cosmic_ape

Claims should be verifiable in principle by humans, obviously. I do not suggest paper be stored as binaries. Perhaps replace “should not be human readable” with “should not be designed for human reading convenience”. They can be written by LLM, or dumped as a list of supported claims, or anything like that. Does that make sense?

You know how doctors use ultrasound to check the state of inner organs of a patient? I think they are gullible too, trusting their equipment like that. They should be cutting into the patients and seeing with their own eyes.

Seriously though. My aim would be that I understand the paper. Its not one query where the LLM tells me something, its an iterative process. So the failure mode, for the LLM to be incorrect here, its not that easy to stumble upon. And there are opportunities for this to be corrected later.

I agree that there needs to be some quality control. But why impose it this way? This penalizes also people who have really good contributions. I've made some related points in the other thread above.

What is important is that these are academic papers. I would not be interested in reading AI generated texts here, for instance. But academic papers already have a sort of dry purposeful language, intentionally.

but I already understand it after LLM. why would I read it?

What might be lost is, for instance, small side observations, not central to the main message of the paper, that had no place in the intro. But they were lost before too, in the appendices.

What is gained is at least a week of the researcher's (the paper author's) time, sometimes more.

I would go even further, and say that the idea that an academic paper should be readable by humans in 2026 is strange. The abstract and the introduction should be. The rest of it, as a reader, I can extract by talking to an LLM and asking what goes on in the paper.

I have reviewed tens of papers this way by now, and one thing I'm certain of is that I get much better understanding of the paper this way. This does not save me time usually. If anything, I now spend more time reviewing a paper, but when I'm done I know more about the paper and if I need to provide feedback, it is more substantial.

I would still prefer the abstract and the intro to be written for humans, as a sanity check. But whether the authors pored over every word, every word is "theirs" does not seem to matter much. If the ideas are good they are good, and if they are bad good writing is not a way to compensate.

Two additional observations here:

* That separation of high level and technical content is not new. Conference papers for a long time now are limited to ~9 pages and Appendix, and it was a well known secret that no one ever reads the Appendix, which contained all proofs, for instance (for all practical purposes. and I did, for the record). Up to now that is, now we do read it again with LLM. No point not to do it.

* There may be a finer point of whether one can separate the ideas from the language used to convey them. But I think it can, and again, we did it for a long time with the appendices.

It’s still alive on most cameras at least.

Hi, It sounds that its not only the fact of being alone that is difficult, but that you are going through changes. You were used to a certain way of viewing life and maybe youself as part of that life, and this changed. It usually will take time to figure what do you want now. Long term there probably are no tricks to this, as there are no tricks to building a house. You just put them bricks one after another, see if they fit. But you are in control of this, and in some ways it may be exciting.

In practical terms: (a) Give yourself time to figure what you want to do now. You are in no hurry. (b) Talking to a psychologist may help a lot. If you think talking to a psychologist is wierd, know that at least most psychologists talk to a (another|) psychologist regularly. (c) Exercise is good. If you are not into it, start slow. Walking, situps in the morning, a few pushups. Do it because its something that is generally good for most people, and now you have the opportunity to build that habit. Not because its a trick or distraction. Also, you don't need a gym for this, but if you go to a gym, you'll be around other people. You wouldn't need to interact with them, but you might. (d) There are hobbies that are quite social. Theatre classes, Karate or BJJ, writing workshops or photography classes. Depends on what you're into.

Enjoy

2013 But still cool

Second that. There were articles a year or two ago about TVs trying to connect to any open Wi-Fi they can find, without you asking them. But hopefully LG wouldn’t go that far.

most papers have slides with audio, and some, including the awards ones will have short frontal talks. this will be released at some point after the conference, but right now looks like you'd have to be registered to see it.

This isn't about me though, no point in making this personal. I'm just trying to understand what's interesting about this picture. You are saying that just because nobody took that particular combination before, it is enough of a reason, right?

A lot of logistics, as mentioned. If you’d like to explain what’s in there beyond that, I’ll be glad to hear.

Care to explain? I actually do take pictures with a camera from time to time.

Again, this was staged. Also, when Tom Cruise performed his own stunts in Mission Impossible, that value I can understand. That is better than photoshop. Because they were hard stunts. This on the other hand seems to be standard.

Tbh, do not quite get the excitement around this picture. It was staged, and the stunt doesn’t appear to be particularly complex. A lot of logistics, sure. But seems like all there is to it is that someone just bothered to do it. So not clear what’s the additional value over photoshop.

Is there a way to listen to the actual content? This is only a brief summary of the radio show.

you mean Pocket? I was wondering how well the search there works. Turns out it is possible to search the pages themselves, but only with paid subscription. In the free version it only searches the titles and urls, like the Bookmarks.

This is funny, because I recently descided to try Pocket, owned by Mozilla, and it adds

>?utm_source=pocket_saves

to all saved links.

Do you mean the boosting graphs in Fig. 1, 4 of that paper?

It looks though that they have a double descent on the train set too, so it might not be the same phenomenon.

Nevertheless, good to know, thanks for sharing! I knew both papers but never thought giving much attention to such details of the figures of the 1998 one. Is the connection between the papers well known, i.e. something people talk about?

This is like a workshop at a usual conference, no proceedings, right?

I haven't followed the original debate, so not sure if this was a part of the argument against LeCun, but it could be:

The issue is LeCun makes an argument about fundamental research, but is not exactly a fundamental researcher, and does not necessarily represent fundamental research.

As an analogy, if you are a researcher of general chemistry, persumably there is no issue. However, if your research is specifically about chemistry for improving bullets, and you produce working prototypes, then some might say you should be subject to some regulation as a part of the arms industry. I'm not saying this is the right thing, just that such a point could be made. LeCun is arguably much more the second kind of researcher than the first. The research he represents is "better ways to recognize faces", not "statistical properties of natural images".

To take another example, there is an enormous amount of regulation in, say, medical research. And in this case there are good reasons for that. Gebru could be possibly arguing for something similar in say face recognition.

Isn't it strange that the periods mentioned are integer multiples of days?

Don't known much about astronomy, but day as a unit of time is just a constant specific to our particular solar system. I guess a function of the sizes of the sun and the planets here. There should be nothing special about it. To think that 500M light years away there is something that has similar time proportions to be observed here as periodic is amazing by itself.

Not sure about that specific function. But the circle of ideas to which this function belongs can be loosely described as "how do we deal with non-smooth objects"? How do we construct them, measure and use them. Some such objects are sometimes called fractals, this function is a simple example.

There are concrete application examples, but in general one guy that was particularly interested in fractals in the real world was Mandelbrot. He published several influential books on fractals in nature (and in finance). Another famous name is Taleb, although imho he's much more a populariser than a researcher.

Not a standard source for algebraists, but Percy Diaconis's book, mentioned in other comments, is an excellent introduction. Also contains applications in probability.

This is of course just a hyperbole, especially in its original context.

For one thing, people routinely use representation theory to study groups of matrices. That is, represent the group of matrices by other matrices, acting on other spaces. If matrices were this elementary "thoroughly understood" object, one presumably wouldn't have needed to do that.

The real power of the representation theory is not that its matrices, but the notion of irreducibility. This allows you to decompose a complex action into simpler blocks and understand it that way.

Well, it depends. Some people would say, justly I think, the way to come at Fourier transform is as a commutative representation theory.

Data like this could (or might have the potential to) be used by the authorities to determine in which areas to test more. But otherwise it shouldn't send people anywhere, I don't think that was the intention.

Not a big fan of the idea of assigning credit to individual features, but as far as this idea goes, this seems like a very useful improvement.

The blog mentions an evaluation on some loan related data, and this data does not appear in the paper. I wonder wether it is available somewhere.

As an aside, it seems you're interested in speech recognition, or speech to text, not voice recognition. Voice recognition is a different problem, where the particular speaker needs to be recognized from voice.