@Jigsaw_You Not everything in preprint is pseudoscience or worthless, the added value of LLMs /generative AI is less and less disputed, at least in some contexts, if you have red the paper the scientists themselves are quite nuanced and careful , with regard to limitations of their research and generalizing.
@Jigsaw_You@ErikJonker the preprint system emerged because review was bottlenecked by publishers, covid was an accelerant, and i’d argue the preprint system saved lives by shortcutting review.
i think a lot of people have a hard time with preprints because it looks like a real paper, in fact a lot of peer reviewed papers are also published as preprints due to access.
but it’s not, it’s just the scientific equivalent of Slack. a faster way to collaborate, trial and error, and find the way
@Jigsaw_You@kellogh , as long as the status is clear (pre-paper, and there are some issues with how clear that is sometimes),discussion and debate, also in early stages is part of the process of science in my view. It's also difficult to define "non-experts", i am not a technical expert in the field but i am perfectable capable to judge methodology, scientific process and other aspects.
@Jigsaw_You@ErikJonker idk i really dislike the “only certain ordained ministers of science are allowed to interpret science” take. the whole point of our system is that science happens in the open. and yeah, one downside is that it gets misinterpreted sometimes, but it’s still better that it’s open and accessible to all, regardless of background or credentials, imo
for reference — i’ve read 100+ machine learning papers over the years, mostly pre-prints. i’ve got no higher level education beyond bachelor’s, so presumably i don’t qualify as an expert. a lot goes over my head, but i’ve built up a bit of an instinct for what’s bullshit and what’s real, why a paper is important, how it fits into my work, etc.
i think this sort of engagement increases impact of research, so it’s good
@Jigsaw_You@ErikJonker i understand how peer review works, but you keep using "assess", like "non-experts can't assess the scientific value", and that seems like a much broader statement than "peer review should be done by experts".
@Jigsaw_You@ErikJonker i hear you, but i'm not sure i agree that abandoning preprints is an adequate response. for one, there's enough experts that believe in this bullshit that this stuff can easily make it through peer review. all the ML greats are fully bought in. second, preprints emerged for a reason, and you can't ignore the problems that led here
that's typically reserved for fields of study that aren't provable. e.g. string theory is considered pseudoscience because it can't be disproven, so experimentation isn't feasible. Note that it's been heavily peer reviewed.
being wrong isn't pseudoscience, it's just bad research, but fully in the realm of science. we follow lots of dead-ends for a lot of reasons
@Jigsaw_You@ErikJonker i don’t think that paper is pseudoscience, it included experiments, and it was also responded to, positively and negatively, with papers that had experiments.
@Jigsaw_You@kellogh ..that's a bit too easy , using generic terms, please be specific about what could be improved in this paper, that could lead to an interesting discussion.
@Jigsaw_You@ErikJonker what paper are you talking about? this quote is taken from the sparks paper that you most recently linked. the researchers that wrote the paper did all experiments without access to any of that
@Jigsaw_You@ErikJonker you’re talking about the literal definition of machine learning, being able to generalize and extrapolate beyond the training dataset, that’s not AGI and we’ve done that for decades with varying success. sure, small perturbations lead to bad results, but we’re very good at measuring this and reasoning about it. that’s not AGI at all.
@Jigsaw_You@ErikJonker i think you might be talking about emergent behavior, which is when a model learns to optimize for an objective that it wasn’t trained to do. that’s one component of AGI, and it’s quite measurable and results can be invalidated. it’s not pseudoscience, because you can talk and reason about it with experimental data
@Jigsaw_You@ErikJonker you can talk about, “do the results of this paper actually exhibit emergent behavior”, and you can do it with real science and experimentation. frankly i don’t care which side of the debate you’re on, if you’re using science to talk about it, you’re engaging productively
@Jigsaw_You@ErikJonker you don't need it, find something that's never made it into public. Or some piece of news or tech that emerged after the training data was captured. This really isnt' hard...
@Jigsaw_You@ErikJonker 🤔 sure, they could be, but if there's plenty of experimental science that can be done without access to the training data. It would certainly be very helpful if they did that, and positive for a whole bunch of reasons, but there's plenty that can be done without it. I'm really not sure what your point is. It certainly doesn't imply that it's all pseudoscience
@Jigsaw_You@ErikJonker you sent me a link. there’s text that can be read. read it. it answers the questions your asking me. if you can’t engage in good faith, i think it’s best we end this conversation
@Jigsaw_You@ErikJonker stuff like this post, you've been very clear that you think practitioners like me shouldn't have direct access to the scientific process. tbh, i've worked with a lot of scientists and i've never seen this sentiment
@Jigsaw_You@ErikJonker if you look at fields like distributed systems — most of the readership is by practitioners, to the point where the writing style is more like a blog post. in fact, a lot of the authorship is even done by people without PhDs, even though they're posted to academic conferences. so this idea that only designated "experts" can partake in peer review seems sketchy at best
Add comment