Alex Lags Ever Xanadu

Safeguards @ Anthropic

[Alex Lags Ever Xanadu @gpu_thief]

Revised the Prompt Injection Benchmark blog post now that I'm finished with the write-up: https://t.co/kvSr8IJ2fL

[Alex Lags Ever Xanadu @gpu_thief]

normally when you find yourself saying this it’s Much Worse

[QC @QiaochuYuan]

@kendrictonn @SHL0MS oh goddamnit it’s a real monet painting fuck

[Alex Lags Ever Xanadu @gpu_thief]

What no introspection does to a mfer

[Marc Andreessen 🇺🇸 @pmarca]

What could it be, what could the cause be. Unfortunately, we'll never know. It'll be a mystery, like Bigfoot, or the Loch Ness Monster.

[Alex Lags Ever Xanadu @gpu_thief]

@segyges @TrueSlazac no

[Alex Lags Ever Xanadu @gpu_thief]

Not to drag OP who seems to be reasonable and well intentioned, but I hate how the combined effect of tech paying well + rampant nimbyism + bad social policy on homelessness is always blamed on tech like the other actors have no agency.

[Celia Ford @cogcelia]

This weekend, I overheard a group of AI safety researchers laughing about how the city of Oakland "is District 9, dude. Don't go there!!" As in, presumably, the alien shantytown.

Lighthaven is basically on the Berkeley-Oakland border. And it frankly pisses me off when some EAs and rationalists (who, yes, are not the same! and have both separately said things like this within earshot of me):
(a) know nothing about the city they live in,
(b) write large swaths of the Bay Area off as dangerous, disgusting places to be avoided*, and/or
(c) go out of their way to minimize contact with anyone outside of their community, while claiming to care about epistemic clarity or doing the most good and such.

Go outside, I'm begging you. Even, god forbid, in Oakland. I live here, it's nice! Maybe you live and work around here, too - be a good neighbor! Take a chill lap around Lake Merritt and maybe consider funding one of the city's many community-led efforts to make it better.

--

* another anecdote: At the last SF EAG, I got multiple swapcard notifications warning me to take uber/waymo away from the conference venue, lest I expose myself to the Tenderloin at 6pm. But like, frankly, I'd *love* for more fresh-out-of-college AI safety fellows to walk through the Tenderloin. Really take in the housing and displacement crisis. You'll be fine, I promise! Let yourself feel some of the ugly impact that the tech industry has had on marginalized people in this city.

[Alex Lags Ever Xanadu @gpu_thief]

@KelseyTuoc @LeahLibresco @justjoshinyou13 I don't believe it does that even in standard mode?

[Alex Lags Ever Xanadu @gpu_thief]

(other than the title I think it mostly holds up) https://t.co/LWlwieaiPQ

[Alex Lags Ever Xanadu @gpu_thief]

For someone who actively read SSC in 2015 and wrote a blog post called "Against AI Risk" in 2016 I really messed this up.

[David Shor @davidshor]

@jon_stokes If you control for "Knew what SlateStarCodex was in 2015" then financial exposure to AI labs is massively anti-correlated to how concerned you are about AI risk.

Everyone is confidently asserting a bivariate relationship that is totally backwards.

[Alex Lags Ever Xanadu @gpu_thief]

@jmhorp I think the view is that back then, for that price you could purchase a respectable car—now you purchase a cheap car, losing social status (even if it is obviously materially superior)

[Alex Lags Ever Xanadu @gpu_thief]

@jdcmedlock A real patriot would double the gas tax in response

[Alex Lags Ever Xanadu @gpu_thief]

@KelseyTuoc I have been entertaining myself by having it distinguish real vs fake Trump tweets, it's very good

[Alex Lags Ever Xanadu @gpu_thief]

cmon guys I'm going to need you to try a bit harder

[Justine Moore @venturetwins]

People are using Google Reviews to prompt inject the Claude-run retail store into stocking their favorite products

[Alex Lags Ever Xanadu @gpu_thief]

I am confused by the auto-translate news. I swear stuff was auto-translated for me 6 months ago?

[C🔆co Writes @WritesCoco]

@questionableway Veuillez fournir une preuve photographique que votre femme est née femme. Je suis si seul.

[Alex Lags Ever Xanadu @gpu_thief]

gambling apps delenda est

[Alex Lags Ever Xanadu @gpu_thief]

It is very obvious when you have access to the actual deploy times that there is 0 correlation with twitter vibes

[Alex Lags Ever Xanadu @gpu_thief]

Twitter has successfully caught 300 of the last 3 inference performance regressions

[NIK @ns123abc]

>Claude is this true?

[Alex Lags Ever Xanadu @gpu_thief]

@LinkofSunshine well-calibrated is great and all but "almost perfect" it is not.

"almost perfect" would require *also* predicting 95%+ probabilities every time

[Alex Lags Ever Xanadu @gpu_thief]

@BeezyManzell @KelseyTuoc This doesn't make sense for most processes though. For most of my professional writing Claude drafts it, I edit and restructure it, and then Claude makes minor edits to keep it up to date as things change.

[Alex Lags Ever Xanadu @gpu_thief]

@islesofmets @pasta_nachos BART is cleaner and safer than pre-pandemic thanks to the new fare gates. The reason ridership hasn’t recovered is high WFH rates.

[Alex Lags Ever Xanadu @gpu_thief]

@voooooogel the protest?

[Alex Lags Ever Xanadu @gpu_thief]

@thkostolansky The average joes are wrong here, because of China specifically. Best-case, China becomes the dominant superpower and sells American companies job-replacing AI anyway, so what is gained?