diz

In case of code, what I find the most infuriating is that they didn't even need to plagiarize. Much of open source code is permissively enough licensed, requiring only attribution.

Anthropic plagiarizes it when they prompt their tool to claim that it wrote the code from some sort of general knowledge, it just learned from all the implementations blah blah blah to make their tool look more impressive.

I don't need that, in fact it would be vastly superior to just "steal" from one particularly good implementation that has a compatible license you can just comply with. (And better yet to try to avoid copying the code and to find a library if at all possible). Why in the fuck even do the copyright laundering on code that is under MIT or similar license? The authors literally tell you that you can just use it.

No no I am talking of actual non bullshit work on the underlying math. Think layernorm, skip connections, that sort of thing, changes how the neural network is computed so that it trains more effectively. edit: in that case would be changing it so that after training, at inference for the typical query, most (intermediary) values computed will be zero.

I dunno, I guess I should try it just to see what the buzz is all about, but I am rather opposed to plagiarism and river boiling combination, and paying them money is like having Peter Thiel do 10x donations matching for donations to a captain planet villain.

I personally want a model that does not store much specific code in its weights, uses RAG on compatibly licensed open source and cites what it RAG’d . E.g. I want to set app icon on Linux, it’s fine if it looks into GLFW and just borrows code with attribution that I will make sure to preserve. I don’t need it to be gaslighting me that it wrote it from reading the docs. And this isn’t literature, theres nothing to be gained from trying to dilute copyright by mixing together a hundred different pieces of code doing the same thing.

I also don’t particularly get the need to hop onto the bandwagon right away.

It has all the feel of boiling a lake to do for(int i=0; i<strlen(s); ++i) . LLMs are so energy intensive in large part because of quadratic scaling, but we know the problem is not intrinsically quadratic otherwise we wouldn’t be able to write, read, or even compile the code.

Each token has the potential of relating to any other token but does only relate to a few.

I’d give the bastards some time to figure this out. I wouldn’t use an O(N^2) compiler I can’t run locally, either, there is also a strategic disadvantage in any dependence on proprietary garbage.

Edit: also i have a very strong suspicion that someone will figure out a way to make most matrix multiplications in an LLM be sparse, doing mostly same shit in a different basis. An answer to a specific query does not intrinsically use every piece of information that LLM has memorized.

Isn’t it part of the lawsuit that one of the developers literally said that downloading torrents on a corporate machine feels wrong?

That they routinely use bittorrent protocol for data only makes it more willful, since they know how it works while your average Joe may not understand that he is distributing anything.

Film photography is my hobby and I think that there isn’t anything that would prevent from exposing a displayed image on a piece of film, except for the cost.

Glass plates it is, then. Good luck matching the resolution.

In all seriousness though I think your normal set up would be detectable even on normal 35mm film due to 1: insufficient resolution (even at 4k, probably even at 8k), and 2: insufficient dynamic range. There would probably also be some effects of spectral response mismatch - reds that are cut off by the film’s spectral response would be converted into film-visible reds by a display. Il

Detection of forgery may require use of a microscope and maybe some statistical techniques. Even if the pixels are smaller than film grains, pixels are on a regular grid and film grains are not.

Edit: trained eyeballing may also work fine if you are familiar with the look of that specific film.

Hmm, maybe too premature - chatgpt has history on by default now, so maybe that's where it got the idea it was a classic puzzle?

With history off, it still sounds like it has the problem in the training dataset, but it is much more bizarre:

https://markdownpastebin.com/?id=68b58bd1c4154789a493df964b3618f1

Could also be randomness.

Select snippet:

Example 1: N = 2 boats

Both ferrymen row their two boats across (time = D/v = 1/3 h). One ferryman (say A) swims back alone to the west bank (time = D/u = 1 h). That same ferryman (A) now rows the second boat back across (time = 1/3 h). Meanwhile, the other ferryman (B) has just been waiting on the east bank—but now both are on the east side, and both boats are there.

Total time

$$ T_2 ;=; \frac{1}{3} ;+; 1 ;+; \frac{1}{3} ;=; \frac{5}{3}\ \mathrm{hours} \approx 1,\mathrm{h},40,\mathrm{min}. $$

I have to say with history off it sounds like an even more ambitious moron. I think their history thing may be sort of freezing bot behavior in time, because the bot sees a lot of past outputs by itself, and in the past it was a lot less into shitting LaTeX all over the place when doing a puzzle.

Now we need to make a logic puzzle involving two people and one cup. Perhaps they are trying to share a drink equitably. Each time they drink one third of remaining cup’s volume.

Yeah that's the version of the problem that chatgpt itself produced, with no towing etc.

I just find it funny that they would train on some sneer problem like this, to the point of making their chatbot look even more stupid. A "300 billion dollar" business, reacting to being made fun of by a very small number of people.

Oh wow it is precisely the problem I "predicted" before: there are surprisingly few production grade implementations to plagiarize from.

Even for seemingly simple stuff. You might think parsing floating point numbers from strings would have a gazillion examples. But it is quite tricky to do it correctly (a correct implementation allows you to convert a floating point number to a string with enough digits, and back, and always obtain precisely the same number that you started with). So even for such omnipresent example, which has probably been implemented well over 10 000 times by various students, if you start pestering your bot with requests to make it better, if you have the bots write the tests and pass them, you could end up plagiarizing something identifiable.

edit: and even suppose there were 2, or 3, or 5 exfat implementations. They would be too different to "blur" together. The deniable plagiarism that they are trying to sell - "it learns the answer in general from many implementations, then writes original code" - is bullshit.

I'm kind of dubious its effective in any term whatsoever, unless the term is "nothing works but we got a lot of it".

I think if people are citing in another 3 months time, they’ll be making a mistake

In 3 months they'll think they're 40% faster while being 38% slower. And sometime in 2026 they will be exactly 100% slower - the moment referred to as "technological singularity".

Yeah, the glorious future where every half-as-good-as-expert developer is now only 25% as good as an expert (a level of performance also known as being "completely shit at it"), but he's writing 10x the amount of unusable shitcode.