to build is available in the training data, or can't determine if it matters whether the thing you're building is in the training data, then it's not a very useful test. Most of the time when you consider agentic approaches, you're not trying to rebuild something the agents involved are trained on.

It's clearly vibecoded if you look at the commit history.

Right, I think the problem with a lot of the economics-based analyses in this thread is that they rely on the assumption that Valve ought to, or wants to, sell the device at its market value.

Well, I suppose that's true strictly speaking. What I'm reacting to is your claim that this is a simple "choice" that everyone else is making:

> Embracing or shunning the society you live in is a choice.

When the reality is that it's more complicated. You're able to make this "choice" because you've spent years cultivating a quasi-religious attitude of equanimity toward things that are, from the perspective of most, annoying, troubling, or frightening. So what you're asking of most people is more than a choice (taking their AirPods out), it's more a matter of converting to a different way of life entirely.

Surely the solution to this social problem, however, can't be "everyone should simply convert to my religion / achieve a higher state of mind where they're not bothered by any form of inconvenience, irritation, or interruption." If it comes to that, most people will continue to wear their AirPods. It's a non-answer.

> about how Claude or even Codex have refused to perform some normal program development tasks

> a lot of reports that the US LLMs refuse to answer questions

I think the specific ask is for a case where the LLM is trained to lie about something. What you've come up with are cases where it refuses to do something, possibly for legal reasons but maybe not (you can come up with plausible non-legal reasons why a company training an LLM might want it to refuse to give you instructions on making a bomb, even if instructions on making a bomb are protected First Amendment speech).

An LLM that responds with "I'm sorry, due to legal requirements placed on my creators, I'm unable to answer questions about events at Tiananmen square in 1989." strikes me as much less problematic than one that pretends there is no relevant or reliable information that exists, or explicitly supports a regime narrative. But I'm also of the opinion that an LLM refusing to help you build a fertilizer bomb is much more reasonable than one that suppresses information of a political nature. I can't think of a case where information that reflects the broad consensus of experts is suppressed by US based LLMs for political reasons.

Looks like the SVG was converted from an EPS file, and the resulting SVG contains individual glyph positions (advances) for the characters in "Personality score", but it doesn't specify a valid font, probably because the font name was mangled in the original EPS file (which is pretty typical).

So whether the resulting file looks right depends on whether the rendering engine chooses the correct font. Looks like it's supposed to be Nimbus Sans or something metric compatible with that, but the serif font chosen by Typst looks obviously wrong.

> i reckon Meta would have found it challenging to download very large files without sharing. It's certainly much faster if you don't get throttled or banned by many peers.

You're not that likely to get throttled by seeds though, and most torrents that are downloadable at all have a few seeds. Seeds have no way of verifying whether you're contributing the network, they're just there because someone (implicitly) decided to make the file available to whomever drops by and asks for it.

I believe they changed the app since Trixie was released (Trixie has KDE 6.3, the changes were in 6.4) and buried a lot of the really common settings behind menus. E.g. you might want to take a screenshot on a delay, and that's now hidden behind a menu whereas they used to surface the most common features on a panel.

I don't think this is quite right. It's not that the question is inherently underspecified, it's that the context of being asked a question is itself information that we use to help answer the question. If someone asks "should I walk or drive" to do X, we assume that this is a question that a real human being would have about an actual situation, so even if all available information provided indicates that driving is the only reasonable answer, this only further confirms the hearer's mental model that something unexpected must hold.

I think it's useful to think about it through the lens of Gricean pragmatic semantics. [1] When we interpret something that someone says to us, we assume they're being cooperative conversation partners; their statements (or questions) are assumed to follow the maxim of manner and the maxim of relation for example, and this shapes how we as listeners interpret the question. So for example, we wouldn't normally expect someone to ask a question that is obviously moot given their actual needs.

So it's not that the question is really all that ambiguous, it's that we're forced (under normal circumstances where we assume the cooperative principle holds) to assume that the question is sincere and that there must be some plausible reason for walking. We only really escape that by realizing that the question is a trick question or a test of some kind. LLMs are generally not trained to make the assumption, but ~70% of humans would, which isn't particularly surprising I don't think.

[1] https://en.wikipedia.org/wiki/Cooperative_principle#Grice's_...

I like this article a lot, but if I can put forward one mild criticism, it seems to depend entirely on having exactly the same measure of semantic distance for word pairs as the original generator. In that case, as the post shows, you only need several guesses to eliminate all possibilities other than the correct one, just like you only need a few GPS satellite locks to pin down your location.

It would be interesting to see a solver that works more like a human player, where it requires the "warmer" "colder" information from different guesses to hone in, rather than being able to simply look up which words have the exact semantic distance (+/- some fudge factor) from the guess.

This is totally hypothetical, but I wonder if a system whereby your dollars went to the publications you actually read, but you could immediately, at any time read anything else you wanted for free would work. There would be an obvious reason to subscribe (you get past the paywall for any publication that is part of the bundle) but you would have the feeling that you're not "wasting" money because your money only goes to the publications you actually support.

(In reality, of course, cable providers were mostly doing this under the hood along with pocketing a big cut for themselves; television is just expensive to produce. But it didn't help the feeling of unfairness when you didn't watch any sports but ESPN was probably the most expensive channel in your "package".)