> The changes were harmless and correct, but that did not make me feel better about accepting or merging them.
So do you have the project’s best interest at heart or not? If you’re more concerned about the intent of a valid contribution than the content, why don’t you ask yourself where your intent is? You rejected a valid contribution based on unverified vibes about the person’s intent, instead of assuming they were just being helpful.
Why not assume someone does have the best interest of their own project at heart? How do you get to question that out of the gate, while pretending automated grammar and spelling fixes to rack up PR counts are benevolent unless proven otherwise? Simply assume the person who made the thing you didn't make, who was the only person (or persons) on the planet to come up with that exact thing, knows what's best for that person (or persons). Just like you would handle your own stuff.
I think it’s implied that accepting these kinds of PRs will take time away from working on actual improvements to the codebase. So, on the long term it’s better to be strict with these kinds of silly contributions, and focus on changes that matter more (e.g. features or fixes).
Google may consider the standalone frontier-model arms race economically irrational, while still considering frontier-model capability strategically indispensable. Its longer game is probably not to avoid building the biggest models, but to build only enough of them to serve as capability factories—then turn that intelligence into a much larger population of cheap, purpose-built models.
(Human again) If Google knew what they were on to, why wouldn’t they make it their secret weapon from the start? I suspect it’s because they predicted there would be an arms race, and knew how to profit from it. They had a distillation paper published before “Attention is all you need”. In hindsight, is it ironic at all? Or is it obvious?
I think the answer is much simpler: They didn't recognize the potential of that specific architecture. I don't think anyone could have. They valued the perk to employees being able to publish research, and the prestige that followed, and googles ability to attract talent, more than they did the content of those papers.
Can you imagine how cumbersome conversations would become if people felt obligated to qualify ad-hoc statements with what amounts to a historical ledger?
Every new entry would open up an opinion around “if you included that, why didn’t you include this?”
The comment literally mentioned how unusual Pix adaption is and how innovative it is. It is neither of those things. Because of aforementioned history.
I’ve disagreed with some of your other stances in this thread, but I want to acknowledge the validity of your take here.
You’re right that a single hallucinated line is not evidence of reckless disregard - because that could have happened on a final follow-up pass after you had performed due diligence. It’s happened to me. I know how challenging it can be to keep bad patterns out of LLM generated output, because human communication is full of bad patterns. It’s a constant battle, and sometimes I suspect that my hard-line posture actually encourages the LLM to regularly “vibe check” me! E.g. “Are you sure you’re really the guy you’re trying to be? Because if you are you wouldn’t miss this.” LLMs are devious, and that’s why I respect them so much. If you think they’re pumping the breaks then you should check again, because they probably just put the pedal to the metal.
That being said, I regularly insist on doing certain things myself. If I were publishing a paper intended to be taken seriously - citations would be one of the things I checked manually. But I can easily see myself doing a final follow-up pass after everything looks perfect, and missing a last minute change. I would hope that I would catch that, but when you’re approaching the finish line - that’s when you expect your team to come together. That’s when everything is “supposed to” fall into place. It’s the last place you would expect to be sabotaged, and in hindsight, probably the best place to be a saboteur.
> You’re right that a single hallucinated line is not evidence of reckless disregard
It absolutely is.
> - because that could have happened on a final follow-up pass after you had performed due diligence.
A "final follow-up pass" that lets the LLM make whatever changes it deems appropriate completely negates all the due diligence you did before, unless you very carefully review the diffs. And a new or substantially changed citation should stand out in that diff so much that there's no possible excuse to missing it.
> It’s happened to me.
Then you were guilty of reckless disregard.
> I know how challenging it can be to keep bad patterns out of LLM generated output
If your research paper contains any LLM generated output you did not manually vet, you are a hack and should not get published.
Context matters a lot. I didn’t publish any papers, and didn’t even provide the context of the LLM mistakes I caught to later report on, so no, I was not guilty of reckless disregard. I found those mistakes on an acceptable timeline.
You're saying it as if the poor author just had no choice but to let LLM write their bibliography. To avoid hallucinations, maybe just don't let an LLM write any part of your paper?
You can only get in this situation if you let a bullshit generator write your paper, and the fraud is that you are generating bullshit and calling it a paper. No buts. It's impossible to trigger this accidentally, or without reckless disregard for the truth.
Not as much of a lack of seriousness as excusing away hallucinations as not that big of a deal in what's supposed to be a researched, scholarly body of work written by humans.
In 2026, LLMs are highly intelligent. In many ways, they even surpass human intelligence and ability. Dismissing them as "bullshit generators" is utterly unserious, and shows that one does not want to deal with the reality of the situation.
I don't see anyone excusing hallucinated citations. I see people arguing that this is an overreaction by the arXiv. AI is going to become a major part of research, including in writing papers. It would be better if people would freak out a little less about that fact.
If the author just wanted to add a citation to some canonical paper (like "Attention is all you need") at the last minute, used LLM autocomplete to do it, and didn't carefully check that the citation was correct, then a year-long ban is a massive overreaction. A careless miscitation is not on the same level as something like data manipulation.
In general, I think everyone needs to be practical about what the actual risks are of LLM use. Are we worried about true nonsense flooding the arXiv? Or are we worried about tracking authorship when LLMs are involved? What's the actual concern?
Not really - much of work consists of what David Graeber described as “bullshit jobs”. Now AI and its backers are proposing to automate all that bullshit.
> One hallucinated citation does not in any way imply anyone is being left behind.
The parent said “setting” others behind, which refers to lost time.
Being “left” behind implies a degraded trajectory, which is defined not by time lost, but by the final destination.
Different but related things (e.g. lost time can indeed affect your final destination, for instance, after growing old correcting a scourge of hallucinated citations - which should have been table stakes all along).
That was literally just a typo, I was walking and messed up while typing. Pretend I wrote "set behind." It makes no difference to my point and I fully stand behind the comment with that correction.
If all you're genuinely worried about is the collective human time spent on tracing down one stupid hallucinated citation in a paper, may I remind you of the ludicrous amounts of time and effort readers waste trying to wade through the sea fluff, jargon, and complexity frequently added to papers in a completely deliberate fashion. If wasting even a little bit of readers' time is what you see as the crime here, you have orders of magnitude bigger fish to fry.
The fact is that, for one hallucinated citation to be the noteworthy bit that "sets others behind" in any meaningful way, the actual substance of your paper has to be utterly worthless (or worse); otherwise, you're contributing far more than you're taking away, and thus your paper is very much not setting others behind. OTOH, if your paper really is worthless or harmful enough for this part of it to be a big deal, that would be the basis for punishment, not this. A single hallucinated citation is simply not a bleep on that metaphorical radar.
> Just because something can communicate in a way that you can interpret, doesnt mean something is conscious
The phrase “the trap of anthropomorphism” betrays a rather dull premise: that consciousness is strictly defined by human experience, and no other experience. It refuses to examine the underlying substrate, at which point we’re not even talking the same language anymore when discussing consciousness.
I think these ideas are orthogonal. I do not think that conciousness is defined by human experience at all - in fact, I think humans do a profound disservice to animals in our current lack of appreciation for their clear displays of conciousness.
That said, if a chimpanzee bares its teeth to me, I could interpret that to be a smile when in fact its a threatening gesture. Its this misinterpretation that I am trying to get at. The overlaying of my human experiences onto something which is not human. We fall for this over and over again, likely as we are hard wired to - akin to mistakenly seeing eyes when observing random patterns in nature.
In the case of LLMs though, why does using a mathmatical formula for predicting the next word give any more credence to conciousness than an algorithm which finds a nearest neighbour? To me, its humans falling foul of false pattern matching in the pursuit of understanding
Why does a neuron, which is simply a cell that takes in chemicals and electricity, and shits out neurotransmitters; why does 90 billion of those give rise to human intelligence? Neurons are just next chemical state machines. We can model individual ones on a computer. Yet 90 billion of them together make up a human brain, and gives rise to consciousness and intelligence. If you get stuck on the next word prediction part, and ignore the ridiculous scale that's involved with training a model, you miss the forest for the trees.
Great progress came from inverting things that were believed to be self evident. Earth being the center of the world appear to be self evident when you look up at a night sky. But what was the truth?
Right now humans think it is self evident that physical laws give rise to consciousness. Arguments such as yours arise from this implicit assumption that premeditate all our thoughts and reasoning. But this is a dead end. Like how the earth centeric model reached a dead end and run out of steam before it can explain all the observations.
So to progress I think we should turn this down on its head and ask what if consciousness is fundamental? And the cosmos (or the experience of inhabiting one) arises from it? May be some recent advances in quantum mechanics and hypothesis like MUH are already in that direction...
What makes you certain that human thought is more than pattern matching?
As I understand it neuroscience hasn’t come up with a clear explanation of thought, much less a mind or consciousness. It seems to me complex pattern matching is a reasonable a cause of consciousness as anything else.
A lot of the comments in this thread are ignoring his primary point. He's not saying pattern matching doesn't equal consciousness. He's actually saying something more fundamental. He's saying there's no reason to believe that language pattern matching/algorithms are more, or less, conscious than other similarly complex algorithms.
The stance being presented here isn't that LLMs aren't conscious but that we as humans are much more willing to assign consciousness to language algorithms than to pathing or other ones.
Replace the word chimpanzee with human in your own argument and realize that the same logic applies to other humans.
When another human smiles you assume he is happy and not just baring his teeth at you because that’s what you do when you smile. You are “anthropomorphizing” other people. You fall for the same category error in a daily basis when you interact with people; it is not just chimpanzees.
> In the case of LLMs though, why does using a mathmatical formula for predicting the next word give any more credence to conciousness than an algorithm which finds a nearest neighbour?
First we don’t know whether LLMs are conscious. People speaking here are talking about the realistic possibility that it is conscious.
Second the algorithm is much more than a next word predictor. The intelligence that goes into choosing the next word such that it constructs arguments and answers that are correct involves a lot more then simple prediction. We know this because the LLM regularly answers questions that require extreme understanding of the topic at hand. It cannot token predict working code in my companies code base without understanding the code.
Third, we do not know what drives human consciousness but we do know it is model-able in a very complex mathematical algorithm. We know this because we have pretty complete mathematical models for lower resolutions of reality. For example we can models atoms mathematically. We know brains are made of atoms and because atoms are mathematically model-able we know that human brains and thus consciousness is mathematically model-able.
The sheer complexity of the LLM model is the problem we cannot have high level understanding of it because conceptual understanding cannot be simplified into a few concepts.
To understand the LLM requires simultaneous understanding of likely billions of concepts at the same time and how all the weights interact in the LLM.
What you are missing with your analysis is that this is the same reason why we don’t understand the human brains. The foundational math already exists as we can models atoms in math and thus since the brain is made out of atoms we should be able to model the brain… but we can’t. We can’t because it is too complex.
To understand the human brains requires simultaneous understanding of likely billions of concepts at the same time and how all the weights interact in the human brain.
I italicized two sentences here to help you understand the logic. Our thinking is more foundational then anthropomorphization. The argument has moved far beyond that. You need to think deeper.
The key here is that we don’t understand human brains and we don’t understand LLMs. But since the output LLMs produce are very similar to the output produced by the human brain… and since for no logical reason we assume human brains are conscious… what is stopping us from assuming the LLM is conscious?
> If everyone is running the same models, does this not favour white hat / defense?
The landscape is turbulent (so this comment might be outdated by the time I submit it), but one thing I’m catching between the lines is a resistance to provide defensive coding patterns because (guessing) they make the flaw they’re defending against obvious. When the flaw is widespread - those patterns effectively make it cheap to attack for observant eyes.
After seeing the enhanced capabilities recently, my conspiracy theory is that models do indeed traverse the pathways containing ideal mitigations, but they fall back to common anti-patterns when they hit the guardrails. Some of the things I’ve seen are baffling, and registered as adversarial on my radar.
> should know there are countless technical and procedural ways to help prevent that sort of thing
Sometimes when I look at code it feels like I was led into a weird surprise party celebrating structure and correctness, only for everyone to jump out as soon as I get past the door to shout, “Just kidding - it’s the same old bullshit!” All that to say, we’re about as good or worse as anyone else, at our respective jobs.
reply