I find this extremely sad.
-
@aud right like maybe the problem with understanding data is that we have so neglected the fundamentals of our infrastructure in pursuit of mass data and papers that treat the process as some isolated herculean event instead of a collaborative synthesis that we simply can't imagine anything but magic but idk
-
Asta [AMP]replied to Asta [AMP] last edited by [email protected]
@[email protected] I feel like an LLM is the opposite of a good scientific model: it is stochastic where no stochasticity is necessary; it is weakly bound in both domain and range; it is inherently incapable of predictions and has an unstable, chaotic output (i.e, a small change in input may produce substantial changes in output, including complete negation). No amount of training is going to going to bind the LLM’s range to pure truth, particularly when the truth is unknown.
Just trash. -
@[email protected] only way to come up with a theory of quantum physics coming from “hey look at the rock falling”? Chatbot, obviously!
Statements like these are so insulting and demeaning. -
@aud what you think "you know what would make our total absence of models better? shoving shit into a model that is hundreds of billions of times less constrained and unpredictable" is a bad argument? i could have sworn...
-
@aud @jonny yeah the places people are using it successfully would have done much better with a stochastic method that actually optimizes towards controllable criteria. currently being used as an incredibly expensive random number generator which has been used for many decades in science with more success but less marketing than we're seeing now
-
jonny (good kind)replied to d@nny "disc@" mc² last edited by
@hipsterelectron @aud fitting models is great, even fitting models that are universal function approximators is great. getting an intimate sense of the limitations and capacities of those algorithms is even better. tailoring them to a specific domain where you understand how their design relates to the underlying generating process you are trying to understand or emulate is the best.
shoving a pre-conditioning prompt into a react template and hooking it up to the openai API is... not that.
-
@[email protected] @[email protected] and, of course, we all know the most efficient way to do science is sift through thousands of papers with a sufficiently sized Markov model specialized in language and hope our fitting process magically filters out the fundamental scientific concepts that the words are trying to express and not, you know, the eight trillion other possibilities.
-
d@nny "disc@" mc²replied to jonny (good kind) last edited by
@jonny @aud yeah we saw flashes of this earlier in bioinformatics everyone loves the idea of models that just work if your data looks similar and it makes you feel terribly intelligent to unify all these concepts but you quickly run into the problem that validating new methods receives zero interest from journals while it's very easy for journals to accept work using a method they've seen before even if it's completely inapplicable. only difference is that LLMs require much less technical effort to hook up which is a great business success but it depends on marketing that generality free of domain constraints
-
@[email protected] into an unconstrained model that is designed to model an entirely orthogonal set of concepts and is likely to reproduce linguistic noise rather than scientific truths.
Haaaaate all this. -
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
-
@[email protected] I think I'd rather draw the tarot than use a chatbot (and I mean this with all respect to drawing the tarot): because with the tarot reading I could use it as a storytelling framework to sort out my instincts, feelings, and conflicts, and go on to do more experimentation. Can't do that with a chatbot.
-
jonny (good kind)replied to d@nny "disc@" mc² last edited by
@hipsterelectron @aud you know i actually like react, and it is both "a cultural signifier for a kind of guy/kind of site" as well as "something i use all the time and find genuinely useful for making websites that are not a total disaster"
-
@[email protected] @[email protected] this is a really good way to put it
-
d@nny "disc@" mc²replied to jonny (good kind) last edited by
-
d@nny "disc@" mc²replied to d@nny "disc@" mc² last edited by
-
samir, mad idea generatorreplied to d@nny "disc@" mc² last edited by
@hipsterelectron @jonny @aud Most of the problem that people have with React isn’t React, it’s using Webpack and friends to build a blog with no HTML, which requires JS and AJAX to render anything at all. You can do this with any framework.
-
@[email protected] @[email protected] @[email protected] I also have to admit that I was forced to work on Redux, not just React, which I hated