conputer dipshit

Text to me: What are ou going Doing Text from me: sorry. dancing

In my headcanon, The Barbican killed
Leon Krier, don’t try to convince me otherwise

increpare releases a game → I buy it

Surveillance is not a purely philosophical issue with purely esoteric harms. Surveillance is a sensitive and often faulty trigger that summons armed agents of the state with impunity that can take away a person's freedom.

unfortunately I was born to be this kind of nuancebro. I am just becoming more myself

should have asked what people think of ChatGPT too. people hate "big tech" but when you ask about individual companies they love them all
xcancel.com/davidshor/st...

YouTube video by jeff mohlenbrok Aldous Harding - Lawn - Live - Webster Hall NYC 06/09/2022

she is committed to the bit. enjoy the tambourine dance at the end

when claude is good enough, "just a claude agent SDK wrapper" can do a whole lot!

No, it doesn't cost Anthropic $5k per Claude Code user The viral claim that Anthropic loses $5,000 per Claude Code subscriber doesn't survive basic scrutiny. Let's do the actual maths.

heh, did you catch this? I don't think it's the most rigorous way to make the argument but I agree with the direction — I think people are way overstating the amount of subsidy
news.ycombinator.com/item?id=4731...

I agree however that breaking things up into better pieces is an important part of dealing with this

a beautiful old greenish-gray stone building on the upenn campus. statue of ben franklin in foreground

and what a building!

Trump Antisemitism Inquiry Demanding List of Jews at Penn Heads to Court

what a headline

this doesn't solve the problem because even if you break them up into more reviewable pieces, we're still producing more of them, plus they can interact

this is what I was getting at here — there are only returns to someone spending more time to review code if they are uniquely well positioned to review that exact code. I think that describes fewer actual code reviewers than we would like

I think it depends a lot on the quality of review you get for that $30. even at $100k, $25 is 30 minutes of dev time (generously). I don't find it hard to imagine the bot review being significantly better than 30 min of $100k dev review time. it's probably better than 30 min of my time

I know you're kidding, but that is exactly what this is doing for the $20

the relevant unit isn't really the individual PR, it's aggregate bugs found. you'd pay $30 per PR because you don't know which ones will turn up bugs and which won't. if you knew a PR didn't have bugs you would just merge it

right, exactly. $30 is like 10 minutes of my time and what's the likelihood I have the exact expertise to review your PR

well I‘m sure part of it is checking initial positives against the code and running tests and stuff

yeah their innovation here is doing it 10 times lol

Again Zitron—to his credit—accurately describes the situation, but doesn’t realize he’s giving away the game entirely if he admits that people are choosing to pay for higher quality. Because model developers hit a wall of diminishing returns, and the only way to make their models do more was to make them burn more tokens to generate a more accurate response (this is a very simple way of describing reasoning, a thing that OpenAI launched in September 2024 and others followed). As a result, all the “gains” from “powerful new models” come from burning more and more tokens.

this is why I say here Zitron is giving the game away when he says oh, people are only getting better results by paying more for them. that’s the whole thing working as well as it could!
crespo.business/posts/cost-o...

YouTube video by KEXP Aldous Harding - Full Performance (Live on KEXP)

really hope this round of touring gets us some high-quality live studio recordings. this is 6 years old. the world deserves more

YouTube video by Aldous Harding Aldous Harding - The Barrel (Official Video)

spotify listener counts suggest not enough of you know who Aldous Harding is

wholesome midwest content: my 7yo reading about all the native plants in the Prairie Moon Nursery catalog www.prairiemoon.com

I don't think people have grasped yet that if you were really confident the tokens were good, you would never stop jamming money in the machine

same reason successful companies grow: they're machines that turn labor and money into more money. add more labor and more money and you get more money

can't believe you didn't think of it yourself. too perfect

anyone can make an LLM barf out tokens — what’s hard is useful tokens. benchmarks like ARG-AGI 2 show the effect of more tokens. note the S-shapes: you have gains for a while, but eventually more tokens don’t help. the smarter the model is, the farther more tokens take you
arcprize.org/leaderboard

they might skim over it for egregious errors or confusing explanations, but there are rapidly diminishing returns to more time because their expertise does not make those extra minutes useful. but a UI expert can spend hours reviewing and still turn up problems and worthwhile considerations

it's clarifying to compare this to people working in or out of their area of expertise. say I have a pretty big UI change — if I ask someone who is not a UI dev to review it, how much time is it worth it for them to spend on that? probably not much