Crabbix

I play league of legends poorly and swear a lot.

[DPRK News Service @DPRK_News]

Rabid boar humanely euthanized in North Hamgyong province.

[Crabbix @CrabbixTV]

@420joke @thejoeydavis Lava Glaciers still bangs

[Crabbix @CrabbixTV]

@NoahChrein The research discipline of 'mechanistic interpretability' does in fact exist. I guess 'why aren't more maths/physics types flooding into it' is still a fair question but I think the answer is just that we're still early

[Crabbix @CrabbixTV]

@BrainsAndTennis Also, 'most problems don't need more than 2 layers of decomposition' is obviously untrue for sufficiently large problems. 'invent the LLM starting from 16th century math and tech' is a solved problem that involved dozens of layers of human 'subagents'

[Crabbix @CrabbixTV]

@BrainsAndTennis I don't think you can call it an RLM if it doesn't recur. At depth 1 your subagent is not the same as your main agent cause it can't call subagents, so it's not a recursion. That said, it seems like "task decomp is hard" is gonna get swallowed by RL task decomp and RLMs just work

[Crabbix @CrabbixTV]

@max_spero_ Given this premise, which I wholeheartedly agree with, a very interesting question is: how much do category 3 tasks actually matter? Like, maybe 1 and 2 can get you interstellar travel but world peace/generally utopian outcomes actually require a lot of 3?

[Crabbix @CrabbixTV]

@lucastohdev I'd like a soft usage limit per chat session, if I kick off a task that I'm expecting to use 2-3% of my week I don't want to come back 6 hours later to find it's been stuck in an infinite loop and burned 30% (has happened a couple times with Sol + underspecified goals)

[Crabbix @CrabbixTV]

@voooooogel

[Crabbix @CrabbixTV]

@doomslide If you find an opaque counterexample to a notorious conjecture, you just create a new open problem "why does this class of counterexample exist". You can't kill maths by making progress

[Crabbix @CrabbixTV]

@celestepoasts I think the only remotely convincing steelman is that this is good *despite* the real harms of open cyber capability. That those harms are outweighed by this basically being a hedge against concentration of power in closed labs, and that glasswing ameliorates those harms somewhat

[Crabbix @CrabbixTV]

@andrewmccalip Claude reverse engineer this guy's circadian rhythm/humanness detection stack and build me 1000 humanlike bots make no mistakes ultrathink

[Crabbix @CrabbixTV]

@macksypowwa You mean the doojig?

[Crabbix @CrabbixTV]

@eusexuant It's like a light 6 if we're being objective, Ego and Lonely is the Muse are strong, continuing to respond to Fantano at this point makes her look pathetic. This is the only correct opinion

[Crabbix @CrabbixTV]

Ceci n'est pas une pipe

[field encounters @nickmorgenbach]

The good thing about AI will be that we will finally, rightfully, distrust all images as fake, which they always were and forever will be

[Crabbix @CrabbixTV]

@quasicoh I tried doing this hackily with o1-mini when it came out and it was a cool toy but not reliable enough to really do anything useful

[Crabbix @CrabbixTV]

@seconds_0 >he thinks mythos is gonna be available on the API

[Crabbix @CrabbixTV]

@jltgonmyphone Amy has beaten the not having a passport allegations 🗣️

[Crabbix @CrabbixTV]

@nicole_clash Hit Master most sets, work full time on (voice) agent development and have built a benchmark for fun before, interested in contributing.

https://t.co/CJOKIjYspA

https://t.co/5nhgVdYOQO

[Crabbix @CrabbixTV]

@extramicrowave Buy puts lol

[Crabbix @CrabbixTV]

@aliceisplaying I think we should start calling kids epsilons. We already call people alphas betas and sigmas