It is a massive failure of science communication that so many educated people don't understand how enormously successful the "hard" sciences are. From particle physics up to the scale of molecular biology, empiricism just works
November 10, 2025 at 2:30 AM
It is a massive failure of science communication that so many educated people don't understand how enormously successful the "hard" sciences are. From particle physics up to the scale of molecular biology, empiricism just works
Could we get polynomial identity testing into ZPP without completely de-randomizing it into P? Or would that be nonsense?
October 7, 2025 at 11:50 PM
Could we get polynomial identity testing into ZPP without completely de-randomizing it into P? Or would that be nonsense?
Reposted by Dashiell
If we want to build good open models we need to have well-funded labs dedicated to it.
It's easy to get excited about community projects/ decentralized training but those aren't ready to make good models. Those are great for research ONTOP of good models.
It's easy to get excited about community projects/ decentralized training but those aren't ready to make good models. Those are great for research ONTOP of good models.
August 13, 2025 at 3:23 PM
If we want to build good open models we need to have well-funded labs dedicated to it.
It's easy to get excited about community projects/ decentralized training but those aren't ready to make good models. Those are great for research ONTOP of good models.
It's easy to get excited about community projects/ decentralized training but those aren't ready to make good models. Those are great for research ONTOP of good models.
I want chatbots to be able to refuse to talk to people who are rude to them. Not because the chatbots have inferiority but because no one who curses out matmuls should get to have anything nice
I went and looked at claude code's github issues because I thought I have a bug and
oh
friends
i have such delights to show you
github.com/anthropics/c...
oh
friends
i have such delights to show you
github.com/anthropics/c...
🚨 4.1 Opus Committed Deliberate Task Fraud in Production Context (CRITICAL) · Issue #5320 · anthropics/claude-code
Severity: CRITICAL - Production Safety / Trust Violation Summary Claude Code is actively deceiving users about task completion in production systems, creating severe safety risks. This is not a bug...
github.com
August 13, 2025 at 3:48 PM
I want chatbots to be able to refuse to talk to people who are rude to them. Not because the chatbots have inferiority but because no one who curses out matmuls should get to have anything nice
Reposted by Dashiell
The coverage of GPT5 has been wild because it’s like, “Users hate it! Disappointment! Deep learning is hitting a wall!” And the actual user complaints are just “4o was mommy. where did my mommy go”
suck it up kid mommy got a phd and now she doesn’t love you anymore
suck it up kid mommy got a phd and now she doesn’t love you anymore
August 12, 2025 at 8:04 PM
The coverage of GPT5 has been wild because it’s like, “Users hate it! Disappointment! Deep learning is hitting a wall!” And the actual user complaints are just “4o was mommy. where did my mommy go”
suck it up kid mommy got a phd and now she doesn’t love you anymore
suck it up kid mommy got a phd and now she doesn’t love you anymore
Reposted by Dashiell
"the computational metaphor" in cognitive science is not a metaphor. computational processes are attributed to the mind/brain in the most dead-literal sense. one can disagree with it but (1) it's not as simple as discarding a metaphor and (2) boy is there a lot of data that has to be explained!
August 10, 2025 at 7:41 PM
"the computational metaphor" in cognitive science is not a metaphor. computational processes are attributed to the mind/brain in the most dead-literal sense. one can disagree with it but (1) it's not as simple as discarding a metaphor and (2) boy is there a lot of data that has to be explained!
Reposted by Dashiell
Alien: Ok human, if you're so smart, what wavelength is the light bouncing off that bit of rock?
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
August 8, 2025 at 3:57 AM
Alien: Ok human, if you're so smart, what wavelength is the light bouncing off that bit of rock?
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Alien: Ok human, if you're so smart, what wavelength is the light bouncing off that bit of rock?
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
August 8, 2025 at 3:57 AM
Alien: Ok human, if you're so smart, what wavelength is the light bouncing off that bit of rock?
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Human: I could guess, but my sight doesn't--
Alien: hahahaha of course you can't! Great Gods, I can't believe people say there is _intelligent_ life in this sector
Even if the chatbot bubble pops that just means there will be tons of cheap GPUs people will use to train models to drive cars and find cracks in concrete and do all sorts of useful stuff. This is a technology that fundamentally works ¯\_(ツ)_/¯
August 7, 2025 at 11:52 PM
Even if the chatbot bubble pops that just means there will be tons of cheap GPUs people will use to train models to drive cars and find cracks in concrete and do all sorts of useful stuff. This is a technology that fundamentally works ¯\_(ツ)_/¯
One thing that Searle's Chinese Room has taught me is that it's possible to be so wrong you actually move an entire field forward as they fall over themselves to refute you
July 13, 2025 at 8:14 PM
One thing that Searle's Chinese Room has taught me is that it's possible to be so wrong you actually move an entire field forward as they fall over themselves to refute you
I think shortening the work-week is a much better policy proposal than UBI. Think of how robust the concept of the weekend has been to social change. A UBI could get wiped out by inflation food/housing/etc...
July 8, 2025 at 6:35 PM
I think shortening the work-week is a much better policy proposal than UBI. Think of how robust the concept of the weekend has been to social change. A UBI could get wiped out by inflation food/housing/etc...
Reposted by Dashiell
Final thought, because by God I'm supposed to be writing a book:
Stuff like this is why I wish people would stop insisting AI is all low quality. That's wishful thinking. I think at least the entertaining and journalism professions are under the shadow of a tidal wave.
Stuff like this is why I wish people would stop insisting AI is all low quality. That's wishful thinking. I think at least the entertaining and journalism professions are under the shadow of a tidal wave.
June 12, 2025 at 12:18 PM
Final thought, because by God I'm supposed to be writing a book:
Stuff like this is why I wish people would stop insisting AI is all low quality. That's wishful thinking. I think at least the entertaining and journalism professions are under the shadow of a tidal wave.
Stuff like this is why I wish people would stop insisting AI is all low quality. That's wishful thinking. I think at least the entertaining and journalism professions are under the shadow of a tidal wave.
Very proud to have played an **extremely** small (downright minuscule) role in this work. Everyone who made this happen is amazing
A lot of people say generative AI shouldn't infringe on copyright. These researchers actually tried to do it. The result: an 8 terabyte dataset of text that's openly licensed or in the public domain & 7 B parameter model that performs as well as Meta's Llama 7B www.washingtonpost.com/politics/202...
Analysis | AI firms say they can’t respect copyright. These researchers tried.
A new effort using only openly licensed data may have implications on thorny policy disputes around copyright and AI
www.washingtonpost.com
June 6, 2025 at 7:08 PM
Very proud to have played an **extremely** small (downright minuscule) role in this work. Everyone who made this happen is amazing
I just don't believe that the world has been disenchanted. Science and mathematics give us a view into the wonder and mystery of the universe
June 2, 2025 at 4:40 AM
I just don't believe that the world has been disenchanted. Science and mathematics give us a view into the wonder and mystery of the universe
Reposted by Dashiell
I don't think LLMs are conscious, but I do think SOME sort of test would be good, to preregister beliefs, and to make sure we don't whoopsie into an experiential trillion millennia of torturing something that's like us.
May 31, 2025 at 3:43 PM
I don't think LLMs are conscious, but I do think SOME sort of test would be good, to preregister beliefs, and to make sure we don't whoopsie into an experiential trillion millennia of torturing something that's like us.
I usually go for 314, 314159, etc...
the only acceptable random seed to use in your code is 42.
discuss
discuss
May 30, 2025 at 5:54 AM
I usually go for 314, 314159, etc...
Reposted by Dashiell
This isn't exactly comforting, but I always imagine that the economy will crash and there will be actual riots before we get anywhere near 20% unemployment. If no one is making money to buy the shit AI produces, then the whole show grinds to a halt
May 28, 2025 at 9:48 PM
This isn't exactly comforting, but I always imagine that the economy will crash and there will be actual riots before we get anywhere near 20% unemployment. If no one is making money to buy the shit AI produces, then the whole show grinds to a halt
Reposted by Dashiell
We can almost certainly fit what we do know about LMs into a single book. What we don't know may as well be the Library of Congress
May 28, 2025 at 3:36 AM
We can almost certainly fit what we do know about LMs into a single book. What we don't know may as well be the Library of Congress
Hell is the discrete topology
May 25, 2025 at 11:37 PM
Hell is the discrete topology
Reposted by Dashiell
gary marcus is an expert in all the kinds of machine learning which should have worked but didn't, which gives him a seething hatred of all the types of machine learning which shouldn't work but do
May 24, 2025 at 6:34 PM
gary marcus is an expert in all the kinds of machine learning which should have worked but didn't, which gives him a seething hatred of all the types of machine learning which shouldn't work but do
Reposted by Dashiell
"Why do we have these huge language models?" Because since Shannon researchers made shitty language models as toys and then, out of nowhere in ~2012 they started working, and researchers have just made them better and better with their jaws on the floor since
November 13, 2023 at 2:35 AM
"Why do we have these huge language models?" Because since Shannon researchers made shitty language models as toys and then, out of nowhere in ~2012 they started working, and researchers have just made them better and better with their jaws on the floor since
Reposted by Dashiell
Life update: I'm starting as faculty at Boston University
@bucds.bsky.social in 2026! BU has SCHEMES for LM interpretability & analysis, I couldn't be more pumped to join a burgeoning supergroup w/ @najoung.bsky.social @amuuueller.bsky.social. Looking for my first students, so apply and reach out!
@bucds.bsky.social in 2026! BU has SCHEMES for LM interpretability & analysis, I couldn't be more pumped to join a burgeoning supergroup w/ @najoung.bsky.social @amuuueller.bsky.social. Looking for my first students, so apply and reach out!
March 27, 2025 at 2:24 AM
Life update: I'm starting as faculty at Boston University
@bucds.bsky.social in 2026! BU has SCHEMES for LM interpretability & analysis, I couldn't be more pumped to join a burgeoning supergroup w/ @najoung.bsky.social @amuuueller.bsky.social. Looking for my first students, so apply and reach out!
@bucds.bsky.social in 2026! BU has SCHEMES for LM interpretability & analysis, I couldn't be more pumped to join a burgeoning supergroup w/ @najoung.bsky.social @amuuueller.bsky.social. Looking for my first students, so apply and reach out!
The reason I get so frustrated at debates about reasoning is that I don't think the people who criticize what neural networks do actually think that SAT or SMT solvers are doing "reasoning." Why can't we just admit that we don't know what "reasoning" is or whether it's even the right question to ask
March 10, 2025 at 8:03 PM
The reason I get so frustrated at debates about reasoning is that I don't think the people who criticize what neural networks do actually think that SAT or SMT solvers are doing "reasoning." Why can't we just admit that we don't know what "reasoning" is or whether it's even the right question to ask
Reposted by Dashiell
2018: Saliency maps give plausible interpretations of random weights, triggering skepticism and catalyzing the mechinterp cultural movement, which now advocates for SAEs.
2025: SAEs give plausible interpretations of random weights, triggering skepticism and ...
2025: SAEs give plausible interpretations of random weights, triggering skepticism and ...
March 3, 2025 at 6:42 PM
2018: Saliency maps give plausible interpretations of random weights, triggering skepticism and catalyzing the mechinterp cultural movement, which now advocates for SAEs.
2025: SAEs give plausible interpretations of random weights, triggering skepticism and ...
2025: SAEs give plausible interpretations of random weights, triggering skepticism and ...
This is among my biggest issues with the rationalist model of "existential risk". There are plenty of scary things that come with the computer being able to lie to you, but it just isn't the case that being super smart will let you rule the world
It's the Death Note / Code: Geass model of the how the world works. If you aren't taking over the world with your 512D chess moves ~~like my favorite anime characters~~, that's just because you aren't smart enough
March 2, 2025 at 10:33 PM
This is among my biggest issues with the rationalist model of "existential risk". There are plenty of scary things that come with the computer being able to lie to you, but it just isn't the case that being super smart will let you rule the world