Mastodon Feed: Posts

jwz wrote:

XScreenSaver thumbnails.

Dear Lazyweb, any idea why XScreenSaver thumbnail images are no longer showing up in System Settings on macOS 27? Works fine on macOS 14.7. The .saver bundles have "Resources/thumbnail.png" or @2X files. I've even tried using...
https://jwz.org/b/ylAQ

Boosted by baldur@toot.cafe ("Baldur Bjarnason"):
mrs_malice@mastodon.world ("It's me, Alice ❤️‍🔥✨") wrote:

it took *ONE* Black woman at google to foresee this LLM shit was bad, and she was fired for that foresight

but i gotta sit here and watch almost every white man over 45 either cackle with glee at his own coming demise or disappear into the bushes when he has a chance to speak up

how can white patriarchy take itself even the slightest bit seriously when it's so painfully clear our society is built on stunting white men's social/emotional/spiritual development & rewarding them for it

Boosted by ireneista@irenes.space ("Irenes (many)"):
JoscelynTransient@chaosfem.tw ("Joscelyn Transpiring") wrote:

I remember when assholes were telling the queer and trans folks standing against genocide of Palestinians that Muslims over there would kill us. More than a year out though, you know who is standing in solidarity with us against anti-trans shit in the US? Muslim Americans who have our backs like we had theirs.

It’s not universal and there are still tensions in certain corners, but like…Mamdani and Al-Sayed are out here standing up for us, Muslim friends of mine show up for us, and Muslims aren’t the ones throwing us under the bus right now but white Christian liberals are.

This is coalition and solidarity. When we stand with others who also are hurt and face violence from these systems, we build bridges that make us all more powerful in fighting for a better world.

Boosted by andrewnez ("Andrew Nesbitt"):
githubchangelog ("Unofficial GitHub Changelog") wrote:

Opt-in dist-tag permissions for npm trusted publishing https://github.blog/changelog/2026-09-30-opt-in-dist-tag-permissions-for-npm-trusted-publishing
#github

Boosted by pixxl:
animesh ("baksignal") wrote:

"Programming is building a mental model. It is not mechanical code production. The shared mental model in developers' minds is the real product. A program is "living" only while the team understands this theory. When programmers leave, it becomes "dead". Without the original theory, new maintainers treat software as a black box, creating patchwork fixes that cause side effects and architectural decay."

Absolute truth.

Originally from Peter Naur's "Programming as Theory Building" (1985)

Boosted by pixxl:
chrisamaphone@hci.social ("chris martens") wrote:

cs department/college leaders who touch finances will draw a breath to handwring about how enrollments are down and grants are waning, then exhale to tell us we should change our curriculum to "prepare students for the AI future" and even rename entire programs e.g. from data science to AI. they say we should "have careful conversations about what this means for the future". we should interpret these dicta as from the mouths of hostages.

Boosted by pixxl:
chrisamaphone@hci.social ("chris martens") wrote:

i'm really unimpressed with how quickly academics seem inclined to throw *ourselves* under the bus with this "ai product supremacy is inevitable" bullshit. you realize this propaganda is part of the active political mission (which has been made quite explicit by certain people in power) to destroy public perception of universities and end the possibility of people like us having the lives and careers we have, right?? we don't have to lie down and take it.

Boosted by pixxl:
ricci@discuss.systems ("Rob Ricci") wrote:

I kind of hate that *gestures at everything* is providing such fertile ground for thinking about @viveksrikumar and C Thi Nguyen's Ethics in AI class (which I'm sitting in on), but it is, so, here I go again.

One of the ideas they've been talking about in class is risk tolerance. Pretty basic in concept, people make decisions about how much risk they are willing to take on. A lot of this is very mundane and implicit (is the second patty worth the cholesterol risk? worth it to dash across the street here instead of going to the corner?) Some of it is explicit in things like investment decisions. The risk tolerance that you take on is a reflection of your values - play it safe? shoot for the moon? Pretty standard stuff.

One way, however, that this shows up in unexpected (to many of us) ways is in things like like science and medicine: it's a value judgement that we don't often acknowledge. A threshold for p-values of 0.05, which most folks use unquestioningly, means (approximately, it's complicated) a tolerance of a 5% chance of being wrong. Is this ... a good amount of risk to tolerate? It's a nice round amount, that's cool. Is it too high? Is it too low? Depends, probably?

Here's a pretty standard illustration of "it depends" that was used in class: risk tolerance in policing. Let's say you've got some technology, totally hypothetical here, that reads license plates with cameras, let's call it Flook. Somebody, somewhere, gets to set thresholds for what's a certain enough match that the cops should roll out and pull over the driver. This might seem like it's just some number, but it's a value judgement about the risk of being wrong.

The cops? They have a pretty high risk tolerance in this situation. Sure, they don't (let's assume, to be generous) want to spend all day pulling over the "wrong" people. But the cost to pulling over the wrong person is just not that high for them. A bit of wasted time, maybe a confrontation with driver that's mad at you, possibly a bit of paperwork. Pretty unlikely to be something that gets you in serious trouble, plus, hey the machine started it. And (I know, assumptions...) they'd like to not let the "right" people get away. So they want that threshold to be pretty loose.

The people being policed? This looks very different for them. Nobody wants to be pulled over when they did nothing wrong. Low risk tolerance. And it definitely isn't uniform. For some, it's a hassle, but a bigger one than it was for the cops - it's the cops' job to hassle people, it's not everyone else's job to be hassled. Pretty low risk tolerance. For others every interaction with the cops carries a *huge* amount of risk. Way more hassling. Higher risk of being detained. Possibly violence. Even if the only thing you did wrong was to have a license plate that looked too blurry from the top of a 10-foot pole, or got typoed when it was entered into a database. Super low risk tolerance. These folks do want the "right" people caught, but they want these thresholds set so low that no cop ever pulls over the "wrong" person - if this means sometimes missing the "right" person, well, they know that the consequences for a bad stop can be super high, so it's worth some risk in the other direction.

So how should Flook set thresholds? How should they balance the risk of incorrect identifications with the desire to not miss correct ones?

Trick question, because Flook has a customer, and the customer is the cops, so they do what the cops want.

This system contains embodied values, and they are a fundamentally un-democratic ones. Only the cops get any direct influence over the secret number that reflects how much "we" value catching criminals vs. surveilling and wrongly pulling over people just trying to have a day. You might see why people are getting increasingly mad at the definitely-made-up Flook.

This is how a number that none of us even know expresses values that ... well, maybe a lot of us strongly disagree with.

And then, one day in early autumn, a Perfectly Normal Guy who's a high-up engineer at Anthropic (the Good Guys of AI, remember), writes this, for everyone to read:

"Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.

To be clear, as we say in our latest Risk Report [link] I think the risk from present models is low. What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought [link]"

So THERE'S some risk tolerance for you.

(Sorry for making you read that, but I want to give you the full quote.)

Apparently, it is the calculation of the Good Guys of AI that a 90% chance of survival for all humans is Just Fine and an acceptable risk to take for whatever they think we are getting out of it. We know they think they are getting rich. I guess they think there are other benefits, but damn, you can sure see why they will say literally anything to justify a 10% risk of human extinction. Better be a bigger deal than 10,000 new TODO apps!

I want to pause for a second here to say that I think that number is utter bullshit (the 10% one, the number of TODO apps is probably in the right order of magnitude). I do not think there is a measurable chance that all humans will be dead Because AI by 2030. I have no idea whether they actually believe this or are lying. Doesn't matter. They are telling us, directly, what their values are, or at the very least what values they would like us to think they have.

They are telling us that their risk tolerance is at a deranged, misanthropic level. I'm not excited to use the word "psychotic" but it's really hard to find something else that fits. Because they do, in fact, have the option to Just Not. If you believe that the product you are working on has a 10% chance of killing everyone, and you are, like, a person who gives a shit about - well, anything, really - then what you actually do is turn off the servers, delete all the files you can find, beat the GPU swords into Sweet Gaming Rig plowshares, and save the human race. You do not prepend Google's long-abandoned slogan (you know the one) to all the prompts and keep Doing the Thing.

What about risk tolerance for the rest of us? Where are we at on this? What do you think the benefits are or could be, what do you think the harms are or could be? What risks do you think are appropriate to take for the possible benefits? What's a reasonable way to spread out both the benefits and harms such that ...

HAHA, trick question again! The Good Guys of AI have a customer, that customer is The Future, and The Future demands that maybe we all die, very sad, but that's a risk it's willing to take. Only the Good Guys of AI can possibly make sure that the sacrifices are the right ones (not that they know how, but they'll figure it out), and the only thing that beats A Bad Guy With an AI is a Good Guy With An AI.

I probably don't have to point this out, but this is as nakedly authoritarian as it gets. We all bear the risks, The Good Guys of AI must be the ones to call the shots.

So yeah, these are their values. Doesn't matter if their assessment is wrong (probably is). Doesn't matter if they're lying about what they really believe (who knows). Who cares how many whitepapers and reports they write. They get to set the risk tolerance for everyone. At least the number isn't secret: it's 10%. Flook almost certainly has a tighter threshold for pulling over the wrong car than "we" apparently have for total human extinction.

Who signed up for this?

Who died (only p(10%) though) and made them king?

Are these your values?

Boosted by pixxl:
amy@types.pl ("Amélia Liao 🍊😈") wrote:

really the only innovation that we can credit the llm labs with is that they managed to invent a class of person even more annoying than evangelicals. the cult of the frontier model

Boosted by brib@bribstodon.xyz ("brib :neofox_floof:​ :Nonbinary:"):
angst_ridden@turtleisland.social ("angst-ridden wanderer") wrote:

Was trying to fix a complex, widely used "continuous integration" server that had an unexpected configuration change in the last update that breaks our 15-year-old pipeline.

Friends, I got so lost that after three hours, I resorted to asking one then another of the leading AI chatbots for help.

Both gave me precise steps, specifying nonexistent UI panels, nonexistent configuration files, and/or command-line parameters that haven't been supported in the last 10 versions ...

What a shambles.

Boosted by brib@bribstodon.xyz ("brib :neofox_floof:​ :Nonbinary:"):
noy@mastodon.world ("น้อย คงคะดี") wrote:

Adding #sketching / #drawing to #fediverse

…on first day of #October

jsonstein@masto.deoan.org ("Jeff Sonstein") wrote:

“… enrollees of Arkansas’ Medicaid expansion — the focus of the “welfare to work” program — are overwhelmingly white, according to an analysis of data from August that the State Department of Human Services provided to The New York Times. They are slightly more likely to be female than male.”

https://www.nytimes.com/2026/09/29/us/sarah-huckabee-sanders-campaign-ad.html?smid=nytcore-ios-share

Boosted by brib@bribstodon.xyz ("brib :neofox_floof:​ :Nonbinary:"):
platinumtulip@sunny.garden ("tulip 🌷:BlobCat_uwu:") wrote:

thank you! here's a secret tip: saturation is an underrated color trait! contrast a light, muted greyish tone with a deeper, more saturated color. or vice versa! sites like https://coolors.co/ can be helpful for getting inspiration and building color schemes too

Boosted by pixxl:
mcc wrote:

Also can I just say how annoyed it makes me that html at some point made /> on certain html elements (like, ) optional or even disfavored. XHTML was a good idea. Why didn't we stick with that. How the fuck am I supposed to remember if /> is banned or required

Boosted by pixxl:
amy@treehouse.systems ("ames") wrote:

also like seriously does nobody else remember when bitcoin was the inevitable future of banking and then NFTs were the inevitable future of ownership and then the metaverse was the inevitable future of, like, existing in public, or whatever it is Zuck thought? how does this scam keep working?? how did you fall for it on the fourth attempt in the last 10 years! you fucking idiot!

Boosted by pixxl:
emilymbender@dair-community.social ("Prof. Emily M. Bender(she/her)") wrote:

From @timnitGebru and me on what we should take from the 2026 summer of AI hype:

https://www.technologyreview.com/2026/09/22/1144867/dont-be-fooled-summer-ai-hype/

Boosted by pixxl:
haskal@types.pl ("witch hat hacker :verified_dragon: :vt1: :vt2:") wrote:

"no way to generate JSON output" says only software product which regularly has trouble generating JSON output

Boosted by pixxl:
haskal@types.pl ("witch hat hacker :verified_dragon: :vt1: :vt2:") wrote:

reminder that the way popular LLM products generate for example JSON output is that they run the LLM, check if the output has any parse errors, and if so they just fucking run the LLM again in a loop until the output is syntactically valid
like
tell me this is normal

Boosted by pixxl:
haskal@types.pl ("witch hat hacker :verified_dragon: :vt1: :vt2:") wrote:

the amount of magic rituals associated with LLMs is like genuinely mind boggling. it's so fucking stupid

Boosted by pixxl:
haskal@types.pl ("witch hat hacker :verified_dragon: :vt1: :vt2:") wrote:

reminded of Gloaguen et al "Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?" https://arxiv.org/abs/2602.11988

Surprisingly, we find that providing context files does not generally improve task success rates, while increasing inference cost by over 20% on average. This observation holds across different LLMs, coding agents, and for both LLM-generated and developer-committed context files.

Boosted by pixxl:
haskal@types.pl ("witch hat hacker :verified_dragon: :vt1: :vt2:") wrote:

"if you create a file called GUMBIES.txt and write the fuckin iliad in it and then go to your random number generator it'll generate better random numbers" "if that doesn't work i switch from the normal RNG to the one with the turbo encabulator version 3.7" "the RNG can create 5 more RNG sub-instances to see if any of them generate the good numbers" like genuine clown shoes behavior. are u literally stupid

Boosted by pixxl:
SnoopJ@hachyderm.io wrote:

gonna take that as a 'no', then

jwz wrote:

The continuing downfall of USPS.

For the third or fourth year in a row, it is October and USPS has issued new Christmas stamps but there are *no* Halloween stamps on sale. I'm sure this is also DeJoy's fault.
https://jwz.org/b/ylAO

timnitGebru@dair-community.social ("Timnit Gebru (she/her).") wrote:

@Meron Lovely to see the people from all over the world who have joined us. I see Siri Lanka, South Africa, Germany, US, Argentina at least so far.

We are discussing horrific events but I'm glad to have this community around the world commemorating this horrific event with us.

Boosted by pixxl:
aeva@mastodon.gamedev.place wrote:

@mcc my favorite html trick is you can just put the little badge on your website that says it is conformant without actually passing validation

Boosted by pixxl:
lina@neuromatch.social ("Lina") wrote:

stop SHAMING people for using the unfit for purpose, ecocide-accelerating, labor-stealing, software-degrading, research-corrupting, privacy-invading, disinformation-producing, CSAM-generating, suicide-encouraging, mass violence-enabling, global economy-precaritizing, democracy-eroding, deskilling, brainrotting, futures-foreclosing tech product that was made by the eugenic death cult fascists who bankroll race science promoters and nazi politicians! it's just a value-neutral tool!

glyph ("Glyph") wrote:

Anyway it's a real bummer that I think she's the most popular journalist on the fediverse

Boosted by jsonstein@masto.deoan.org ("Jeff Sonstein"):
mary1kathy.bsky.social@bsky.brid.gy ("KittyKatGirl😻🙏💞🐞") wrote:

glyph ("Glyph") wrote:

You can also see her telegraphing this a little bit in her defensiveness about Substack and her constant complaining about how hard it is to get paid as a journalist and how she deserves so much more money. Which, again, not wrong, and people should not have been critical of her for pointing it out, but also, if your worldview is that you _personally_ deserve more money rather than deserving more money for your role as a quality journalist, then "pivot to grifting" is just the next logical step.

Boosted by pixxl:
amy@types.pl ("Amélia Liao 🍊😈") wrote:

by the way if you would like to participate in this absolute crime against typesetting today i learned that pdfmsym provides \suum and \prood