Harvard particle physicist Matthew Schwartz drops 36 papers authored with Claude
reddit.com
reddit.com
I was confused too. Whoever wrote the title(s) doesn’t know a second language, otherwise, they would have realized how confusing that was, even to native English speakers.
Edit: the papers do seem to have other names on them, and he says that all papers "were reviewed by humans for accuracy", but I still think there's a very serious risk of people being gulled into okaying something that's plausible rather than rigorously checking everything. There's enough human-generated slop in science already!
We do of course need to watch for hallucinations, but I in general expect AI to be very good at theoretical physics. AI can take a lot of known equations and prove/propose (these are different things!) generalizations to that may or may not match reality. AI can suggest experiments to see if the predictions match the real world. Maybe string theory can finely make a non-trivial prediction that we can test in the real world...
However AI is terrible for other parts of physics (at least so far) and those are also important we shouldn't lose track of them despite the excitement that we can get from progress in things AI is good at.
I wouldn't count on that to last in any area.
Not one sided at all
Three words: Cambridge, Jason Arday.
I suspect there will be citations galore shortly, faster than anyone can think.
Good judgement takes time, but if the time to judge exceeds the time to spam -- and it does, by many orders of magnitude -- then we should expect the approach "judge each work by its merits" to result in being overwhelmed by spam, which we already are. Reputations let you amortize the cost of acceptance and reduce the cost of rejection, giving you a fighting chance against spam that extreme open-mindedness does not.
This is literal slop.
It literally is just spam, and nothing else.
It's like the Nobel disease: https://en.wikipedia.org/wiki/Nobel_disease
Surely, this is not correct. The monkeys are drastically less likely to produce valid words; the set of monkeys that produce valid words are much less likely to produce valid sentences, etc. etc. Very low bar to clear, but of course this is vastly better than infinite monkeys at producing good research.
Now imagine that’s hitting the right letter on a keyboard many thousands of times in a row.
I recommend the story A Short Stay in Hell
Seems like he realized that as well. I suspect that you may be suffering under the same burden as the Redditors.
> Last December, I tried using Claude as a research assistant, and found that Claude Opus 4.5 performed like a strong graduate student at 20 times the speed. Despite Claude producing a high-quality paper at the end of the experiment, it was a slog to get there. I had to correct every sentence it wrote, steer it away from irrelevant threads, and pull it back from dead ends.
Doesn't sound like someone that's producing "slop". But we probably have different definitions of the word, because I've noticed a lot of people use to mean anything AI was involved with at all.
(I did learn much of my QFT from Schwartz’s QFT textbook though, so I hold him in higher regard than <insert random Harvard Physics prof here>.)
It's worth a read, and very interesting. All of the papers he posted have relevant experts of said fields on them as I understand it.
I didn't read the papers, and I am no specialist in any of the fields that guy meddles in, but it really does sound too good to be true. And if not, well, we're entering a new era of humanity right now.
Wisdom of the crowd has finally proven to be more valuable than some obscure title in academia. Crowds are relentless and ruthless, kind of a red team instead of a cronyism.
PS: The infamous replication crisis deserves a honorable mention here.
May someone grant me an ounce of his hubris and I'll become president of the united states. And I'm not even a citizen.
How many different fields would you imagine a Harvard professor to be an expert in?
LLMs making otherwise extremely intelligent people veer off in a strange direction isn't new. When reassured by LLMs, there have been many cases where someone starts posting papers on a plethora of fields they know nothing about, making unified theories that try to tie several fields together, or get convinced they made a real breakthrough when they haven't. I don't think there's been a single counterexample so far where someone used LLMs to push out so many papers on so many topics and have them be correct. So until someone who knows something in these fields pitches in, I'm remaining skeptical by default.
I read the Anthropic post yesterday. I read the Reddit comments today. I observed that many of them are being overzealously anti AI and likely didn't even click the link they were commenting on, let alone look beyond it.
It's good to be skeptical. It's bad to jump to conclusions based on an emotional response to AI in general. It's bad to be a Redditor.
Should they pretend that the papers are great and the guy is infallible just because of his position? No, but his prior achievements are probably a signal that they should not jump straight into the usual circle jerk.
Now in the age of AI this problem will be heavily accentuated. Who's going to review the quality and validity of dozens and dozens of papers being generated by a researcher who used to publish sporadically?
It's not hard to see science going through a "slop crisis" in the next few years, in the same way that now projects are getting inundated with PRs.
That's not entirely true. There's a whole genre of "response" and "response to response" papers that voice concerns when there's a methodological issue. That's somewhat common in my fields (computational biology and biomedical research).
And even if a "response" isn't published, people still try to reproduce published techniques if those seem useful.
You'd be surprised how rarely (overall) reproduction / verification is attempted. Sadly, a lot of research is taken at face value. That's done quite often with AI/tech stuff these days, and even more so for scientific research. It's a hard problem.
I didn't say it's common, I said that stating it's "never" reproduced is false.
How about this: instead of going through publishers, researchers are required (by their funding organizations) to put a bounty on falsifying the result (for, say, 10 years). The higher the bounty, the higher the 'impact factor'. Put your money where your mouth is. A bounty of at least the cost of the study should be expected for serious research.
I can imagine a 'falsification insure' industry. They'd look at reputation as well as employ people close to the subject matter to asses risk, taking over the role of peer review. But with aligned incentives.
This way researchers and institutions who don't care too much about truth are marginalized; they either have low impact factors or they'll need to pay huge falsification insure fees.
And researchers who are sceptical of certain results will have an incentive (money!) to try to replicate it.
Remaining question: who is the authority on whether a publication has been falsified?
I suppose if this framework allows physics and other sciences to rapidly progress (e.g. publish more findings), that is good? It destroys a certain model we had, that science should be arduous and require a genius many years of work to uncover something, but if these papers truly add knowledge via this new tool, then this is for the good.
Not sure implications for researchers, but from the public perspective, now we have 36 new papers that increase our knowledge. But maybe an expert can weigh in, the papers might be slop?
I guess on a theoretical level, if you had a button which would create breakthroughs in a field, but you yourself couldn't understand it, would you push it? I would because otherwise the field may not discover it, and others will be able to understand the breakthrough.
on mobile, it's cancer like the rest of the internet, unapologetically growth-hacked with shit like "warning!!1 mature content!!1 log in with official reddit app to keep our community safe" modal you get regardless of which benign subreddit you try to browse.