conputer dipshit

not sure this website will be able to survive this discourse

Are We Under-Bubbled? Why the future needs more people willing to be duped

there’s also a “bubble are good, actually” take

there are also high-level constraints like power. there’s only so fast you increase the amount of power going to these things. I think we’ll have plenty of more ordinary problems in the meantime

The Shape of AI: Jaggedness, Bottlenecks and Salients And why Nano Banana Pro is such a big deal

a slightly more optimistic view is that when intelligence stops being the bottleneck on some thing, you are left with some other bottleneck

I think there used to be more people who took bubble to mean it pops and goes away, like crypto appeared to ( though it didn’t really go away). there are fewer such people now

the 2026 capex numbers are definitely a little alarming

more importantly with the cost of inference falling so fast comstantly, today’s intelligence level would be a lot cheaper to serve in a year. so if they were subsidizing and wanted to stop, they could phase it out and no one would notice (and they may already have)

I wish they would release clearer numbers but personally I think API prices are well above break-even on a marginal basis at least. though it’s possible the fixed costs are so huge they can’t be recouped

they’re going to tell you you only think you’re doing this. nope! I’ve seen it too

we are so inclined toward either the tiny micro level or global macro level of analysis that we barely even see the mesoscale phenomena that may well determine most social dynamics

on this, part of the facile thinking I am trying to push against is the idea that there‘s no middle between the singularity and it being all bullshit. it’s much harder to think about how things really are

Building pro-worker AI | Brookings Daron Acemoglu, David Autor, and Simon Johnson ask: What is pro-worker AI, and how can we build it?

full paper available here

huh! how did nobody tell me about this one

I don’t know, I wouldn‘t say they change completely every model gen. and even a 6 month out of date understanding has to be better than a 2 year old understanding if you’re trying to write laws

obviously I should link the bill
www.nysenate.gov/legislation/...

something else I’m curious about: these regulations would almost certainly only apply to hosted models at the big providers. there’s no way they could stop people from using open models locally or on rented cloud GPUs, right? so these restrictions serve to heavily incentivize the latter

this is also the first one of these I’ve heard of that wasn’t 4o

this is the kind of thing we need to figure out asap

I guess I aspire to be some kind of practical intellectual and I feel like I'm in an interesting position to theorize the present because I am on the one hand a software engineer and on the other hand a great books philosophy guy and on the other other hand some kind of marxist

like, I think we're going to need a bunch of different laws like this and it would be better if they were informed by a better understanding of the activity of using LLMs

I think the most likely outcome is we muddle through between the two extremes and the actual course we take could depend on people getting wise 3 or 6 months sooner. I agree it’s possible that it doesn’t really matter

jesus

I'm not saying Gemini _has_ CPTSD in any true sense, but it certainly does display all the telltale signs of someone with CPTSD spiraling into self loathing/crashing out when it fails to accomplish a goal

We Sat Down With The First Solitaire App For Heterosexual Men, And Now? We're True Believers.

By StraightFlush Staff

The world of heterosexual mobile card playing broadens its horizons with a smartly produced, intuitive experience that takes all the strangeness out of playing with yourself.

is this equivalent to `diff-formatter = ":git"` ? I use that anyway to make delta work as the pager

jj (Jujutsu) NEVER use git unless jj has no way to do the thing. Always use jj. jj status, jj diff, jj diff -r @-, jj log, etc. to view a file at a revision, use jj file show <path> -r <rev> (not jj cat). to exclude paths from a jj command, use fileset syntax: jj diff '~dir1 & ~dir2' or jj restore '~package-lock.json' prefer squash workflow in jj over editing, where if you're trying to update rev A, work in a rev on top of A and periodically squash what you've done into A for parallel approaches, use jj new <base> to create siblings from a common base, implement each approach, then compare. bookmarks are unnecessary for this workflow use jjw to manage jj workspaces: jjw create (or jjw c) creates a workspace and cds into it, jjw ls lists workspaces, jjw rm interactively removes one Non-destructive jj operations are generally allowlisted. When working on a complex change, use jj new or jj commit (equiv do jj desc + jj new) after chunks of work to snapshot each step in a reviewable way when using jj squash, avoid the editor popup with -m 'msg' or -u to keep the destination message. These flags are mutually exclusive. don't try to run destructive jj ops like squash or abandon unprompted. intermediate commits are fine; just note when cleanup might be needed --ignore-immutable may be needed when abandoning divergent commits from other authors, e.g., after rebasing on their branch and force pushing jj dt and jj dts are custom aliases that diff a rev against its fork point from trunk (like a GitHub PR diff). jj dt shows the full diff, jj dts shows --stat. Both default to @ but accept an optional rev argument.

hm I might have too much in there. but most are idiosyncrasies of how I want to work or fixes for common mistakes it makes
github.com/david-crespo...

I learned about a cool way to mitigate this from Joe Weisenthal of all people
xcancel.com/TheStalwart/...

🚨New Paper!🚨 How do reasoning LLMs handle inferences that have no deterministic answer? We find that they diverge from humans in some significant ways, and fail to reflect human uncertainty… 🧵(1/10)

I am finding it very useful so I find it plausible that others are. the amount we're spending per user is a lot, yet still a very small proportion of our salaries

the reason to think this is that if you zoom into the current superusers, a year ago they looked exactly like the population at large: probably a power law distribution with small proportion of superusers and a long tail. now they're all superusers