sometimelurker

6 ms·

91.
by sometimelurker 4mo ago
no they 100% use MTP with a cheaper model alongside opus, and it would infact be unprovable if they just sometimes switched to auto-accepting everything from the MTP. its true that if they did anthropic would need to hide that they do this,
92.
by sometimelurker 4mo ago
no it might. a high reasoning task is probably harder than a low reasoning task, so the same MTP LLM will predict more correct tokens on the low reasoning task. to compensate for this, big labs likely have different MTP LLMs for different c
93.
by sometimelurker 4mo ago
they use multitoken prediction behind the scenes, that might interact with the CoT in a strange way. maybe for different thinking modes they have different MTP models? if so thats interesting
94.
by sometimelurker 4mo ago
I really freaked out once I stopped seeing AI gen videos on those scrollyapps, that fear is what got me off them. (the videos got so good I couldn't tell if they were real or not)
95.
by sometimelurker 4mo ago
> easy(ish) to detect 100% on small models, but frontier models (at the level ddeepseekv4pro) can tell when their being tested so it becomes harder to check. you can always finetune them to remove CCP propaganda from them
96.
by sometimelurker 4mo ago
you can finetune the ccp propaganda out of them, then your mostly fine. if you want to be more safe you can finetune their public base models to not have ccp propagnada, and then proceed with the rest of the training (costs more tho)
97.
by sometimelurker 5mo ago
cant wait til someone figures how to run Linux on one of these
98.
by sometimelurker 5mo ago
reminds me of kolmogorov complexity
99.
by sometimelurker 5mo ago
I like AI but I'd like it to be a different ratio of non-ai/AI posts here bc its a lot and kinda irritating by now
100.
by sometimelurker 5mo ago
AI arguably yes (specifically modern really big LLMs)
101.
by sometimelurker 5mo ago
thank you for maintaining LessWrong
102.
by sometimelurker 5mo ago
you can actually learn a lot about the world by just guessing whats happening. you don't even need to read the news, you can just guess correctly.
103.
by sometimelurker 5mo ago
which is reasonable (everyone watches screens) and if they want to communicate this to kids who watch too many screens, is there even a better way?
104.
by sometimelurker 5mo ago
maybe markets can be intelligent, and if so, why not 'weights'? btw love this piece, thank you
105.
by sometimelurker 5mo ago
Persepolis was really good, read it a few years ago. really recommended
106.
by sometimelurker 5mo ago
llama.cpp + llama-serve, I use it. if you like terminal, pi
107.
by sometimelurker 5mo ago
i heard that a lot of anti-ai stuff is written by AI because the people reading it aren't prepared to recognize if it was AI-gen because they themselves don't use ai
108.
by sometimelurker 5mo ago
not long ago I asked a bunch of people about this and they were totally fine talking to LLMs about how they feel. blew my mind, and now the 6/10 figure in the headline seems believable. there should be a push to educate people about th
109.
by sometimelurker 5mo ago
AI safety cant be automated so this claim isn't 100% true
110.
by sometimelurker 5mo ago
maybe windows should actually just become linux
111.
by sometimelurker 5mo ago
uh... in a very literal sense upvoting this helps Ellison. I think
112.
by sometimelurker 5mo ago
> Disney in this context, 'Disney' represents a plurality, and it's likely that there's people at Disney that want their kids off screens
113.
by sometimelurker 5mo ago
intuition says this isn't hard with SAEs. find the feature that corresponds to appscript code and train against it.
114.
by sometimelurker 5mo ago
I think the navigators were also gmo'ed a bit
115.
by sometimelurker 5mo ago
I said this same message on another sama quote "article' ( https://news.ycombinator.com/item?id=48280333 ) and it remains relevant here: I don't see any benefit to quoting sama. I'm trying to think of cont
116.
by sometimelurker 5mo ago
I think he knows this is too vague and unimplementable, he just wants to shift the conversation to more UBI-ish things.
117.
by sometimelurker 5mo ago
adding Leo XIV to that list would be pretty funny
118.
by sometimelurker 5mo ago
> An LLM could try to do something along the lines if either it gets fine-tuned to do it, or somebody instructs it to do it. long horizon RL teaches LLMs behaviors that incentize power-seeking and lying and other unethical actions to ach
119.
by sometimelurker 5mo ago
> AI psychosis I think most of us have an acute sort of it > next token predictor model (nondeterministic) next token predictor = capable of outputing any text if MLPs optimize over the space of functions, and sequence to sequence mod
120.
by sometimelurker 5mo ago
is meta low on $? why would they do this?

More ›