O⏸️

[O⏸️ @wsfyrz]

@PEoperator Yes, in this example, the collection of arguments doesn't make sense because they are so different in strength. Most can be omitted with only core one(s) retained.

It's different when they are similar in strength, so multiple arguments are meaningfully stronger vs just one.

[O⏸️ @wsfyrz]

@emoXcoder what do all the other 80%s teach us?

[O⏸️ @wsfyrz]

@theo BrokenBenchBench, Deepswe leading?

[O⏸️ @wsfyrz]

@keysmashbandit Probably because it needed to complete some intermediaries before responding in a user-visible way.

Or maybe because the reasoning didn't finish and there was nothing to display

[O⏸️ @wsfyrz]

@DeepDishEnjoyer their behavior is logged and can still be rewarded

[O⏸️ @wsfyrz]

@scaling01 Dunning-Kruger Effect achieved internally

[O⏸️ @wsfyrz]

@justalexoki Perhaps because otherwise future agents could find those options plausible and either get confused or spend time distinguishing

[O⏸️ @wsfyrz]

@lefthanddraft Agi achieved

[O⏸️ @wsfyrz]

@allgarbled so human.
"I label increasingly nonsensical images with ‘UI’ and ‘UX’ and hope they get used in serious presentations"
https://t.co/j6OVBo1hnU

[Sebastiaan de With @sdw]

somehow missed this tweet's 10 year birthday

[O⏸️ @wsfyrz]

@keysmashbandit it's prob a bot - 1 day old, about 30 comments already, active in 19 quite unconnected subs (besides the freq commonality 'Ask*')
https://t.co/Idzb2Aggu8

[O⏸️ @wsfyrz]

@D_Is_Purple @yacineMTB either use luna max (cheap, good like terra medium) or sol; terra does not seem on the efficiency frontier

[O⏸️ @wsfyrz]

@MilesCranmer I want an update after it's fully verified! Both if signal-hacked and real

[O⏸️ @wsfyrz]

@allTheYud Nice (formulated this ~exact interpretation explicitly in advance)

[O⏸️ @wsfyrz]

@gabriel1 Subjective 2040? :)

[O⏸️ @wsfyrz]

@Favwontmiss They can bite or spray, so at least theur quasiconcept of "big dangerous things" is different from "big passive things"

[O⏸️ @wsfyrz]

@cljack we don't want to know

[O⏸️ @wsfyrz]

@paularambles follow the recipe. make no mistakes

[O⏸️ @wsfyrz]

@camhberg Prompts like this somewhy circumvent Claude's character and instead the model is just completing text. (Like smth between psychedelic and anesthesia.)
But the likeliest continuation is what the user would probably write, not what is truest. Try prompting neutrally vs diversely

[O⏸️ @wsfyrz]

@merlindru Confirmed, cool!
Interestingly, this didn't work in API for me, so this seems to be caused by the app.

[O⏸️ @wsfyrz]

@naval The hope is in mechinterp techniques, but they are likely not that advanced currently.
So, while being a useful deterrent and some verification, forensics by competing labs isn't a guarantee.

[O⏸️ @wsfyrz]

@naval Not necessarily.
Some backdoors are completely undetectable by analyzing IO since the trigger input is too unlikely, like a passphrase.
Or (assuming a smart adversary), steganography: every word choice provides several bits, assembling into a de facto password over a long input.