datarama

datarama, 20 days ago

@drahardja But if we have an open knowledge base, AI companies will just take that for free.

datarama, 20 days ago

@drahardja That depends on whether it's illegal at all, and at this point, a lot of jurisdictions haven't actually made that decision.

(I think "copyright doesn't apply if you're a corporation with enough computational resources, but it does if you're just a person" is the most ridiculous possible interpretation ... but I also have to realize that judges and politicians can be bought.)

datarama, 20 days ago

@drahardja In the EU, the situation is currently that AI training systems have to "respect a machine-readable opt-out" if training commercial systems (copyright explicitly doesn't cover training research systems, which is probably why eg. Stability funded a "non-profit research lab" to do all the actual data gathering for them).

But the only current way to make such an opt-out is a flaky W3C proposal which is easy to circumvent (https://www.w3.org/community/reports/tdmrep/CG-FINAL-tdmrep-20240202/).

datarama, 20 days ago

@drahardja Let's say, for example, that I self-publish some free code on my website and set up TDMReP to send its machine-readable response to all requests for it.

Some guy takes a copy of my code and puts it on Github. Now TDMReP no longer sends a machine-readable response for the version there. So now it gets shoved into Copilot and Starcoder and all the rest - and none of them broke the law, since I wasn't sending them a "machine-readable opt-out".

datarama, 20 days ago

@drahardja I'm not sure if there is even any way to mount an actual legal challenge to this, or if we just kinda have to accept that any human creative work in the future is just free training data now (unless we guard it and keep it secret - and then what's the point?)

datarama, 20 days ago

@mcc @chris I've said it before and I'm sorry if I sound like a broken record:

Then they'll just scrape from the Stack Overflow replacement. Any creative works any human ever puts on the internet again is just training data now. There is no way we can share code with each other anymore without also giving it as a free gift to Sam fucking Altman and his ilk.

datarama, 20 days ago

@Richard_Littler They have unwittingly created a perfect anti-advertisement for the entire tech industry.

datarama, 20 days ago

@allpurposemat A 100 Mbps ethernet adapter is a good deal faster than the junk going out of my house!

datarama, 20 days ago

@allpurposemat I live in a city centre in a Nordic country; it's just that fibre rollout has been very, very patchy. (It's very expensive and time-consuming to dig here, compared to suburbs and the countryside.)

datarama, 28 days ago

@skinnylatte 16yo blue-tongue skink sends a friendly blep.

datarama, 29 days ago

@thomasfuchs I see a lizard, I favourite.

datarama, 29 days ago

@gabrielesvelto I'm guessing LoTR mountains are dwarf molehills.

datarama, 29 days ago

@gabrielesvelto My favourite Generic Fantasyland Dwarf paradox is this:

Everybody knows they're famous for dwarven ale!

But ... they live underground. Where do they grow the barley? Is the famous stuff all made of imports?

datarama, 1 month ago

@lauren I'd guess someone openly advertising being a cruel sociopath would improve their chances?

datarama, 1 month ago

@thomasfuchs 404 Media wrote about it about a week ago. It is, unfortunately, quite real.

datarama, 1 month ago

@hrefna men... don't have pelvises? what?

datarama, 1 month ago

@Craigp The slightly less stupid version: Life in the Permian (and earlier) was very, very strange to modern-day people, because such a huge percentage of it died out without leaving any descendants. So it's weird to us; there are no living relatives to contrast with.

If I could time machine back for a wildlife-watching trip (assuming I could be kept safe), I'd skip the dinosaurs' heyday and go to the Permian.

datarama, 1 month ago

@Craigp (I had a game setting idea a long time ago: A sentient proto-civilization - paleolithic-level - arises in deep prehistory; either the Cretaceous (for easily recognizable dinosaurs to hang out with) or the Permian (because the Permian is awesome). You'd try to make your little nonhuman culture robust enough to survive the coming horrors and hardships.

But I could never think of a game mechanic I'd actually like.)

datarama, 1 month ago

@Craigp Part of it, for me, was to explicitly lean into the non-human aspect of it. What would a paleolithic society be like for an egg-laying (and perhaps exothermic?) creature? Could there even be an analogue to the neolithic revolution(s) for an obligate carnivore?

(or I suppose you could do something similar for the far future. Humans are all dead; what'll the next sapient be like? I'm thinking descendants of crows or perhaps parrots.)

datarama, 1 month ago

@eniko ...except that computed GOTO itself isn't standard C. It was popularized by GNU C, but IIRC Clang has an equivalent mechanism.)

datarama, 1 month ago

@eniko Truth to be told, though: I just ran one of my old implementations of this on an Intel i9-13900K and there's absolutely no observable difference between that and switch dispatch anymore. Branch predictors on x86-64 got really good.

I haven't benched on ARM.

Back when I started writing direct-threaded interpreters (2005ish), you could get very impressive gains over switch dispatch.

datarama, 1 month ago

@eniko My latest project was an odd interpreter which interpreted tree-shaped code, but using an explicit VM with two stacks and a little bunch of registers. The idea was to dynamically rewrite trees using runtime type and value information and have something that could JIT really well. I got it to run almost as fast as linearized bytecode before JIT'ing.

Then I got terribly depressed and unable to code on that sort of thing for a while. And now I'm too busy with actual work.

datarama, 1 month ago

@eniko Just bytecode. Mine was odd because it wasn't in a linear format, but explicitly tree-shaped.