555watch · HackerTrans

hace 9 mesesdiscuss

In my country there are obvious clones of known portals with clickbaity titles, visually hard to distinguish from real. Always with some random url. Sometimes just fake news and propaganda, sometimes selling fake things. Under paid sponsorship. Reported them multiple times. Always the response is as follows: we looked at it and found nothing wrong. All the disputes are also killed. So its done on purpose

555watch

el año pasadodiscuss

Has anyone commented already about the absurdity of watching Youtube Shorts? A wide empty white space, with a narrow vertical strip of content that is often stretched and split into two smaller videos.

555watch

el año pasadodiscuss

Can't believe I never heard about this. This is a life saver, I specifically installed ton of different browsers to emulate this....

555watch

hace 2 añosdiscuss

I don't know, maybe it's a bit off topic, but at least in cases that I'm imagining, I would always hire a human than fully rely on AI. Let the human consult with AI if needed, but still finalize the decision or result. The human will be thinking about the problem for months or years, even if passively during a vacation, an idea will occasionaly pop up. AI will think about its task for seconds, in case it missed some information or whatever, it will never wake up in the middle of the night thinking "s**, i forgot about X"

555watch

hace 2 añosdiscuss

Your use of the word stochastic here negates what you are saying.

Stochastic Generative models can generate new and correct data if the distribution is right. Its in the definition

555watch

hace 2 añosdiscuss

My primitive understanding is that we approximate a Markovian approach and indirectly model the transition probabilities just by working through tokens.

555watch

hace 2 añosdiscuss

Finding rotations that maximize variance is not that far from euclidean based clustering.

555watch

hace 2 añosdiscuss

Not discussing the trivial example (since for any model there exists a distribution and a dataset on which a model will not perform). Just a general thought. Intro to ML teaches us this: if we want to "learn" a hypothesis class reasonably well with a finite sample, the class shouldn't be too complex. Otherwise we lose precision and/or any guarantees. This implies that for any DL algorithm A(S) on sample S, there exists a data transformation g such that B(g(S)) will result in a lower* risk for some simpler algorithm B. The question is not whether linear models are good or bad, but how complex/expensive is the transformation g.

555watch

hace 3 añosdiscuss

Yeep, stopped reading after that proposal. I'm not sure OP should be advocating stuff like this without at least a warning. It's not the rubbing, the pressure is creating the "stars". Some people already have increased eyeball pressure, as well as increased blood pressure.

I'm not a doctor and don't know what I'm talking about, but would definitely advice against eye rubbing.

555watch

hace 3 añosdiscuss

I think the previous poster claims that different measurement process yields different measurement errors (spread). Since correlation coefficient is a function of spread, if the measurement errors are random, even if the underlying relation is the same, it suffices to increase the spread a little bit and get a subsequently smaller correlation coefficient.

Confidence bounds for every correlation coefficient would add value and _might_ change some of the interpretations.

E.g.: "its average correlation with the other measurements is only 0.03, which is not just small, it is substantially smaller than the next smallest, which is ear breadth, with an average correlation of 0.13."

If the former is 0.03 +- 0.02 and the latter is 0.13 +- 0.07, we could claim that both are equal to 0 (or just equal).

555watch

hace 3 añosdiscuss

You miss a key point... Not sure if deliberately or due to distance from all of that.

Allow Russia to back down with positive land gains today and you create a precedent that any invasion on a non-NATO non-nuclear country is possible and viable just because you're a large world superpower.

This should not be the case, and a large part of Europe understands that. Hopefully, you will too.

555watch

hace 3 añosdiscuss

Sorry if it's a dumb question, since It's quite hard to keep up with all the recent developments in custom GPT/LLM solutions.

Do I understand correctly that the GPT4All provides a delta on top of some LLAMA model variant? If so, does one need to first obtain the original LLAMA model weights to be able to run all the subsequent derivations? Is there a _legal_ way to obtain it without being a published AI researcher? If not, I'm not sure that Gpt4All is viable when looking for legal solutions.

555watch

hace 4 añosdiscuss

Maybe someone that knows the topic well could elaborate a little bit on how do the authors arrive to their conclusion? From what I'm gathering, by Theorem 2 they show, that with n approaching infinity, the probability that the function $f$ exists approaches zero very quickly. I understand that, for this reason, they require that the variables $d$ and $k$ also approach infinity reasonably fast (or faster than $n$) in order to bound that probability with anything other than zero.

We can already stop here and question, why does an upper bound on any similar probability interest us? If we want to show that the function $f$ exists with high probability, we should also consider lower bounds, not only the upper bounds as is done in the paper (clearly, I can bound any probability by 1 and would not be wrong).

But even leaving this question aside and going back to Theorem 2, they essentially show that for any sample of size $n$, they can find (and overfit) a smooth function $f$, given a sufficiently large model space $d$ and $k$. Assume I launch such a model in production, and continue generating further observations $n$. It follows from the theorem 2 that $f$ will quickly become unsuitable and require retraining on a larger training space $d$ and $k$.

However, from a statistical standing point, unless the data generating process is non-differentiable at every point, we should be able to assume that there exists N_epsilon, such, that for every n > n_epsilon, we should be able to find $f$ such that the errors would be controllably small (< epsilon) against the true data generating process f + sigma. So, beyond a certain $n$, further increasing of observations should not affect the initial fit of the model.

This is not at all what follows from the Theorem 2, suggesting that any such $f$ is still fitting on the errors, not necessarily the true process.

What am I missing (or assuming incorrectly)? Would be very interested to discuss this paper further!

555watch

hace 5 añosdiscuss

Probably that they completed an IPO (initial _public_ offering)

555watch

hace 5 añosdiscuss

It's not that difficult to fix package versions for any prod deployment, as long as you deploy the projects as R packages. You can fix the versions in the ``DESCRIPTION`` files, which are easy to manage through, e.g., Docker and ``devtools::install_version``. Essentially, not a huge difference from ``requirements.txt`` that we're used to.

555watch

hace 5 añosdiscuss

> It was then that I first learned that my conscious is only a small part of my mind, and not in direct control of everything.

That's probably the craziest takeaway from these. In my experience, losing half the vision + the middle part was the most often one. However, the craziest attack was when I temporarily lost the control of speech. The mind and thoughts were there, but nothing verbally said made sense when listening to myself. I'm so glad that taking a nap was enough to resolve that...

comments