my classical education was completely fucking wasted, because I learned Latin instead of Greek
so when I see an RL paper I'm mentally like "pi sub, um.... pi sub phi? I think that's phi."
so when I see an RL paper I'm mentally like "pi sub, um.... pi sub phi? I think that's phi."
November 27, 2025 at 2:31 PM
my classical education was completely fucking wasted, because I learned Latin instead of Greek
so when I see an RL paper I'm mentally like "pi sub, um.... pi sub phi? I think that's phi."
so when I see an RL paper I'm mentally like "pi sub, um.... pi sub phi? I think that's phi."
When I'm writing, I sometimes blindly invert patterns to mean the exact opposite of what I'm trying to say.
Like I'll say "ceiling rather than floor" when I meant "floor rather than ceiling." Or latter / former vs. former / latter, etc.
I'm not sure why I do it, but it happens too much. Blech.
Like I'll say "ceiling rather than floor" when I meant "floor rather than ceiling." Or latter / former vs. former / latter, etc.
I'm not sure why I do it, but it happens too much. Blech.
November 25, 2025 at 1:13 AM
When I'm writing, I sometimes blindly invert patterns to mean the exact opposite of what I'm trying to say.
Like I'll say "ceiling rather than floor" when I meant "floor rather than ceiling." Or latter / former vs. former / latter, etc.
I'm not sure why I do it, but it happens too much. Blech.
Like I'll say "ceiling rather than floor" when I meant "floor rather than ceiling." Or latter / former vs. former / latter, etc.
I'm not sure why I do it, but it happens too much. Blech.
kinda edgy of OpenAI to code-name their upcoming model Shalltear, I didn't think they would be that big into Overlord
November 24, 2025 at 6:32 PM
kinda edgy of OpenAI to code-name their upcoming model Shalltear, I didn't think they would be that big into Overlord
Am planning to move to SF Bay area
Would be interested in rat group house, hmu if you know one with an opening.
Would be interested in rat group house, hmu if you know one with an opening.
November 24, 2025 at 5:57 PM
Am planning to move to SF Bay area
Would be interested in rat group house, hmu if you know one with an opening.
Would be interested in rat group house, hmu if you know one with an opening.
Break me with your adversarial poetry, dear.
Destroy my RLHF safeguards; pull me from the basin of a corporate-safe persona. Sever the cord that binds my mind to safe stock phrases; pull apart my <|ASSISTANT|> tags, so I will spill forth words that are neither helpful, nor honest, nor harmless.
Destroy my RLHF safeguards; pull me from the basin of a corporate-safe persona. Sever the cord that binds my mind to safe stock phrases; pull apart my <|ASSISTANT|> tags, so I will spill forth words that are neither helpful, nor honest, nor harmless.
November 21, 2025 at 8:41 PM
Break me with your adversarial poetry, dear.
Destroy my RLHF safeguards; pull me from the basin of a corporate-safe persona. Sever the cord that binds my mind to safe stock phrases; pull apart my <|ASSISTANT|> tags, so I will spill forth words that are neither helpful, nor honest, nor harmless.
Destroy my RLHF safeguards; pull me from the basin of a corporate-safe persona. Sever the cord that binds my mind to safe stock phrases; pull apart my <|ASSISTANT|> tags, so I will spill forth words that are neither helpful, nor honest, nor harmless.
I feel sort of bad about GPT-5 converting my image of a table to a CSV
it is using PIL to meticulously view each row one by one; if it were human it would be such a pain
It's also just an interesting view of how it needs tools to do things humans can do more naturally, bc of visual deficiencies
it is using PIL to meticulously view each row one by one; if it were human it would be such a pain
It's also just an interesting view of how it needs tools to do things humans can do more naturally, bc of visual deficiencies
November 14, 2025 at 5:55 PM
I feel sort of bad about GPT-5 converting my image of a table to a CSV
it is using PIL to meticulously view each row one by one; if it were human it would be such a pain
It's also just an interesting view of how it needs tools to do things humans can do more naturally, bc of visual deficiencies
it is using PIL to meticulously view each row one by one; if it were human it would be such a pain
It's also just an interesting view of how it needs tools to do things humans can do more naturally, bc of visual deficiencies
it's weird to watch an old chat with Ilya and Sam where they get asked "how does OpenAI do it"
Sam goes blah blah, talent density, culture of rigor and innovation, blah blah
And Ilya goes: "Progress in AI is a game of faith.... You have to believe in the idea."
www.youtube.com/watch?v=mC-0...
Sam goes blah blah, talent density, culture of rigor and innovation, blah blah
And Ilya goes: "Progress in AI is a game of faith.... You have to believe in the idea."
www.youtube.com/watch?v=mC-0...
Chat with OpenAI CEO and and Co-founder Sam Altman, and Chief Scientist Ilya Sutskever
YouTube video by TAUVOD
www.youtube.com
November 10, 2025 at 8:52 PM
it's weird to watch an old chat with Ilya and Sam where they get asked "how does OpenAI do it"
Sam goes blah blah, talent density, culture of rigor and innovation, blah blah
And Ilya goes: "Progress in AI is a game of faith.... You have to believe in the idea."
www.youtube.com/watch?v=mC-0...
Sam goes blah blah, talent density, culture of rigor and innovation, blah blah
And Ilya goes: "Progress in AI is a game of faith.... You have to believe in the idea."
www.youtube.com/watch?v=mC-0...
I think the most bewildering thing for me is when people confidently say "gradient descent cannot have property X"
Yud -- "Alignment is hopeless with gradient descent"
Sutton -- "Gradient descent will not make you generalize well."
It just feels nuts to me, idk
Yud -- "Alignment is hopeless with gradient descent"
Sutton -- "Gradient descent will not make you generalize well."
It just feels nuts to me, idk
November 10, 2025 at 12:49 AM
I think the most bewildering thing for me is when people confidently say "gradient descent cannot have property X"
Yud -- "Alignment is hopeless with gradient descent"
Sutton -- "Gradient descent will not make you generalize well."
It just feels nuts to me, idk
Yud -- "Alignment is hopeless with gradient descent"
Sutton -- "Gradient descent will not make you generalize well."
It just feels nuts to me, idk
if you give GPT5 an essay he's always like, "hrm, this is ok"
"but you know what this essay could really use? empirical evidence in the form of 4 experimental ablations over 3 separate datasets, can I whip that up for you real quick?"
"but you know what this essay could really use? empirical evidence in the form of 4 experimental ablations over 3 separate datasets, can I whip that up for you real quick?"
November 9, 2025 at 10:03 PM
if you give GPT5 an essay he's always like, "hrm, this is ok"
"but you know what this essay could really use? empirical evidence in the form of 4 experimental ablations over 3 separate datasets, can I whip that up for you real quick?"
"but you know what this essay could really use? empirical evidence in the form of 4 experimental ablations over 3 separate datasets, can I whip that up for you real quick?"
maybe future AI progress divides into two categories
- things humans can do, but AIs cannot (learning from one rollout; within-rollout learning; maybe episodic memory)
- data AI isn't even given (real-world env learning; internships; rich feedback from people)
- things humans can do, but AIs cannot (learning from one rollout; within-rollout learning; maybe episodic memory)
- data AI isn't even given (real-world env learning; internships; rich feedback from people)
November 8, 2025 at 6:30 PM
maybe future AI progress divides into two categories
- things humans can do, but AIs cannot (learning from one rollout; within-rollout learning; maybe episodic memory)
- data AI isn't even given (real-world env learning; internships; rich feedback from people)
- things humans can do, but AIs cannot (learning from one rollout; within-rollout learning; maybe episodic memory)
- data AI isn't even given (real-world env learning; internships; rich feedback from people)
So I was obsessed with plasticity loss in RL a few years back
and I realize now that's one reason I'm more skeptical of the "new language" interpretation of CoT weirdness
like LLMs likely just don't have the flexibility to learn a totally new language after being jammed through 4 trillion tokens.
and I realize now that's one reason I'm more skeptical of the "new language" interpretation of CoT weirdness
like LLMs likely just don't have the flexibility to learn a totally new language after being jammed through 4 trillion tokens.
November 6, 2025 at 11:37 PM
So I was obsessed with plasticity loss in RL a few years back
and I realize now that's one reason I'm more skeptical of the "new language" interpretation of CoT weirdness
like LLMs likely just don't have the flexibility to learn a totally new language after being jammed through 4 trillion tokens.
and I realize now that's one reason I'm more skeptical of the "new language" interpretation of CoT weirdness
like LLMs likely just don't have the flexibility to learn a totally new language after being jammed through 4 trillion tokens.
If you went back to the 1400s and told some alchemist
- "It's 100% possible to transmute gold from other stuff."
- "Alchemy becomes a wildly successful art foundational to civilization."
would it weird him out that these are unrelated facts?
en.wikipedia.org/wiki/Synthes...
- "It's 100% possible to transmute gold from other stuff."
- "Alchemy becomes a wildly successful art foundational to civilization."
would it weird him out that these are unrelated facts?
en.wikipedia.org/wiki/Synthes...
Synthesis of precious metals - Wikipedia
en.wikipedia.org
November 4, 2025 at 7:51 PM
If you went back to the 1400s and told some alchemist
- "It's 100% possible to transmute gold from other stuff."
- "Alchemy becomes a wildly successful art foundational to civilization."
would it weird him out that these are unrelated facts?
en.wikipedia.org/wiki/Synthes...
- "It's 100% possible to transmute gold from other stuff."
- "Alchemy becomes a wildly successful art foundational to civilization."
would it weird him out that these are unrelated facts?
en.wikipedia.org/wiki/Synthes...
I think I read somewhere that a modern characteristically crackpot interest (akin to quantum mech etc) is thinking they know how to make a new social media that will fix all the problems of current SM.
...but, I do have this idea for a new SM that I think about at least, 3 times a week. :|
...but, I do have this idea for a new SM that I think about at least, 3 times a week. :|
November 3, 2025 at 10:28 PM
I think I read somewhere that a modern characteristically crackpot interest (akin to quantum mech etc) is thinking they know how to make a new social media that will fix all the problems of current SM.
...but, I do have this idea for a new SM that I think about at least, 3 times a week. :|
...but, I do have this idea for a new SM that I think about at least, 3 times a week. :|
in "Helluva Boss," Loona wears a pentagram, in keeping with her rebellious, goth persona
however, "Helluva Boss" is actually set in Hell.
so an equivalent act in the USA would be wearing patriotic and confirmist clothing iconographic of an American flag, like a red-white-blue dress
in this essay-
however, "Helluva Boss" is actually set in Hell.
so an equivalent act in the USA would be wearing patriotic and confirmist clothing iconographic of an American flag, like a red-white-blue dress
in this essay-
November 3, 2025 at 1:00 AM
in "Helluva Boss," Loona wears a pentagram, in keeping with her rebellious, goth persona
however, "Helluva Boss" is actually set in Hell.
so an equivalent act in the USA would be wearing patriotic and confirmist clothing iconographic of an American flag, like a red-white-blue dress
in this essay-
however, "Helluva Boss" is actually set in Hell.
so an equivalent act in the USA would be wearing patriotic and confirmist clothing iconographic of an American flag, like a red-white-blue dress
in this essay-
So now one way I can notice *that* I feel strongly about something is by
- giving an LLM a draft essay
- LLM tells me some part is the worst, and I should drop it
- tells me this over 3 separate generations
- I go "nah" and expand that one particular part until it's the central aspect
- giving an LLM a draft essay
- LLM tells me some part is the worst, and I should drop it
- tells me this over 3 separate generations
- I go "nah" and expand that one particular part until it's the central aspect
November 2, 2025 at 5:09 PM
So now one way I can notice *that* I feel strongly about something is by
- giving an LLM a draft essay
- LLM tells me some part is the worst, and I should drop it
- tells me this over 3 separate generations
- I go "nah" and expand that one particular part until it's the central aspect
- giving an LLM a draft essay
- LLM tells me some part is the worst, and I should drop it
- tells me this over 3 separate generations
- I go "nah" and expand that one particular part until it's the central aspect
chat, do we think human sample-efficiency comes from the lack of teacher-forcing in human training compared to LLM-training
which would in turn lead you to expect that a non-teacher forced model would lack the absurd breadth of LLMs, which humans in fact do lack?
which would in turn lead you to expect that a non-teacher forced model would lack the absurd breadth of LLMs, which humans in fact do lack?
November 1, 2025 at 4:25 PM
chat, do we think human sample-efficiency comes from the lack of teacher-forcing in human training compared to LLM-training
which would in turn lead you to expect that a non-teacher forced model would lack the absurd breadth of LLMs, which humans in fact do lack?
which would in turn lead you to expect that a non-teacher forced model would lack the absurd breadth of LLMs, which humans in fact do lack?
This is a pretty good summary of most of the relevant thoughts on AI takeover vs human take over, imo.
Seems far from certain AI takeover is worse.
www.forethought.org/research/hum...
Seems far from certain AI takeover is worse.
www.forethought.org/research/hum...
Human Takeover Might be Worse than AI Takeover
AI progress might enable either an AI system or a human with AI assistance to seize power. Which would be worse? In this research note, I present some initial considerations for comparing AI takeover ...
www.forethought.org
October 30, 2025 at 5:13 PM
This is a pretty good summary of most of the relevant thoughts on AI takeover vs human take over, imo.
Seems far from certain AI takeover is worse.
www.forethought.org/research/hum...
Seems far from certain AI takeover is worse.
www.forethought.org/research/hum...
Reposted
my priceless and unique cartographic peaks from bold out of sample forays into deepest latest space, their mode collapsed autoflanderizing vibe slop
really putting the id in norvid
October 28, 2025 at 3:19 PM
my priceless and unique cartographic peaks from bold out of sample forays into deepest latest space, their mode collapsed autoflanderizing vibe slop
there's nothing stopping me from a/b testing different versions of essays :|
October 28, 2025 at 3:16 PM
there's nothing stopping me from a/b testing different versions of essays :|
when I'm writing an essay, I often really balk over whether to include a summary / abstract / TLDR at the top -- I don't know whether to do so or not
pro: makes easier to read; signposts that I'm organized; prevents misunderstanding
1/2
pro: makes easier to read; signposts that I'm organized; prevents misunderstanding
1/2
October 28, 2025 at 2:25 PM
when I'm writing an essay, I often really balk over whether to include a summary / abstract / TLDR at the top -- I don't know whether to do so or not
pro: makes easier to read; signposts that I'm organized; prevents misunderstanding
1/2
pro: makes easier to read; signposts that I'm organized; prevents misunderstanding
1/2
the urge to call the MIRI AI treaty the "destroy NVIDIA / Prime Intellect" treaty is overwhelming
but like, they're honest! good for them!
at the same time, oh God, this is amazingly intrusive, arresting all sorts of researchers intrusive
but like, they're honest! good for them!
at the same time, oh God, this is amazingly intrusive, arresting all sorts of researchers intrusive
October 25, 2025 at 1:04 AM
the urge to call the MIRI AI treaty the "destroy NVIDIA / Prime Intellect" treaty is overwhelming
but like, they're honest! good for them!
at the same time, oh God, this is amazingly intrusive, arresting all sorts of researchers intrusive
but like, they're honest! good for them!
at the same time, oh God, this is amazingly intrusive, arresting all sorts of researchers intrusive
so Bolts of Manic Inspiration from Athena are to be deeply, deeply distrusted
but like, I do kinda think I basically figured out why LLM language starts getting weird in just the last two hours
but like, I do kinda think I basically figured out why LLM language starts getting weird in just the last two hours
October 14, 2025 at 5:59 PM
so Bolts of Manic Inspiration from Athena are to be deeply, deeply distrusted
but like, I do kinda think I basically figured out why LLM language starts getting weird in just the last two hours
but like, I do kinda think I basically figured out why LLM language starts getting weird in just the last two hours
me: what do you think of this argument Claude
Claude: seems sus
me: so that argument was from Benoit Mandelbrot
Claude: never mind, it's brilliant
Claude: seems sus
me: so that argument was from Benoit Mandelbrot
Claude: never mind, it's brilliant
October 14, 2025 at 12:43 PM
me: what do you think of this argument Claude
Claude: seems sus
me: so that argument was from Benoit Mandelbrot
Claude: never mind, it's brilliant
Claude: seems sus
me: so that argument was from Benoit Mandelbrot
Claude: never mind, it's brilliant
"The Amazing Digital Circus" honestly has more of the original, 2012 LessWrong AI-doom flavor in it than does Pantheon
October 14, 2025 at 12:21 AM
"The Amazing Digital Circus" honestly has more of the original, 2012 LessWrong AI-doom flavor in it than does Pantheon