I don’t agree with everything that Freeman Dyson writes, but this review of a book by Columbia physicist Brian Greene was pretty good. It’s on the topic of string theory, which has come up before on this blog.

Dyson begins his review on a judiciously positive note:

I recommend Greene’s book to any nonexpert reader who wants an up-to-date account of theoretical physics, written in colloquial language that anyone can understand. For the nonexpert reader, my doubts and hesitations are unimportant. It is not important whether Greene’s picture of the universe will turn out to be technically accurate. . . . Even if many of the details later turn out to be wrong, the picture is a big step toward understanding. Progress in science is often built on wrong theories that are later corrected. . . . Greene’s book explains to the nonexpert reader two essential themes of modern science. First it describes the historical path of observation and theory that led from Newton and Galileo in the seventeenth century to Einstein and Stephen Hawking in the twentieth. Then it shows us the style of thinking that led beyond Einstein and Hawking to the fashionable theories of today. The history and the style of thinking are authentic, whether or not the fashionable theories are here to stay.

After quoting from Greene’s description of string theory, Dyson continues:

This is a fine beginning for a theory of the universe, and maybe it is true. To be useful, a scientific theory does not need to be true, but it needs to be testable. My doubts about string theory arise from the fact that it is not at present testable.

I guess not much has changed in the past twenty-one years!

More interesting than the string-bashing or string-skepticism is Dyson’s age-period-cohort take on the sociology of the cutting edge of physics:

In the history of science there is always a tension between revolutionaries and conservatives, between those who build grand castles in the air and those who prefer to lay one brick at a time on solid ground. The normal state of tension is between young revolutionaries and old conservatives. This is the way it is now, and the way it was eighty years ago when the quantum revolution happened. I [Dyson in 2004] am a typical old conservative, out of touch with the new ideas and surrounded by young string theorists whose conversation I do not pretend to understand. In the 1920s, the golden age of quantum theory, the young revolutionaries were Werner Heisenberg and Paul Dirac, making their great discoveries at the age of twenty-five, and the old conservative was Ernest Rutherford . . . a great scientist, left behind by the revolution that he had helped to bring about. That is the normal state of affairs.

Fifty years ago, when I was considerably younger than Greene is now, things were different. The normal state of affairs was inverted. At that time, in the late 1940s and early 1950s, the revolutionaries were old and the conservatives were young. The old revolutionaries were Albert Einstein, Dirac, Heisenberg, Max Born, and Erwin Schrödinger. Every one of them had a crazy theory that he thought would be the key to understanding everything. Einstein had his unified field theory, Heisenberg had his fundamental length theory, Born had a new version of quantum theory that he called reciprocity, Schrödinger had a new version of Einstein’s unified field theory that he called the Final Affine Field Laws, and Dirac had a weird version of quantum theory in which every state had probability either plus two or minus two. . . . Each of the five old men believed that physics needed another revolution as profound as the quantum revolution that they had led twenty-five years earlier. Each of them believed that his pet idea was the crucial first step along a road that would lead to the next big breakthrough.

Young people like me saw all these famous old men making fools of themselves, and so we became conservatives. The chief young players then were Julian Schwinger and Richard Feynman in America and Sin-Itiro Tomonaga in Japan. Anyone who knew Feynman might be surprised to hear him labeled a conservative, but the label is accurate. Feynman’s style was ebullient and wonderfully original, but the substance of his science was conservative. He and Schwinger and Tomonaga understood that the physics they had inherited from the quantum revolution was pretty good. The physical ideas were basically correct. They did not need to start another revolution. They only needed to take the existing physical theories and clean up the details. I helped them with the later stages of the cleanup. The result of our efforts was the modern theory of quantum electrodynamics, the theory that accurately describes the way atoms and radiation behave.

This theory was a triumph of conservatism. We took the theories that Dirac and Heisenberg had invented in the 1920s, and changed as little as possible to make the theories self-consistent and user-friendly. Nature smiled on our efforts. When new experiments were done to test the theory, the results agreed with the theory to eleven decimal places. . . .

This is fascinating. I’d never thought of the history of twentieth-century physics this way, and it leaves me with some thoughts:

1. The age-period-cohort nonidentifiability, something we’ve seen before when studying public opinion. Dyson is talking about the scientific views of leading physicists, but it’s the same general thing, that you can explain the observed data in multiple ways.

One story is that attitudes of the young were driven by the logic of events: in the 1910s-20s, the foundations of physics were in a mess and so the young physicists were radical, recognizing the need for revolutionary science; in the 1940s-50s, the foundations were strong and much progress could be made using normal science, hence the young physicists were conservative; in the 1980s-90s, the advances of conventional methods in fundamental physics had trickled to a halt, hence the young physicists were motivated to be radical. Another logic-of-events thing that Dyson could’ve mentioned, but didn’t, is that the 1940s-50s were special in that a huge amount of effort in theoretical physics was going into the design of atomic bombs: for that, what was needed was creativity in the application of existing fundamental theories, not a fundamental restructuring. After the 1950s, the military remained a major funder of physics, but bomb design was no longer the cutting edge.

A different explanation is based on cohorts. The cohort of Einstein, Dirac, etc., achieved success with new fundamental theories and so they kept wanting to do that–they remained radicals all their lives. Reacting against this, the cohort of Feynman, Dyson, etc., achieved success while working within existing theories, so they remained conservatives all their lives. The cohort of Greene, etc., reacted against their fathers and became radicals.

These two stories can coexist; that’s part of the nonidentifiability.

2. Similar things have gone on in statistics. From the 1950s-1980s there was a conservative movement within academic statistics featuring opposition to Bayesian methods (see discussion here), with some of this attitude lingering even into the 1990s (as discussed here). My generation was more radical, favoring developments in many different directions. In recent decades I and others have become more aware of misuses of statistics but we have not framed this as an anti-modeling stance. My point here is not that I’m right and these other people are wrong, but just that, as with Dyson, I see generational differences. Indeed, I’ve talked with some young statisticians who express to me what seem like naive old-fogey attitudes demanding statistical guarantees. Again, these are legitimate differences of opinion, in the same way that it’s perfectly fine for Dyson and Greene to differ on the value of string theory. It’s just interesting to see these sorts of age-period-cohort effects in these contexts.

This is Jessica. There’s been some debate among computer scientists about what policies conferences should adopt for papers with hallucinated references. An independent analysis turned up at least 53 NeurIPS 2025 papers that were accepted (and presumably presented) at the conference in December but which had at least one hallucinated reference.

The question is, what should the default policy be if a paper is found to have at least one hallucinated reference? Should we conclude that these papers should have been rejected, and retract them? Should we instead let authors correct them? Going forward, should we desk reject papers with at least one hallucinated reference? What exactly can be concluded about the quality of the rest of the paper if you find at least one hallucinated reference?

The NeurIPS board statement suggests leadership is uncertain what to do about these papers:

“The usage of LLMs in papers at AI conferences is rapidly evolving, and NeurIPS is actively monitoring developments. In previous years, we piloted policies regarding the use of LLMs, and in 2025, reviewers were instructed to flag hallucinations. Regarding the findings of this specific work, we emphasize that significantly more effort is required to determine the implications. Even if 1.1% of the papers have one or more incorrect references due to the use of LLMs, the content of the papers themselves are not necessarily invalidated. For example, authors may have given an LLM a partial description of a citation and asked the LLM to produce bibtex (a formatted reference). As always, NeurIPS is committed to evolving the review and authorship process to best ensure scientific rigor and to identify ways that LLMs can be used to enhance author and reviewer capabilities.”

To make things concrete, consider a hallucinated reference to be a citation listed in the references section of the paper where the average reader cannot (in a reasonable amount of time) determine the identity of the cited paper well enough to track it down. That is, even if the hallucination is a transformation of what was originally a valid citation, the transformation is severe enough that it’s not obvious what the paper is. Hallucinated references are distinguishable from more minor errors like syntax issues or other errors that affect the citation but don’t prevent you from still easily tracking down what was intended. 

I think we should be asking ourselves: What would we do if we found there was hallucinated evidence, such as experiment results? And we should treat these papers with hallucinated references as equivalently problematic. It doesn’t matter how many hallucinated references. It doesn’t matter how “valid” most of the paper is, or the probability that the main conclusions are correct conditional on finding a hallucinated reference. As most of us learn in primary school, a key reason authors cite relevant prior work is to help establish support for claims they make. If we don’t necessarily require that those references link to real research, then what are we even doing? 

For the NeurIPS board to say that “the content of the papers themselves are not necessarily invalidated” suggests that they think some degree of fictionalized evidence is tolerable, if it happened through honest mistake. A friend recently relayed to me such a horror story, in which, in a last minute rush before a deadline, they gave an LLM the full correct citations for their paper and prompted it to fix some minor formatting issues to conform with the required format. They submitted the results in time, only to find that the model had added a single citation to a non-existent paper, listing them as an author alongside some renowned researchers in the field. Imagine anyone you look up (much less one of the big wigs you associate yourself with in the fake citation!) reading your paper and discovering this. Yikes. 

So yes, these kinds of mistakes can happen. But I disagree with the board that it matters whether the hallucinations were accidental and most of the paper is ok. Sure, the proofs might be correct, or the paper’s experimental results unaffected. But if authors are using LLMs to help with their citations and are not building time into their process to check the results, it seems fair to conclude that either 1) they don’t understand the errors LLMs tend to make very well, or 2) they don’t consider it a priority to get the facts right. In most cases we can rule out #1 with ML researchers, suggesting that not everyone is on the same page about the importance of not making things up. When you imply that hallucinated references do not necessarily affect the validity of the paper, you signal tolerance for some amount of hallucinated evidence. You tempt authors to keep taking their chances with how much responsibility they can offload to models, rather than encouraging them to retain, regardless of tool use, a sense of personal accountability for the factuality of what they submit. 

Ultimately, I don’t think it matters that much whether NeurIPS allows authors of the affected 2025 papers to correct or retract. It would not surprise me if leadership decides to err on the side of the authors and let them correct, given that policies about LLM usage are evolving rapidly. What does matter is that they signal a lack of tolerance going forward, and this is where they missed an opportunity. 

This is Jessica. In “A Glass-Bottomed Cadillac”, David Hickey describes the advice Hank Williams Sr. received from his father and passes on to his son: “Don’t get any on you, pipsqueak.” By which he meant, don’t let the moral fallout of the road permeate your sense of self. Will Oldham describes the same struggle to retain yourself while the world presses in:

I am still what I meant to be
And I’m losing my mind
But our burdens must lessen
Though our enemies thrive

These days it seems there’s plenty of compromising mess to get on a person who isn’t being careful. By using certain services or purchasing from certain companies, you may be implicitly empowering forces you don’t agree with. Take X, fka Twitter, for example. Until a few days ago they were still enabling anyone willing to pay a small fee to generate child sexual abuse material. To Elon Musk, stopping users from subjugating whoever they please, regardless of their age, was unnecessary censorship. And yet, most of the people I know there went about their business on that platform without so much as a peep. By which I mean breathlessly posting about all the other, apparently more intellectually gratifying look-at-what-the-AI-did-now stuff. Watching AI research conversations play out on social media lately is both exhilarating and exhausting, with the volume of news coming out, and the sense of needing to not miss out on the next source of buzz that it fosters. One could even liken being in AI or ML to being a porn or sex addict, with all the urgency and heavy breathing. But that’s an analogy for a different post.

A less experienced version of myself might have felt personally offended by the thought of friends continuing to support a situation that actively disenfranchised people like me. But one eventually learns that going through life taking such things others do personally is no recipe for peace of mind. I also know from experience that the internal calculus does not always feel so easy when it comes to deciding when to vote with one’s feet. In the case of X’s knowing enablement of child porn, the right thing to do may have seemed obvious (like, don’t send your money every month to an entity that is enabling CSAM at scale!) But I get that severing connections with things you perceive as important to your goals can be difficult, and I think many researchers see their social media influence as part of their identity and evidence of success. Still, the whole situation makes me feel hesitant about going back to that platform.

From a practical perspective, part of the problem with getting too offended by people for not standing up to systems in which they are embedded, even if the principle of it does really does seem obvious, is that it assumes a level of intentionality that they may not have. To be clear, I do not mean that the choice is not ultimately available to them, or that it’s wrong to hold people accountable for their actions. Rather, I have been thinking about how hard it can be to possess oneself, and how many people do not fully possess themselves. By which I mean, they lack either the imagination or the intrinsic motivation to live by their own values. Or as Gram Parsons once put it,

Some of my friends don’t know who they belong to
Some can’t get a single thing to work inside

It’s possible to shut many things out in the name of “making it.” In my experience what gets shut out, and what you let get to you, is often not so much a choice as an attribute of your level of self-knowledge and self-ownership about yourself at that point. Are you capable of imagining a version of yourself that remains true to your values even if it causes friction with other perceived needs? Are you capable of imagining a version of yourself that is sure enough of what you’re doing that the friction disappears? Could you love yourself as much or feel as secure in your position without your Twitter account?

It’s easy to accumulate impurities in the quest to be something greater than your current state. As Rafe Meager writes in their recent essay, “In professional life, one is obligated to traffic in a certain amount of bullshit. We all know that. But it has to be finely calibrated, and that is hard.”

Calibrating our behavior and beliefs is hard in part because we can become conditioned to acting against our sense of self as a normal part of personal and professional growth. Actively moving toward the things that challenge you is not necessarily a flaw, and certain things must be endured to succeed in a new game. And often the things that challenge us are the things that don’t come naturally, that are not, at first glance, clearly aligned with our values.

When I was younger, one of the last things I wanted to be was a computer scientist. This is not to say I’m unhappy, or I don’t enjoy what I do now, because I very much do. I just can’t even imagine having to explain to my high school self how that happened, as what we do in computer science (and to some extent in other disciplines I frequent, like statistics) has never felt like the things that I’m most intrinsically motivated to be good at – I always did very well in math and science classes, but I had no passion for it. But somewhere along the way I got disillusioned with the pursuits that seemed more aligned with my values, like art, and so I started gravitating to the things that seemed most different from what bothered me. After enough years of testing to see how far you can take something, it becomes hard to go back, and it does start to feel like what you’re meant to do. Meanwhile, the Destroyer song loops in my mind,

Don’t become
Don’t become
The thing you hated
The thing you hated
The thing you hated

It’s a dangerous game, getting very good at letting go of things that once seemed important to who you are, in favor of urges to be something else. On the one hand, there’s a real power to be gained in separating yourself from the things you think you need. Sometimes after you take the first step, it’s like something clicks and there’s a high in suddenly realization that something you were stuck on bothers you no more. Or maybe it still does, but somehow now you can find peace in settling back to watch the tides of your cut-off aspirations and desires continue to pulse back toward what was severed. A way of earning indifference. I believe it’s possible to literally remake yourself this way.

And yet part of me, in looking back, feels like maybe I let myself down. Maybe I would’ve been a better writer or philosopher, both of which I always felt more personal proclivity for. Or maybe I would just feel less disillusioned now if I’d kept up certain hobbies that do matter to me—like writing poetry—more consistently, rather than dropping it for so many years because I couldn’t imagine what it would mean to be a person who did both. It was a failure of imagination: I couldn’t conceive of devoting myself to becoming the new thing while also retaining other aspects of myself.

What’s interesting is that this kind of letting yourself down, by letting yourself drift too far from your original purpose or what you feel like you’re best at, is not necessarily so different from lacking the imagination to do the right thing in the current political moment. There is a loss in both cases, a quiet slipping away of who you really are while you think you’re out there proving it. While you think you’re the one who has your priorities straight, while you’re striving to play the game, or maybe you’re even killing it, comparing how many followers you have or how many papers you wrote this year to those around you. The ignorability of it all is terrifying. Kierkegard got it right when he wrote that “The greatest hazard of all, losing one’s self, can occur very quietly in the world, as if it were nothing at all. No other loss can occur so quietly; any other loss—an arm, a leg, five dollars, a wife, etc.—is sure to be noticed.”

How do you tell whether letting go is growth or self-betrayal—especially when what we “need” narrows what we can even see? On the one hand, by moving more towards statistics and formalization over time, my thinking has expanded to encompass new forms of rigor. But it’s a kind of rigor along narrow lines. In another sense I lost imagination, in that it’s now harder for me to take seriously things that I can’t fit into my formal frameworks.

Did I lose ownership of myself? When I think about what it means to fully possess oneself, I think of Aristotle’s preoccupation with explanations of the inner principles that determine an entity’s states of change and rest. He distinguished that which exists by nature (physei, φύσει) and that which exists from other causes (di’ allas aitias, δι’ ἄλλας αἰτίας). It’s the first kind that’s self-possessed: it “contains in itself its own archê (ἀρχή),” the principle and origin of its entry into presence; the second “does not have its principle in itself,” but finds it in the productive activity of human beings.

This is why failures of imagination feel so tragic. To lack imagination is not only to fail to picture alternative versions of the self; it is to lose contact with the inner source that could have animated them. You become legible, optimizable, and perhaps successful, but successful in the way an object is successful when it performs its intended function. You can be moving quickly, and still be at rest with respect to yourself.

Researchers are beginning to ask how, and if, generative AI systems can attain something like intrinsic motivation. How do you get it to devote itself to an open-ended goal, like creative expression, that can’t be boiled down into a simple reward model? In a recent paper, Charness and Grieco find that AI outputs outperform human outputs (as determined by other humans) for tasks that are more clearly specified in terms of how to solve (“closed”), like writing short stories using specific required words. But AI outputs consistently underperform human outputs for open-ended tasks, like inventing things, where the participant is required to find, invent, or discover the problems. They propose a model for the utility function the agent faces as depending on three factors: the output they’d get from simply following instructions, deviation from the instruction-following output due to randomness (e.g., model temperature), and the utility of exercising imagination. Models can only follow instructions more or less closely, and obtain more diverse outputs through higher temperature, but humans alone respond to the pleasure of bringing imaginative ideas to life, a proxy for intrinsic motivation.

Is intrinsic motivation partly a matter of temporal depth—the capacity to care about consequences that don’t pay off immediately, or even in any obvious reward currency? What kind of “vision” allows an agent to see farther ahead, and be moved by what it sees? Yeats, in A Vision, has a line that keeps returning to me: “The Spirit … may know the most violent love and hatred possible, for it can see the remote consequences of the most trivial acts of the living, provided those consequences are part of its future life.”

The need for this kind of vision feels politically relevant now, as programs are dismantled and regulations rolled back or rewritten in ways that will reverberate for years. As a professor, most present in my mind are the moves that affect science in this country—visa regulation, institutional acquiescence when under political attack, the seeding of doubt in the goals of science or value of education.

But Yeats’ quote also feels very personally relevant. What would it take for us to be able to see this way in our personal choices? When we ignore the alliances we support through our various choices, we narrow our vision on purpose. It’s a survival mechanism, but it comes at a cost.

Ultimately I don’t think anyone owes me an explanation of their internal calculus. But neither do I owe them the assumption that they are in control of their drives. There was a time I might have tried to convince myself that the right interpretation was to give everyone the maximum benefit of the doubt. But as Andrew emphasizes, steelmanning is its own hang-up.

Meanwhile, there are days now where I feel like I’m just waking up, from a period of my life where I became laser focused on goals and outcomes and forgot about everything else. It’s quite painful sometimes, existing with this new awareness of what was always there, an inner “archê” that I’d let go quiet. More bluntly, it’s a cliche mid-life crisis with everything but the sports car. But the waking up is also very beautiful: realizing, suddenly, how real things are, how alive, and how little use there have for whatever games you’ve been playing. You don’t get infinite chances to notice.