tauon's recent activity
-
Comment on How I feel about LLM (AI) writing in ~tech
-
Comment on How I feel about LLM (AI) writing in ~tech
tauon Link ParentYes, it’s very much an intuition thing, IMO. (Dick Guindon) I think this quote succinctly captures why we value written thought; it’s been refined at least once from after the thought’s initial...Yes, it’s very much an intuition thing, IMO.
“Writing is nature's way of letting you know how sloppy your thinking is.”
(Dick Guindon)
I think this quote succinctly captures why we value written thought; it’s been refined at least once from after the thought’s initial inception.
-
Comment on Curl will end its bug bounty program by the end of January due to excessive AI generated reports in ~comp
tauon Link ParentAnd the accompanying blog article to that Mastodon post: https://daniel.haxx.se/blog/2026/04/22/high-quality-chaos/ The graphs are… quite something.And the accompanying blog article to that Mastodon post: https://daniel.haxx.se/blog/2026/04/22/high-quality-chaos/
The graphs are… quite something.
-
Comment on The boy that cried Mythos in ~comp
tauon Link ParentNoise/pun galore I prefer Fix (ΦΙΞ) in terms of Greek beer brands, maybe they could contribute some of those to the open-source community!Noise/pun galore
I prefer Fix (ΦΙΞ) in terms of Greek beer brands, maybe they could contribute some of those to the open-source community! -
Comment on The boy that cried Mythos in ~comp
tauon (edited )Link ParentYes, I was considering bringing up that post as well, but ultimately decided against it, since there is no information as far as I can tell on how extensively Mythos was used to find something in...Yes, I was considering bringing up that post as well, but ultimately decided against it, since
- there is no information as far as I can tell on how extensively Mythos was used to find something in
curl: knowing how big “corporate”/PR style projects go, they probably wanted to cover at least some ground with as many “prestige” repos to show for as possible rather than going in-depth on one in particular, maybe even going so far as to actively move on to the next after the first confirmed finding (“look, we found at least one issue in <large number of> huge software projects” sounds better than “we dug up 15 issues incurl, which may or may not have happened anyways”) - like you said the other models are “responsible” for at the very least just as much recent output, both quantitatively as well as qualitatively, so the trend direction for this is very clear regardless of a specific Anthropic marketing plot.
- there is no information as far as I can tell on how extensively Mythos was used to find something in
-
Comment on “Rediscovering” the operating system (AKA: the desktop is the killer app) in ~tech
tauon Link ParentSame for me, pretty much – "files in folders" is a great way to put the mantra. The only thing I do miss sometimes is the Photos (in my case iOS, but I'm assuming most have this) app's "map"...Same for me, pretty much – "files in folders" is a great way to put the mantra.
The only thing I do miss sometimes is the Photos (in my case iOS, but I'm assuming most have this) app's "map" feature, where you can search over the entire library by location, interactively. Sometimes the date (which I usually sort my folders by) is not as relevant, for example if you want to show a "best of" a given location, or the changes it has seen over the years, or even when you know where a photo was taken, but not quite when.
This is probably something that can be replicated with a standalone program and some metadata/index over the entirety of a directory likePictures/, but I haven't bothered checking if such a tool exists yet. -
Comment on San Francisco-based startup Andon Labs has put an artificial intelligence agent nicknamed “Mona” in charge at an experimental café in the Swedish capital in ~food
tauon Link ParentUgh, I love the Manna story. It's such a great juxtaposition of Spoiler, I guess if you wanted to go in knowing nothing the dystopian scenario on one hand, where no collective action was taken and...Ugh, I love the Manna story. It's such a great juxtaposition of
Spoiler, I guess if you wanted to go in knowing nothing
the dystopian scenario on one hand, where no collective action was taken and the status quo of an individualist society amplifies, and what could be possible if people didn't have to fight for making a basic living on the other: artists could be artists, researchers could be researchers, teachers could teach, people could live however they wanted to, health, "wealth", and general quality of life explodes through the roof and so on.Definitely seconding, read it for some food for thought.
Edit: Apparently it's actually still live at https://marshallbrain.com/manna, no need to hit the Internet Archive :-)
-
Comment on The boy that cried Mythos in ~comp
tauon Link ParentThis sounds just like the experience people like Daniel Stenberg of curl or kernel maintainer Greg Kroah-Hartman had up until approximately the start of this year. But ever since that point, both...This sounds just like the experience people like Daniel Stenberg of
curlor kernel maintainer Greg Kroah-Hartman had up until approximately the start of this year.
But ever since that point, both say that things have started to change. Sure, still many, many bogus reports, but increasingly also high-quality findings, and even if the issue is not a security-related finding, some can still "just" be regular, good, hard-to-spot bugs.These are obviously among the two biggest projects in existence in terms of attention by public eyes, so maybe it only takes longer until the higher-quality findings reach software that isn't among the literal biggest (in terms of users) or oldest open-source projects?
-
Comment on Aurora: A leverage-aware optimizer for rectangular matrices in ~comp
tauon Link ParentI feel the same way. For example, the much-publicized TurboQuant from a while back hasn’t found mainstream adoption yet, I don’t think? And then we’ll have hardware solutions upcoming for...I feel the same way. For example, the much-publicized TurboQuant from a while back hasn’t found mainstream adoption yet, I don’t think?
And then we’ll have hardware solutions upcoming for inference performance boosts, too, which typically have way longer lead times than is the case with novel software ideas… Cf. for example ChatJimmy by Taalas.
Then all of that is notwithstanding a completely novel research discovery in the “mechanics” area regarding model architecture or the like, like @FlippantGod alluded to above.
-
Comment on Goldman Sachs flags Amazon and Alphabet for inflating S&P 500 earnings growth figures in ~finance
tauon Link ParentGenuine question: Is your firm checking whether that $2000 per head/month is resulting in more-than-that in benefits (whether directly monetary, i.e. revenue, or otherwise), and if so – how are...Genuine question: Is your firm checking whether that $2000 per head/month is resulting in more-than-that in benefits (whether directly monetary, i.e. revenue, or otherwise), and if so – how are they going about it/using what methods for such measurements?
Because these are some wild numbers.
-
Comment on Tildes Survey #2: What country do you live in? (Results) in ~talk
-
Comment on Tildes Survey #3: What country were you born in? (Results) in ~talk
-
Comment on Apple set to become third-biggest laptop maker this year in ~tech
tauon Link ParentYeah, unfortunately me saying that is a pipedream, of course. I’d imagine there to be a whole lot more goodwill for their hardware from certain audiences if they made (parts of) macOS actually...And I don't forsee Apple ever offering support for anything besides their own OS. They have a rocky enough relationship with the GPL license, exacerbated by their anti-FOSS strategies. An unlocked bootloader is the best I think they'll ever do.
Yeah, unfortunately me saying that is a pipedream, of course.
I’d imagine there to be a whole lot more goodwill for their hardware from certain audiences if they made (parts of) macOS actually source-available, even under some crazy “restrictive” license, as opposed to just the very bare minimum (Darwin and XNU kernel) that’s accessible at the moment.
-
Comment on Apple set to become third-biggest laptop maker this year in ~tech
tauon Link ParentRight. They’d essentially have to, at minimum, offer all iCloud and related products from their services category on Linux first (and probably pre-install them on a “recommended” package list for...Right. They’d essentially have to, at minimum, offer all iCloud and related products from their services category on Linux first (and probably pre-install them on a “recommended” package list for every distro) before the finance/revenue division would even consider allowing such a move.
-
Comment on Apple set to become third-biggest laptop maker this year in ~tech
tauon Link ParentIn addition to @TurtleCracker’s clarification, I think another factor is that with an Apple laptop, you know whether you’re spending $3k or <$1k, it’s not coming with bloatware preinstalled either...In addition to @TurtleCracker’s clarification, I think another factor is that with an Apple laptop, you know whether you’re spending $3k or <$1k, it’s not coming with bloatware preinstalled either way.
But nevertheless, having a few OS options to run on Apple hardware would be great. Not sure why Apple doesn’t just buy up Asahi Linux (peanuts for them, but a night-and-day difference compared to the project’s current funding levels), they’d dominate the developer market even more if MacBooks offered first-class Linux/non-macOS support.
-
Comment on Synthesizing multi-agent harnesses for vulnerability discovery in ~comp
tauon (edited )Link ParentThis is a very cool finding, and/however, does not surprise me much anymore after these two realizations in particular from a bit back: Cursor boosts model performance versus other harnesses There...This is a very cool finding, and/however,
if you have the right harness, you can [do X for almost any given X]
does not surprise me much anymore after these two realizations in particular from a bit back:
- Cursor boosts model performance versus other harnesses
- There are ten separate harnesses that use Opus better than Claude Code itself does [on TerminalBench]
This has been coming on the horizon for a while now, IMO, in hindsight essentially ever since it was found to be beneficial to send LLMs off in a loop.
-
Comment on Apple names insider John Ternus as CEO, Tim Cook to become executive chairman in ~tech
tauon Link ParentWow, I had totally forgotten about that product! A friend of mine had one of these when they were new-ish. Not to overly step into “defending megacorp online” territory, but I think that was a...Wow, I had totally forgotten about that product! A friend of mine had one of these when they were new-ish.
It's almost like someone at Apple was trolling the user base.
Not to overly step into “defending megacorp online” territory, but I think that was a combination of legitimate engineering problems and business/timing issues they had to solve. Some of the constraints for the first-gen Apple Pencil:
- Not enough time for further big hardware redesign changes iPad-side, at least for that generation (the later iPads could charge the pen wirelessly, and had magnets to hold it in place)
- No official lightning-to-lightning cable for the pen, because users would probably attempt to do iPhone shenanigans with it and complain about something not being possible (like one iOS device charging another); and it’ll risk damage to iPad, Pencil, or itself because it will get used on the go while everything’s stuffed in a backpack – not likely to survive for long, especially if it’s a short cable
- Plus you might end up with pencil users leaving it permanently connected, and we know what Apple of that era thought about unintended use cases (and optics)
- No USB-C to USB-C (or USB-C/USB-A to lightning) cable, because users should not need to carry a second, distinct cable for their pen (or one that doesn’t connect to the iPad when you want to charge the pen with no outlet nearby, e.g. in every classroom ever)
… leaving little room for any other implementation. They probably did some internal testing and user research and figured the pen actually breaking off is rare enough to accept the design’s ridiculousness.
-
Comment on I made a website with free and low-cost resources for web development, game development, privacy, graphics, small web, etc in ~tech
tauon (edited )LinkCool stuff! I really enjoyed the (even 2026-updated!) email obfuscation article, since I’ve dealt with that topic before personally (and didn’t think scraping mechanisms would still be this...Cool stuff! I really enjoyed the (even 2026-updated!) email obfuscation article, since I’ve dealt with that topic before personally (and didn’t think scraping mechanisms would still be this dumb/cheap, so was more concerned than I perhaps needed to be :P).
The more things you add, the longer this page will become. Maybe you could consider adding a table of contents, at least for the bigger overarching categories?
-
Comment on Apple names insider John Ternus as CEO, Tim Cook to become executive chairman in ~tech
tauon Link ParentCould also be an optimism that their current good run wouldn’t be ended. :PCould also be an optimism that their current good run wouldn’t be ended. :P
-
Comment on What's a battle that nobody knows you're fighting? in ~talk
tauon Link ParentFor what it’s worth, you don’t know me, and I don’t know you, so feel free to DM me if you want to talk about anything you can’t really or don’t want to discuss IRL. (This is directed at any- and...For what it’s worth, you don’t know me, and I don’t know you, so feel free to DM me if you want to talk about anything you can’t really or don’t want to discuss IRL.
(This is directed at any- and everyone reading. No pressure, though, just throwing it out there. :-))
Just as a heads-up, unfortunately (like @post_below demonstrated) it is fundamentally not possible for these tools to be accurate on a technical level, at least in the general case.
Either you’ll let through a lot of LLM writing, or deny a lot of human writing, but when even the industry “frontier” OpenAI admits the authorship of a piece of text cannot be attributed to an LLM or a human in an automated manner, it’s not looking good.
There have been “watermarking”, uh, attempts, where models were designed to favor certain phrasings on the individual word level, which is a pretty cool tech showcase…, but especially ever since the existence of somewhat capable open-weight models, the cat has been out of the bag – Google would have had to invent this method much earlier (and be confident enough to enable it by default with everything) for it to be relied upon by a majority of models, unfortunately.