Yet another recent example of AI assistance in mathematics: my colleague Ernest Ryu here at UCLA was able to solve an open problem in optimization theory (roughly speaking, an asymptotic convergence result for a certain class of ODEs) in large part through an extended conversation with a large language model, serving both as a "rubber duck" and a stochastic generator of proof ideas. x.com/ErnestRyu/status/1980759… Many of the ideas generated were not usable, but a non-trivial fraction of them did contain some viable strategies that had not been immediately evident to Ernest. There was a significant pruning process to isolate the small number of useful ideas and discard the larger set of non-useful ones, but even so, the tool provided a net time saving in reaching a working argument, which Ernest then polished by hand into a two-page proof. (1/2)
I have had multiple persons tell me recently that they truly hesitated and made really sure they didn't submit slop before they filed their first security reports to #curl.
Meaning: public shaming seems to at least partially work. Banning, taunting and ridiculing the fools works as a reminder for people to maybe think again and make sure.
@tykling wonder maybe, but we need to survive so I see no real alternative here. Plus, we keep getting quality reports as well so we know they still show up.
unfortunately, it mostly works on the wrong people. Those who would already have put care into their work are now even more afraid to do something wrong.
(Edit: Not saying what you're doing is wrong. I certainly don't have a better idea. Just pointing out the conundrum.)
@wyatt the term "slop" is usually used for what the AI outputs, but sure they usually just ask an AI to find a problem and it invents something which to a layman may sound accurate
right or wrong, a response is at least useful. The person can decided whether it's worth their time doubling down constructively, and if they really are honest they likely will because they care (it's kinda like human greylisting) . Importantly, the burden and onus is not on you.
I've submitted many bugs in closed source software (mostly Microsoft) and hear literally nothing back.
The bragging rights & CV / resume enhancement value of “I found a bug in curl” are a powerful motivator. Many will choose the blue pill of AI instead of hard work to get ahead of competitors. The possibility of public shaming and identification as imposters raises the bar
In November 2022, after I had been keeping track and adding names to this slide for a few years already, we could boast about curl having run on 89 different operating systems and only one year later we celebrated having reached 100 operating systems…
One reason I really really love running Gemini in Chrome is that it helps me to deal with many of these pesky dashboards. A press of a shortcut key, a quick prompt, and an accessible easy-to-follow summary. How is that not #accessibility? :)
OpenAI, the company behind ChatGPT, released their own browser called Atlas, and it actually is something new: the first browser that actively fights against the web. anildash.com//2025/10/22/atlas… #Internet #WebDev #AI
Think about it. Rose has an exhaust system leak into her cabin, crashes her jeep. Then she ends up in a town that was abandoned because of the coal fires. It's carbon monoxide all the way down
hey hey you with the text-heavy website c'mere i wanna tell you something closer good now listen carefully [at the top of my lungs] THE BODY COPY SHOULD BE SET NO SMALLER THAN 12-POINT TYPE AND THE TEXT BLOCK SHOULD BE NO MORE THAN FIFTY EMS WIDE
Is there a way to express the former in CSS without saying that no matter what the user has their font size set to, I want it to be either a little bigger (assuming a slightly too small default) or 12pt?
Hey, Windows people, that thing where accessing a UNC path or something like //malicious_server/blah.txt leaks the NTLM blahblah to an attacker-controlled hostname...
Is that a genuine concern? Do people go and patch things to basically disallow that everywhere? Is it assumed that UNC paths should only be used for trusted hosts?
Since Windows Vista, NetBIOS and Windows DNS routing has been blocked for networks the user declares as non-private (ie not home or work networks). So UNC paths basically shouldn't work except on user-trusted networks in the first place.
I have a Windows machine around in my home network and it operates as a network media drive server (for all kinds of legacy reasons that aren't really important), so I still keep some awareness of how all this stuff works with Windows. 😃
Estos días estuve haciendo un mandil de trabajo para hacer #carpintería. Lo hice con gabardina de un pantalón viejo. Tiene partes chuecas pero es parte del encanto 🫢
Honestly not surprised with it costing like $150 not that long ago. I pretty much consider it a bootloader for anticheat games and not a tool for any real use.
saw it just the other day. But why do they charge 30 euro for free operating system? It is fine if they are adding value, like drivers, etc.But are they?
@jarekrozanski we in the RH desktop team work with them pretty closely, so I know they pay a team of Linux engineers and they really make an effort to support their hardware in Linux well. They also provide quite a lot of hardware to the Fedora Project for testing. IMHO the 30 euros are well deserved.
Want Inkscape's Multi-page to work in web browsers (and other tools)?
Well my work today has been on moving our inkscape:page element to the SVG standard "view" element. Already supported in web browsers and hopefully supported someday in SVG viewers that use librsvg (are you interested @federicomena ?)
This will need testing and a migration for older data too. But I like more standards compliant files and less custom stuff in inkscape's svg.
Is there any open-weights generative AI model that is trained to generate natural-language descriptions of images, but, crucially, _isn't_ trained to follow instructions? So something like Microsoft's image description model that has long been used by the Seeing AI app among other things, not the current crop of instruction-fine-tuned vision language models.
What about the llama-vision models? I suppose there must be a base model without instruction following for finetuning, or it might be possible to remove post training somehow.
I had no idea that the Screen Wrapping add-on existed so thank you so much for making the community aware of it. I admit I'm one of those users who know about and use a few add-ons but I don't explore the add-on store all that much and I probably should. I hope you call out an add-on with every issue.
We'll try and highlight an add-on more regularly, yes!
Kev, good question. Could we add it to NVDA itself? Possibly (well, there's an add-on for it, so YES it could be done).
It has been asked for previously, and at the time the rationale against it was more around the feature than the technicalities of implementing it. See this discussion here: github.com/nvaccess/nvda/issue…
Reported by RamboUdin on 2010-08-21 09:43 When navigating webpages by heading,links as well as landmark, and you've reach the bottom of top of document, pressing that keystrokes will say, "No ** he...
@kev @DavidGoldfield We'd like to be able to better track the popularity of add-ons, so that requests like this, where there is an add-on can be evaluated with that knowledge. Because XYZ does it, is, of itself, not necessarily an automatic justification (as an example: Jaws has its own interface for the notification area icons, because that area wasn't accessible originally. By the time NVDA was created, it WAS accessible so having that functionality in a screen reader was no longer necessary).
@DavidGoldfield That makes sense about JAWS and the notification area, but some people still like it for it's convenience. FIgured the screen wrapping could be treated the same way, especially since new users don't always know about addons.
People say it's an urban legend. But it's not. Because I do it. Do what I hear you say? I boil my undies in hotel kettles. I mean, how else am I supposed to clean them? It's not like there's a washing machine in the room.
#curl gets some of the worst #AIslop "vulnerabilities" reported to it via Hackerone: Here we have a fake 90s exploit assuming executable stack and x86 arch. Someone seriously passing this as their own research is stupid beyond belief.
An additional tax on millionaires in Massachusetts has generated $3 billion more than expected without forcing significant high-profile departures from the state. The money is being used to fund bridge repairs, literacy programs and public transport.
Hey Fediverse! Announcing cycle.town, a new instance dedicated to urbanism and bikes!
We're running the glitch-soc patches, including increased character limit. Still more to configure, but the server is ready to take registrations. (Our domain is just a few days new and is still being marked as spam by some email providers; check your junk folder for your confirmation emails.)
Come follow me at @streetmeme for my main account focused on Minneapolis urbanism.
Tuta
in reply to D4v • • •