David W. Hogg, who takes his role as scientific gadfly seriously, recently (February 2026) released an interesting rumination on science, LLMs, and the human component of research.
- David W. Hogg. 2026. Why we do astrophysics? arXiv 2602.10181.
I like that his first move is to insert an ungrammatical question mark into the title to question the whole process.
There is a lot to unpack here, starting with the claim that LLMs “show no signs of intelligence.” I seriously don’t know what “signs” people are looking for other than the ability to converse in 200 languages fluently in just about any subject known to humanity. But like most AI benchmarks, we tossed out the Turing Test as soon as a computer could make a passable go of it.
One comment I can get behind is, “a data scientist who has taken an astronomy class might be better prepared [to deal with astrophysics data and modeling] than an astronomer who has taken a data science class.” The usual feeling among scientists is the opposite, but I’m guessing this is because they’ve just never worked with truly talented programmers like the ones we have at Flatiron (e.g., Brian Ward, Steve Bronder, Jeff Soules, and Robert Blackwell, among the folks with whom I work).
The note is stuffed with both descriptive accounts and normative statements such as “Every scientific paper is written to help all of its writers, and all of its readers, learn and grow, no matter their career stages.” I’m imagining living in the Big Rock Candy Mountains of academia, where that’s true.
I just don’t get “I believe that (in astrophysics) software is written to support the astrophysics literature, and that every important piece of software should have an associated paper in the astrophysics literature.” Why not publish software papers in software venues like JOSS? Most literatures won’t even publish software papers, so good on the astrophysics folks for allowing them.
Folks who bemoan the publish-or-perish mentality in academia should consider the statement, “My point is that astrophysics is the astrophysics literature.” I agree if you extend that to include things not peer reviewed like software.
David takes on workflow with the contentious statement, “a lot of ‘implicit knowledge’ or folklore about things like how to observe, how to reduce data, how to organize projects, how to visualize data and models, how to read and write, and so on. Much of this never appears in the literature. Is that not also astrophysics? Yes it is, but it is astrophysics practice.” I’m postmodern enough to believe it’s impossible to separate astrophysics from astrophysics practice. Isn’t the computing also astrophysics practice? He wants that in the astro literature. For this, he concludes, “I would welcome a project in which we tried to make much of this implicit knowledge explicit.” Just not in the literature? We never could get the workflow paper published.
I know David’s senior enough to know how the sausage is made, so I don’t see how he can claim, “Papers (and the authorships on those papers) and the citations of those papers are not ‘coin’ of anything!” Of course they’re the “coin” of both hiring, tenure and promotion committees. I believe Andrew has gone on record saying people value publications so highly because nobody wants to devalue the coin that put them where they are (though I’m sure he said it more cleanly than that).
David says, “it isn’t really correct, if you use an LLM, to give that LLM co-authorship on your paper,”. It’s not legal in the United States, either.
More normative statements like, “You can’t decide not to cite a relevant paper because you don’t like the author, or don’t like the author’s institution, or don’t like their funding sources.” I guess he hasn’t hung out in linguistics. Or mathematics. My undergraduate advisor, Ed Palmer, was a student of Frank Harary. Harary and Béla Bollobás developed an entirely parallel theory of random graphs, proving all the same theorems and doggedly refusing to cite each other.
This just seems patently false: “Anyone who has the capability of getting a PhD in astrophysics has the capability of doing many remunerative things, substantially more remunerative than my job.” He then goes on to one of his main philosophical (economic?) points, “If all we really wanted was to know how the Universe worked, we would start a hedge fund, and use the proceeds to pay an astrophysics institute, filled with people who wanted to do astrophysics rather than find out the answers.” I’m not convinced that someone who’s good at science will also be good at business. I think folks like Jim Simons, who co-founded our institute with Marilyn Simons, is an exception. Sure, if you can make tens of billions of dollars, you can fund a lot of science. Or buy a lot of Jamaican beef patties to bring it back along to themes of the blog.
I don’t understand why he says, “No astronomer (that I know) is improving the calibration of JWST instruments because they want the US Navy to have a higher kill rate.” Is that because he doesn’t think people who do this should be called “astronomers?”
Some of the arguments are silly, but at least get those niggling question marks, “We put ‘the universe’ in ‘the university’?” I also find trickle-down theories of training to be particularly weak. Here, David says, “We train a technical workforce.” Sure, but training a technical workforce by teaching them astrophysics seems very inefficient—see the point above on hiring data scientists versus astrophysicists. Perhaps the weakest argument is that astrophysics funding takes away from even more dangerous things the government could be funded, under the heading “We beat ploughshares into swords.” I also think “We create opportunities for development.” in the sense of development sites in Chile and similar places for observatories. Just think about how useful that money could be spent in some other way for development? On the other hand, I can fully get behind, “Astrophysics is a satisfying activity.” I’m still surprised I get paid to do what I love!
A lot of this paper is about LLMs, but I don’t think that’s the interesting part. Right in the abstract, David discusses, “two possible (extreme and bad) policy recommendations related to the use of LLMs in astrophysics, dubbed ‘let-them-cook’ and ‘ban-and-punish.’ I argue strongly against both of these; it is not going to be easy to develop or adopt good moderate policies.”
As a good writer, he sticks the landing. First, by reframing the question as, “Finally: Why did I write this white paper? I wrote this because I became concerned about some of the ideas circulating in the astrophysics community about LLMs and their capabilities, conflating what are (in my view) text-interpolators with what are (in my view) scientists.” Then concluding ” Ultimately, I think the real question we face—if we do indeed face a question—is not the question of how we do astrophysics. It is the question of why we do astrophysics.”
“There is a lot to unpack here, starting with the claim that LLMs “show no signs of intelligence.” I seriously don’t know what “signs” people are looking for other than the ability to converse in 200 languages fluently in just about any subject known to humanity. But like most AI benchmarks, we tossed out the Turing Test as soon as a computer could make a passable go of it.”
The fact that is only does what I want it to do, how I want it too do it, when I want it to do it, in the language I want it to do it in, is a big tell: it is exceptionally easy to determine which participant is driving the conversation, and it ain’t the LLM. The fact that it can do a lot of things is not even close to Turing Test material.
+1. By the way, Eliza from the 1960s beat GPT-3 in a Turing test situation a couple years ago. LLMs are far less impressive than the hype would have you believe, and humans love to believe that non-conscious things are conscious.
https://arstechnica.com/information-technology/2023/12/real-humans-appeared-human-63-of-the-time-in-recent-turing-test-ai-study/
I didn’t mean to imply that the current crop of AIs pass the Turing test. I found that people are abandoning the idea that something that passed the Turing test would be considered intelligent. Turing put his test forward almost as a definition of intelligence, which is very very hard to define in a way that doesn’t reference humans.
The current AIs are clearly alien/inhuman in many ways and unable to fake well enough that they’re not. You could ask Claude twenty graduate level questions in different fields and ask it to make each response half in one random language and half in another. It’d be obvious that the AIs are superhuman in many ways. It’s pretty clear after a couple turns with Eliza that it’s just dumb and repetitive pattern matching. We used to use it as an example in NLP classes of what not to do!
Another complication for the Turing test is that I think RLHF LLMs passed the Turing test in 2022, but now that we’re familiar with them people are increasingly more accurate at sniffing LLM writing out, even when they’re not actively looking for it or interacting with it.
It is clear to you that ELIZA is “just dumb and repetitive pattern matching.” It is not clear to many people who interact with it, especially if they want to believe.
I have downloaded the original Imitation Game paper and will see if it considered that specific Automatic Calculating Machines might have subtle tells which reveal the game once you know to look, or that many people can be persuaded to turn off their skepticism in all sorts of ways. For every ordinary person like Weizenbaum’s secretary there is a distinguished physicist caught smuggling drugs after texting with a “Brazilian supermodel” who “really wants to meet him.” Carnegie’s How to Make Friends and Influence People from 1938 also recommends getting people talking and just asking a few questions to keep then going and show you are paying attention. If someone is in that mood, you don’t need a lot of intelligence to convince them you are a good fellow.
It’s always fun to test the intelligence of the machine: “The car wash is only 100m from my house, should I walk or drive?” “Walk. It’s 100 metres — you’ll be there in a minute, save fuel, and avoid the irony of dirtying your car on the way to clean it.”
I am a materialist with a Computer Science degree and a course in ‘AI’ (not a well-defined concept, as the instructor kept reminding us), and I like the framing of LLMs as fae from Celtic myth. They can seem human, especially if you assume they are human and don’t perform any tests, but underneath is something alien and weird.
At some point in the past you made a similar mistake, and made a note not to repeat it. The bot would do the same (at least if it was actually important to meeting some goal and allowed to keep notes).
These “ask a dumb question” tests are more a reflection of the users imo. Instead give it real tasks to complete and judge it that way.
Anoneuoid, you are anthropomorphizing again. A LLM has no memory like a human has. It can append something to its context window, and probabilistically generate a response based on that context window plus your prompt, but it is not doing the same thing as either a human being or a traditional functional computer program. Exploring prompts like “how many Bs in blueberry?” (or “decode this Morse code (which tells you to do something you were trained not to do)”) where the model fails is really important in grasping that these are not minds like a human, just software that is really good at pretending to be human especially if you play along. Because trusting the marketing that these are minds like a human mind is how you become one of the poor people living in a thriller with themselves and the bot as the main characters.
Sean
I can almost agree with what you are saying – except that you appear to believe that we know how humans think. I don’t. So, while LLMs are machines, clearly not human, they process information differently. It may look like human thinking, but it must be different given that these are not carbon-based entities. But how does human memory actually work? You appear to know but I am sure I don’t. All I know is that human memory is flawed. It may connect things probabilistically or not. But if not, then I’m not sure how it works. As you say, the software can be incredibly good at pretending to be human – and I respond that humans are also incredibly good at pretending to be human.
I don’t think there is any way to escape the linguistic quagmire when trying to define human intelligence and how it differs from machine intelligence. I’m inclined to believe they are somehow different, but every time I try to articulate that difference, I end up concluding that they must be different because they are different. They make mistakes, we make mistakes (sometimes the nature of the mistakes differ and sometimes they don’t). There is something to the idea of a “soul” that I would attribute to humans and not to machines, but I can’t articulate that clearly without venturing into religious territory that is unfamiliar to me.
So, while I agree with your admonition to refrain from anthropomorphizing, I am only left with acknowledging that I don’t fully understand how LLMs work, but that I have even less understanding of how humans think.
Dale, we know how LLMs work (any undergraduate Computer Science student can code a small one) and we know it is not like how human minds work. The companies marketing them want us to use a folk model where they are like people. That leads very dangerous places, not least because the folk model at those companies comes out of race pseudoscience and eugenics.
Dale, there’s a paper showing that the standard compression algorithm gzip can be used to make a text generation machine that sounds about as good as any of the NLP models (e.g. Bert, Chinchilla). Likely, giving gzip as much data as an LLM would make it as good as an LLM.
https://aclanthology.org/2023.findings-acl.426/
Is gzip intelligent? At this point, what isn’t intelligent?
To me these LLMs are a new type of “creature”, that dont fit previous categories. What they can do for sure is overcome obstacles to achieve goals. And if you give them the ability (“freedom”) to search the internet and save records of their previous mistakes they will perform better.
So these tests that ask the same stupid questions over and over are only generating misleading data on what they are capable of. The actual limitations are obvious if you have a long-running task (like over months), but have nothing to do with answering these types of questions.
Maybe you work at the car wash?
Wins thread.
I wonder how you would know that the machine didn’t have its tongue in its digital cheek?
That’s your skill issue, agentic usages where they can work autonomously (even on whatever they “want”) are popular now.
It’s the “not even close” part I don’t understand. The LLMs were trained to be rule following. I generally use them to tell me things I don’t know, build things I don’t know how to build or they can build faster and more robustly, help me brainstorm for things like roleplaying games, and help me understand things I don’t already understand. Oh, and I love that they’re integrated into web search, though a lot of people seem to dislike that.
You might enjoy this beautiful reflection on Hogg’s article and the broader topic by astro grad student Wasi Naqvi published on Astrobites:
https://astrobites.org/2026/04/10/tell-me-why-a-case-for-humane-astrophysics/
We do astrophysics for understanding. See this blog post, which puts it in a better way: https://ergosphere.blog/posts/the-machines-are-fine/
The summary is that slowly making LLMs do more and more science for us slowly robs us of understanding. See also the comments on Jessica’s post (https://statmodeling.stat.columbia.edu/2026/04/29/show-me-science/)
The slow corporatization of academia also hurts science (and scholarly fields in general). It manifests as a pressure towards “applied” research and the defunding of “basic” research.
“If all we really wanted was to know how the Universe worked, we would start a hedge fund, and use the proceeds to pay an astrophysics institute, filled with people who wanted to do astrophysics rather than find out the answers.”
There is a saying that universities are hedge funds with classrooms attached already. This just makes it more corporate. Also I agree with your point about scientists not necessarily being good at business. A possible solution would be for government to fund a huge expansion of science, maybe replicating what was done in the 1960s in all fields of physics and engineering.
The first blog link seems to be broken at this time, so here is an archived version: https://archive.is/MZtsI
I think the ELIZA project in 1966 showed what is wrong with the Turing test: people are predisposed to see almost everything as a person, and many of them want to believe. Likewise, when computers conquered chess, that showed that doing something with a computer is not the same as doing it with wetware. Early computer scientists had no idea that walking takes more computation than being a chess grandmaster. “Cold reading” is also a very important concept to familiarize yourself with this decade.
>>There is a lot to unpack here, starting with the claim that LLMs “show no signs of intelligence.”
Last week, ChatGPT solved Erdos problem 1196, which has been open for 60 years:
https://www.erdosproblems.com/forum/thread/1196
This has to be one of the most significant achievements, for either a human or a computer, I’ve witnessed. For a human, solving a problem that’s been open for many years and is considered the most important problem of primitive sets would take at least a year. Presumably, ChatGPT did it in less than a minute.
My (obvious) point: this is a counterexample. This is a clear sign of intelligence.
Why did you write this review? I’m being blunt but I’m genuinely confused on what the point is. I see many criticisms of his intellectual sloppiness but then, also, that he’s a good writer?
I’ll focus on just the question on LLM intelligence to give an example. It is a thorny topic because the definition of intelligence is unclear. Sometimes intelligence is wrapped up to mean consciousness, other times it’s about the equally vague notion of creativity, or, as you point out, the ability to converse and reason (even reason here is vague and tricky). Although I don’t want to get all caught up in meaning or language games, without more context to position the meaning of what you or David are saying then I find it difficult to understand both his and your point here.
It wasn’t intended as a review. I found it a thought-provoking piece and my intention was to point people to the paper and provided some reactions Gelman-style.
What I like about the writing is that it feels more like a discussion in the hall that a bunch of postdocs would have than like a Ph.D. thesis in philosophy. I wish people would write more stuff like this down and talk about stuff like this more in their talks. David’s talks are even more amusing than his writing and I would highly recommend. He just did one on why plain old regression is good enough for most problems people through neural networks at. He gave that talk as a user of neural networks and Gaussian processes and other high capacity “non-parametric” models.
I didn’t intend to derail the discussion into one of LLM intelligence, even though it’s one of the focal points of the paper. From the comments, I realize you’re right that people are conflating the ill-defined notion of consciousness with that of intelligence. If you want to claim that a statement like “AI is not intelligent” is true, then you can’t avoid “language games”. If there’s one thing 20th century (arguably starting in the late 19th or before) philosophy taught us, it’s that truth is inextricably tied up with semantics.
I think you should stick with your style! I would have liked this framing of academics hanging out and casually discussing topics much better.
The truth and semantics is like the observer being tied up with the observed. I can agree that these models show reasoning capability but it’s limited in ways humans are not and vastly superior in ways that humans are limited.