Carlos Scheidegger

Carlos Scheidegger

@cscheid.net

Engineer at Posit, technical lead for quarto.org Touching grass by learning Brazilian music. feet in Minneapolis, heart in some random Brazil beach town I don't even know yet DMs open, please try to be normal out there

4.4K followers664 following4.4K posts

Pinned

Ok, I've seen enough. I can't contribute to this medium anymore. I don't know what happens next. We resuscitate blogs and go from there? I don't know. But this is fucked, and I don't think participating at all is net positive. Back to pure work announcements, I guess. DMs open, love y'all

So the OpenAI bound is kind of hellish in that it’s subexponential but doesn’t survive many poly time reductions? That means we can’t leverage that algorithm for many npc problems because the reduction would make the solution exponential, even if cnf itself isn’t. Yeesh that sucks

Sucks because if the idea is productive we don’t get to leverage all of the reductions work that people did in the past, seems like reality is more fragile than that

Carlos Scheidegger reposted

An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4

Carlos Scheidegger reposted

By the way, I have “reply without confirmation” turned off. Are these people actually this absurdly incapable

Carlos Scheidegger@cscheid.net

This is completely fucking unhinged behavior from whatever product manager that pushed this. On the spot fireable offense if we lived in a reasonable world

Seriously. If you work at Apple, send this along? know that this is why I’m never going to turn Siri on anymore. @domoritz.de, maybe? I’m genuinely livid. Imagine if this was a serious text and I had accidentally triggered a “haha” emoji.

This is completely fucking unhinged behavior from whatever product manager that pushed this. On the spot fireable offense if we lived in a reasonable world

Carlos Scheidegger@cscheid.net

Siri, this is insane. (Context: I just upgraded to the latest phone) I got a text message from our building manager about some other renter doing some bullshit in the common areas. I blurted out “wow” entirely involuntarily. Siri helpfully put a !! emoji reaction on the text. Absolutely fucking not

Siri, this is insane. (Context: I just upgraded to the latest phone) I got a text message from our building manager about some other renter doing some bullshit in the common areas. I blurted out “wow” entirely involuntarily. Siri helpfully put a !! emoji reaction on the text. Absolutely fucking not

Carlos Scheidegger reposted

The thread deriving from @mjskay.com's response to my post reminds me of good, old #dataViz Twitter. I miss this type of conversation!

Matthew Kay@mjskay.com

I think this depends on dialect and audience To me there's a definition of misleadingness (or maybe correctness) where if there are perceptual operations on a chart your audience will tend to perform and those operations do not correspond to valid data operations, arguably the chart is incorrect

The iPhone duo is a real space age gadget but the aspect ratio is just a little too weird for me, having played with it in person a bit earlier today

It’s elegant in principle! 1/sqrt(2), invariant under folding+rotation. Clever! But I think ultimately designer brain. Just make people write two versions of each app if the app cares that much about presentation aspect ratio.

I can’t imagine the sequence of life decisions that ends with someone screaming “you just read too much mainstream media” into a phone in an airport lobby. And yet

Yup, 850GB magically recovered on reboot.

Carlos Scheidegger@cscheid.net

Then when you eventually decide to reboot your laptop, login freezes for a long time. It’s been 15+ minutes tonight I ran into this a few times and didn’t realize that’s what it was. Tonight I found a dir deep into /private/var that took 800GB according to du. I bet it won’t be there after reboot

macOS does something really fucky with its file systems huh? Pretty regularly, if I have a long uptime session that creates lots of temp files (hello 5 parallel agents building a bug rust project for 12h/day over a week), free disk space continually shrinks

Then when you eventually decide to reboot your laptop, login freezes for a long time. It’s been 15+ minutes tonight I ran into this a few times and didn’t realize that’s what it was. Tonight I found a dir deep into /private/var that took 800GB according to du. I bet it won’t be there after reboot

macOS does something really fucky with its file systems huh? Pretty regularly, if I have a long uptime session that creates lots of temp files (hello 5 parallel agents building a bug rust project for 12h/day over a week), free disk space continually shrinks

macOS does something really fucky with its file systems huh? Pretty regularly, if I have a long uptime session that creates lots of temp files (hello 5 parallel agents building a bug rust project for 12h/day over a week), free disk space continually shrinks

Then when you eventually decide to reboot your laptop, login freezes for a long time. It’s been 15+ minutes tonight I ran into this a few times and didn’t realize that’s what it was. Tonight I found a dir deep into /private/var that took 800GB according to du. I bet it won’t be there after reboot

System engineer looking for promotion: “hey, fs cleanup takes forever. What if we hide it behind next login? I bet no one will notice”

oh, no. oh *hell*, no. This is like reading the first paragraph of a @dereklowe.bsky.social's "Things I Won't Work With" article, one which ends with "and they were not wearing any safety gear"

rain 🌦️@sunshowers.io

Oh. The PHP thing is trying to do thread cancellation on OOM through longjmp. You never do thread cancellation, certainly not through longjmp, and especially not on OOM

Carlos Scheidegger reposted

shitposts aside, this one is serious, my loves: if you’re just waking up and you have any service online running PHP 5.4 to latest, you need to take it down RIGHT NOW and don’t put it back online until after it’s patched absolutely catastrophic vulnerability dropped in the middle of the night

Carlos Scheidegger reposted

Replying to @helenkennedy.com

Also, it's ghostwritten by Claude: "We didn't have to reject him. The calendar did." This kind of Surprising Turnaround Assigning Agency To An Inanimate Object is one of its most common tics.

github.comOld versions have been yanked? · Issue #1075 · zip-rs/zip2Just wanted to check this is correct - everything other than the latest version (v9.0.3) has been yanked on crates.io.
Carlos Scheidegger@cscheid.net

Can someone steelman the current notion/culture of yanking crates from cargo? Today's exasperation brought to you by wasm-pack being broken by someone deep in the dependency chain yanking some random zip2 crates from cargo. Whyyy

Can someone steelman the current notion/culture of yanking crates from cargo? Today's exasperation brought to you by wasm-pack being broken by someone deep in the dependency chain yanking some random zip2 crates from cargo. Whyyy

I'm only 30% of the way through or so, but one thing jumps to my mind: there wasn't one single breakthrough here. Instead, we got substantively better at: - tracking training dynamics - nn architectures - interpretability protocols (for lack of a better name) These interact positively together.

Ryan Moulton@moultano.bsky.social

Haven't read this yet, but I'm posting it because it seems important and I want to. already-clean.goodfire.com

View full thread

Ok, done reading Roughly, "if we examine modern nns using modern tools we have to build nns, it gets easier". That's good news, but (shock!) the oai guy is overselling it. Minimal counter: someone gives you only a pile of weights, says "this is aligned". "good news mechinterp" won't help you

In the limit, I think I want mechinterp to tell me the behavior of _any_ network. This piece argues convincingly that "good old mechinterp is fixable", but not "good old mechinterp lets us solve alignment", unless you have a pretty narrow definition of "solve alignment"

I'm only 30% of the way through or so, but one thing jumps to my mind: there wasn't one single breakthrough here. Instead, we got substantively better at: - tracking training dynamics - nn architectures - interpretability protocols (for lack of a better name) These interact positively together.

Ryan Moulton@moultano.bsky.social

Haven't read this yet, but I'm posting it because it seems important and I want to. already-clean.goodfire.com

This paragraph is interesting. It seems that "bigger, better, newer models train better *and* are easier to keep monosemantic" is a better explanation than "polysemanticity was never in InceptionV1". It's good news either way, but I wonder.

JD Long@jdlong.cerebralmastication.com

It’s awesome that GLM is a LLM model name and a generalized linear model. Because there was not enough confusion in the ML space.

Carlos Scheidegger reposted

No joke. I think this is the most important video I have ever seen about mobile games. This guy breaks down how this game psychologically profiles you, and makes the game harder or easier in order to show you ads

Carlos Scheidegger reposted

Virginia Vassilevska Williams wrote some thoughts about the 3-SUM breakthrough Josh Alman and her wrote an exposition of, how it happened, what it means, and some context about 'what could have been' in a world with more time and fewer models. Read it here: people.csail.mit.edu/virgi/though...

people.csail.mit.eduhttps://people.csail.mit.edu/virgi/thoughts.txtA week has passed since Josh and I posted our write-up on the arXiv, and I decided to write a few things. Much was said about our write-up. Some were excited, others enraged. Most everyone was scar...

I'm genuinely glad I read this and worked through all the stages of grief Before Present Discourse; the paper is easily on the pareto front of the (excellence, inspires-dread) diagram.

Erin Kissane@kissane.myatproto.social

Good morning, I had a terrible time reading this over breakfast