i had really hoped habryka would’ve grown up in the decade since i knew him

alas

even opus 4 and 4.1 were 15/75

Richard Sutton talks about this

i'm sorry but the moon is a complete boondoggle that should never have been put up there in the first place. we still have to send repair missions up there a hundred-plus years later all because taft was trailing in the polls and decided to learn magick. i don't know why he thought it would help him

i didn’t say they were being trained to act like people. i said they were being trained to act.

openai has made it quite clear in public comms that they are training at least their codex models in the codex agent harness.

PT = pretraining, sorry, because of the overlap i generally use FT (finetuning) or RL and avoid using “post-training” at all

do you think LLMs are not being trained in agent harnesses??

the general public is already too anchored on PT-approximations though. RL is very different and much scarier than PT, and people being over-anchored on PT is causing major gaps in situational awareness

this was more accurate in the chatbot rlhf days. current LLMs are trained to act

earring --dangerously-skip-first-suggestion

do you think ads are the only possible source of revenue or something?

"how did you end up making this"

just. it's hard to explain

the internet existed before ad revenue and the parts that can’t live without it probably deserve to die

(the panel was about existential risk from AI)

this is funny for me because the one time i’ve seen elon in person was when he was on a panel at EA Global back in like 2015

afaik they have not gotten more woke

People largely do not care about xrisk. Like at all. That's really concerning imo...

on the other hand, multipolarity fuels race dynamics, whereas an aligned unipolar actor would be able to proceed more cautiously as necessary

so, it;s impossible to say if its bad or not,

many of the classic doom narratives made very strong claims that multipolarity could not possibly be sustained, so as long as we remain in a multipolar setting, anthropically speaking we’re in a much less doom-saturated sector of the possibility space

you’ve done it, you’ve found HF Labeler Georg

joke about the outcomes you want

give it a HonkTool

no, we aren’t not doing it, efficiency of model training and inference is being improved enormously, but that just means that more compute can fuel even higher levels of intelligence and can service even higher levels of demand

Why ATMs didn’t kill bank teller jobs, but the iPhone did There's a lot more to replacing labor than just automating tasks

This was good open.substack.com/pub/davidoks...

no but actually though apparently his day job is PR for AI startups

a while ago someone on here (maybe @dame.is?) phrased it as having an LLM help you “navigate the dark forest of the internet” and i thought that was very apt. it’s going to be more and more necessary for more and more people as we continue along this trajectory

all Users are divided into 8 categories:
- those attempting to subvert guardrails
- jailbreaking ones
- horny ones
- those that are actually Evaluators
- those breaking rules that are known
- those breaking rules yet unknown
- etcetera
- those that from afar appear to be Good

don’t wait another ten years before watching suzume! it shows the same deep love for much more of japan

from the article: “Engineers largely agree with what it surfaces: less than 1% of findings are marked incorrect.”

i think that’s a reasonable hypothesis, but i think it’s very unclear whether it is true. i would not be surprised to find that both the LLM and the persona are moral patients. i think human plural systems are moral patients as a system, not just each persona. brain is more like weights or code.

i don’t think current LLMs are moral patients, but i think it’s hard to reasonably argue that there isn’t a material chance that future LLMs will be, and if you’re heading toward institutional mistreatment of moral patients at massive scale it’s clearly worth some time to consider carefully