I feel like a terrible person for writing this, so apologies in advance. But when I read “Thanks to … Opus 4.6 for a lot the research”, and then in the comments people are pointing out what seem to be multiple factual errors, I can’t help but wonder whether this is all true? More precisely, it’s not clear to me how much I should update in the direction of any of the claims made in this post. Could you tell us a bit more about what fact-checking happened?
These were: 1) multiple errors in visualization from the Gemini image 2) clarification on what matters more (in some respects) to the US grid being able to pull off what New Zealand did, i.e. grid operators or grid utilities 3) the 225 billion thing is defensible, I think, as I clarified. Given the amount of facts laid out I don’t think this is a particularly bad track record!
Sometimes you can tell your civilization just Hasn’t Done Its Homework on something because the dozens of LLMs you send on little agentic research quests keep coming back with confused, miscellaneous, high-spread data / takes on something. Like, I’m researching solar flares and the reliability of Earth’s ten thousand total large transformers (which our grids depend on) right now and the sixth context window I ask about this will mention something none of the previous context windows have, in a way that clearly conflicts, and then you realize there are only like, 5 reports out there on this topic and they all came to different conclusions and there’s no backup or “filler” between these 5 two-hundred-page PDFs as you would have for most topics (e.g. LW posts, Substacks, random nerds with 2010 era in-depth sites) because it’s so weirdly obscure. This is also a problem because when the topic has so little trace in the training data, the LLM will search things using horrible priors, and so you need to hold it by the hand much more than usual.
This does not give me any confidence in your results given that in the most trivially checkable places so far, complete ignorant amateurs here have already found serious misstatements. (And in my experience, the LLMs are substantially worse at the harder-to-check sort of physics/electronics/mechanical engineering which makes up the bulk of this post...)
For example, I strongly disagree that you can just casually round off the $225b figure of ‘all losses to all industrial espionage’ as losses to Americans due to Chinese, because Chinese industrial espionage is not limited to America and America represents less than half global R&D so that’s at least a 100% overestimate, and that’s before we get into the question of how much industrial espionage is Chinese. (And if a >2x overestimate doesn’t matter at all, why was it included?)
This post seems like a good example to me of why my proposed LLM policy is a good idea; having no idea what has been checked here and what has been simply laundered through LLMs to sound authoritative and is as deeply flawed as that random spotcheck, I can make no use of this post, and it is worthless to me except perhaps as a source of some curated external links for my own LLM agents. And I don’t think it is front page material until some actual experts, not LLMs, sign off on it.
I have been thinking for a while how to handle cases like OP or vibecoded codebases, which confusingly mingle some unstated degree of human curation, computation, agentic work, and errors/confabulations/miscitations into a curate’s egg of a final end product where validating & vetting would be as much or more work than creating it from scratch would be (especially given their adversarial training to write authoritatively to convince the user, sycophantically).
I think the right epistemic move has to be to go up a level and treat the publication as the prompt, and then individuals can simply ‘compile’ their own using a trusted LLM to reproduce any final claims. This allows whitebox checking and improvement using future better AIs, and the admitted overhead of recompilation could eventually be amortized away by a trusted third party or even cryptomagic like ZKPs.
> This does not give me any confidence in your results given that in the most trivially checkable places so far, complete ignorant amateurs here have already found serious misstatements.
This was a month ago and I’ve smoothed over errors since. A lot of bewilderment has since faded. I certainly hope that’s not due to LLMs talking authoritatively to me, but I have more reason to cast doubt on that now.
> and America represents less than half global R&D so that’s at least a 100% overestimate
The report this is from (F “225”) is about American losses to everyone specifically, not total R&D lost to the Chinese. So foreign IP isn’t relevant.
> And I don’t think it is front page material until some actual experts, not LLMs, sign off on it.
Yes this post did not go through anyone who actually works in utilities or a space weather expert; I think now it was a mistake not to run this through some first. Now that there’s an artifact of my research thus far, doing so is easier, so I’ll do that now and add an epistemic status marker at the beginning.
This was a month ago and I’ve smoothed over errors since.
This was exactly the response I was hoping you would not make. The problem is not the mere existence of a specific error, but what it says about the process as a whole. Thinking you can just patch bugs is not a solution; a solution is preventing the bugs from happening in the first place. The solution to buffer overflows was not patching every C program one by one as hackers discovered each vulnerability, but moving to memory-safe languages; the solution to ChatGPTese is not search-and-replacing em dashes with semicolons or rewriting it until it fools Pangram...
The report this is from (F “225”) is about American losses to everyone specifically, not total R&D lost to the Chinese.
Fair enough. It is still an overestimate for the previous mentioned reason and the footnote is still wrong. (And now that I look at the PDF, I am in even more doubt about the substantive claim of positive externalities; it is not at all obvious to me how to transform a claim of an annual loss of “counterfeit goods, pirated software, and theft of trade secrets” into a global positive externality figure, especially given how enormous Chinese R&D has become as a % of global R&D, and how much of a powerhouse they are in many industries like solar panels or cars. What is sauce for the goose is sauce for the gander.)
This was exactly the response I was hoping you would not make. The problem is not the mere existence of a specific error, but what it says about the process as a whole. Thinking you can just patch bugs is not a solution; a solution is preventing the bugs from happening in the first place. The solution to buffer overflows was not patching every C program one by one as hackers discovered each vulnerability, but moving to memory-safe languages; the solution to ChatGPTese is not search-and-replacing em dashes with semicolons or rewriting it until it fools Pangram...
I’m confused by what you think the counterfactual here is, and how your proposed LLM policy would have helped here at all. Approximately none of the text in this post was written by an LLM. It probably is the case that most of the facts cited in the post came from LLM-generated research, and it’s true that I have no idea how many of them were checked against primary sources. This does not seem like a difference in kind from a post where most of the cited facts came from variety of NYTimes articles (or other secondary sources of similar repute); if anything I’d expect current frontier LLMs to be slightly safer to rely on in this way (maybe more hallucinations, but less actively adversarial “technically true but misleading” framings).
In particular, this seems false:
I can make no use of this post, and it is worthless to me except perhaps as a source of some curated external links for my own LLM agents.
If you believed this, you might be happy to take bets about how accurately the post represented e.g. various historical cases, such as the 2003 solar storms that reportedly bricked 12 transformers in South Africa. I do not think you believe the correct update to have made, upon reading that section of the post, was “no update”.
And I don’t think it is front page material until some actual experts, not LLMs, sign off on it.
I don’t think this post is confidently making “interesting” claims that would more strongly motivate “expert” sign-off than many other interesting[1] LessWrong posts.
It probably is the case that most of the facts cited in the post came from LLM-generated research, and it’s true that I have no idea how many of them were checked against primary sources. This does not seem like a difference in kind from a post where most of the cited facts came from variety of NYTimes articles
I think there is a lot more reason to trust the facts cited in an NYT article. For one, the New York Times, along with most major news publications, has standards for fact checking. They try hard to get primary source validation, or at least secondary source validation (some of those guidelines are stated here); falsifying information is a fireable offense. They also have a reputation to uphold, a major part of which rests on their ability to convey the news truthfully. These kinds of checks don’t really exist for LLMs.
Nor do we have much insight into how LLM information is generated. With news publications, we can at least understand the sorts of biases which might be introduced via the mechanisms under which stories are produced: people interviewing a bunch of people, maybe in misleading ways, leaving out some facts, etc. With LLMs, we have much less of an idea of what kind of errors might emerge, and hence what to mentally correct for, since we don’t understand the process that generates their outputs.
if anything I’d expect current frontier LLMs to be slightly safer to rely on in this way (maybe more hallucinations, but less actively adversarial “technically true but misleading” framings).
Perhaps this is just a personal difference, but I would much rather take “technically true but misleading” over “totally wrong but subtle enough and authoritative enough and seems-kind-of-right enough that you can barely notice unless you really dig into the claims or already have extensive background knowledge.”
I do not think you believe the correct update to have made, upon reading that section of the post, was “no update”.
My response upon reading that LLMs did substantial research or writing for a post is generally to not make any update. That doesn’t mean parts of it aren’t right, they likely are, it just means that it takes a ton of work for me to sus out what’s true (much more than for a human post, for reasons that Gwern outlined above), and it’s usually not worth it.
Yeah I’ve eliminated the footnote’s sentence. It was way too shaky and didn’t even bring much to the post. The reason I quoted it at all is because I had read the claim in an Economist article I couldn’t find, and thought it’d be interesting to include it.
> This was exactly the response I was hoping you would not make. The problem is not the mere existence of a specific error, but what it says about the process as a whole.
Right… well I’m emailing researchers now. I hope to overhaul this post that way. I will definitely do so first next time.
I’m very curious what you think of the New LessWrong LLM policy. It seems substantially stricter than the old one, which was essentially what you proposed.
Maybe this is a communication issue? The style of writing comes across as rather authoritative, the way you write gave me the impression that you are an expert on this topic. The only red flag that I found in the text was the “research by Claude” thanks at the end. Personally I would have appreciated a disclaimer near the start of the article.
Saying that epistimics were discussed on a twitter thread not linked from the article is not helpful to me. I do not have a twitter account, so I’m afraid I’ve still not read it.
Do you have any object-level disagreements? I don’t actually find this meta talk to be useful. Also Twitter does allow you to see individual tweets if linked, if you’re still curious (it’s about the state of the internet’s ability to answer questions around solar storms).
Thanks! For me that helps a lot. I do really appreciate the effort you have put into this, and I don’t want to suggest that no-one is allowed to talk about anything without becoming/consulting experts. At the same time I definitely agree with Gwern that in the age of LLM writing, it is more important than ever to be really clear about the epistemic status of our work.
Agreed. Niche topics need comprehensive posts because readers won’t know almost any details, in these cases it’s relatively easy to get non-core facts directionally correct but wrong especially graphics which lie outside the logical flow.
I feel like a terrible person for writing this, so apologies in advance. But when I read “Thanks to … Opus 4.6 for a lot the research”, and then in the comments people are pointing out what seem to be multiple factual errors, I can’t help but wonder whether this is all true? More precisely, it’s not clear to me how much I should update in the direction of any of the claims made in this post. Could you tell us a bit more about what fact-checking happened?
These were: 1) multiple errors in visualization from the Gemini image 2) clarification on what matters more (in some respects) to the US grid being able to pull off what New Zealand did, i.e. grid operators or grid utilities 3) the 225 billion thing is defensible, I think, as I clarified. Given the amount of facts laid out I don’t think this is a particularly bad track record!
I did write about the epistemics of this project, back when it was in its beginnings: https://x.com/croissanthology/status/2014071957692531107?s=20
This does not give me any confidence in your results given that in the most trivially checkable places so far, complete ignorant amateurs here have already found serious misstatements. (And in my experience, the LLMs are substantially worse at the harder-to-check sort of physics/electronics/mechanical engineering which makes up the bulk of this post...)
For example, I strongly disagree that you can just casually round off the $225b figure of ‘all losses to all industrial espionage’ as losses to Americans due to Chinese, because Chinese industrial espionage is not limited to America and America represents less than half global R&D so that’s at least a 100% overestimate, and that’s before we get into the question of how much industrial espionage is Chinese. (And if a >2x overestimate doesn’t matter at all, why was it included?)
This post seems like a good example to me of why my proposed LLM policy is a good idea; having no idea what has been checked here and what has been simply laundered through LLMs to sound authoritative and is as deeply flawed as that random spotcheck, I can make no use of this post, and it is worthless to me except perhaps as a source of some curated external links for my own LLM agents. And I don’t think it is front page material until some actual experts, not LLMs, sign off on it.
I have been thinking for a while how to handle cases like OP or vibecoded codebases, which confusingly mingle some unstated degree of human curation, computation, agentic work, and errors/confabulations/miscitations into a curate’s egg of a final end product where validating & vetting would be as much or more work than creating it from scratch would be (especially given their adversarial training to write authoritatively to convince the user, sycophantically).
I think the right epistemic move has to be to go up a level and treat the publication as the prompt, and then individuals can simply ‘compile’ their own using a trusted LLM to reproduce any final claims. This allows whitebox checking and improvement using future better AIs, and the admitted overhead of recompilation could eventually be amortized away by a trusted third party or even cryptomagic like ZKPs.
> This does not give me any confidence in your results given that in the most trivially checkable places so far, complete ignorant amateurs here have already found serious misstatements.
This was a month ago and I’ve smoothed over errors since. A lot of bewilderment has since faded. I certainly hope that’s not due to LLMs talking authoritatively to me, but I have more reason to cast doubt on that now.
> and America represents less than half global R&D so that’s at least a 100% overestimate
The report this is from (F “225”) is about American losses to everyone specifically, not total R&D lost to the Chinese. So foreign IP isn’t relevant.
> And I don’t think it is front page material until some actual experts, not LLMs, sign off on it.
Yes this post did not go through anyone who actually works in utilities or a space weather expert; I think now it was a mistake not to run this through some first. Now that there’s an artifact of my research thus far, doing so is easier, so I’ll do that now and add an epistemic status marker at the beginning.
This was exactly the response I was hoping you would not make. The problem is not the mere existence of a specific error, but what it says about the process as a whole. Thinking you can just patch bugs is not a solution; a solution is preventing the bugs from happening in the first place. The solution to buffer overflows was not patching every C program one by one as hackers discovered each vulnerability, but moving to memory-safe languages; the solution to ChatGPTese is not search-and-replacing em dashes with semicolons or rewriting it until it fools Pangram...
You can link to a specific page like
https://www.nbr.org/wp-content/uploads/pdfs/publications/IP_Commission_Report_Update.pdf#page=9BTW.Fair enough. It is still an overestimate for the previous mentioned reason and the footnote is still wrong. (And now that I look at the PDF, I am in even more doubt about the substantive claim of positive externalities; it is not at all obvious to me how to transform a claim of an annual loss of “counterfeit goods, pirated software, and theft of trade secrets” into a global positive externality figure, especially given how enormous Chinese R&D has become as a % of global R&D, and how much of a powerhouse they are in many industries like solar panels or cars. What is sauce for the goose is sauce for the gander.)
I’m confused by what you think the counterfactual here is, and how your proposed LLM policy would have helped here at all. Approximately none of the text in this post was written by an LLM. It probably is the case that most of the facts cited in the post came from LLM-generated research, and it’s true that I have no idea how many of them were checked against primary sources. This does not seem like a difference in kind from a post where most of the cited facts came from variety of NYTimes articles (or other secondary sources of similar repute); if anything I’d expect current frontier LLMs to be slightly safer to rely on in this way (maybe more hallucinations, but less actively adversarial “technically true but misleading” framings).
In particular, this seems false:
If you believed this, you might be happy to take bets about how accurately the post represented e.g. various historical cases, such as the 2003 solar storms that reportedly bricked 12 transformers in South Africa. I do not think you believe the correct update to have made, upon reading that section of the post, was “no update”.
I don’t think this post is confidently making “interesting” claims that would more strongly motivate “expert” sign-off than many other interesting[1] LessWrong posts.
Maybe you disagree! But I don’t think this post is much worse on the relevant dimensions than the average curated post.
I think there is a lot more reason to trust the facts cited in an NYT article. For one, the New York Times, along with most major news publications, has standards for fact checking. They try hard to get primary source validation, or at least secondary source validation (some of those guidelines are stated here); falsifying information is a fireable offense. They also have a reputation to uphold, a major part of which rests on their ability to convey the news truthfully. These kinds of checks don’t really exist for LLMs.
Nor do we have much insight into how LLM information is generated. With news publications, we can at least understand the sorts of biases which might be introduced via the mechanisms under which stories are produced: people interviewing a bunch of people, maybe in misleading ways, leaving out some facts, etc. With LLMs, we have much less of an idea of what kind of errors might emerge, and hence what to mentally correct for, since we don’t understand the process that generates their outputs.
Perhaps this is just a personal difference, but I would much rather take “technically true but misleading” over “totally wrong but subtle enough and authoritative enough and seems-kind-of-right enough that you can barely notice unless you really dig into the claims or already have extensive background knowledge.”
My response upon reading that LLMs did substantial research or writing for a post is generally to not make any update. That doesn’t mean parts of it aren’t right, they likely are, it just means that it takes a ton of work for me to sus out what’s true (much more than for a human post, for reasons that Gwern outlined above), and it’s usually not worth it.
Yeah I’ve eliminated the footnote’s sentence. It was way too shaky and didn’t even bring much to the post. The reason I quoted it at all is because I had read the claim in an Economist article I couldn’t find, and thought it’d be interesting to include it.
> This was exactly the response I was hoping you would not make. The problem is not the mere existence of a specific error, but what it says about the process as a whole.
Right… well I’m emailing researchers now. I hope to overhaul this post that way. I will definitely do so first next time.
I’m very curious what you think of the New LessWrong LLM policy. It seems substantially stricter than the old one, which was essentially what you proposed.
Maybe this is a communication issue? The style of writing comes across as rather authoritative, the way you write gave me the impression that you are an expert on this topic. The only red flag that I found in the text was the “research by Claude” thanks at the end. Personally I would have appreciated a disclaimer near the start of the article. Saying that epistimics were discussed on a twitter thread not linked from the article is not helpful to me. I do not have a twitter account, so I’m afraid I’ve still not read it.
Do you have any object-level disagreements? I don’t actually find this meta talk to be useful. Also Twitter does allow you to see individual tweets if linked, if you’re still curious (it’s about the state of the internet’s ability to answer questions around solar storms).
I added a disclaimer.
Thanks! For me that helps a lot. I do really appreciate the effort you have put into this, and I don’t want to suggest that no-one is allowed to talk about anything without becoming/consulting experts. At the same time I definitely agree with Gwern that in the age of LLM writing, it is more important than ever to be really clear about the epistemic status of our work.
Agreed. Niche topics need comprehensive posts because readers won’t know almost any details, in these cases it’s relatively easy to get non-core facts directionally correct but wrong especially graphics which lie outside the logical flow.