Also: asked the AI to implement HVM5 in Bend2 and it worked! 2x-3x slower than the manually optimized C, but the fact it worked increases my confidence on Bend's compiler. Also crazy how we're getting used to AI doing that kind of monster work...

I've asked the AI to implement HVM5 in Bend2. It worked!

It is 2x-3x slower than the C version though. But that's a manually optimized file, by me and by dozens of AIs, including all sorts of micro-optimizations. So not bad, and increases my confidence in Bend2's compiler

@Ceoz_1 it is ok to be in denial

I'm almost reaching acceptance

@DaleCloudman the code was just perfect. everything from the prompt was there, flawlessly. the prompt demanded very complex features, things that should take several days to develop, and it all just worked

@EttoreMariotti a next gen ML algorithm will probably need 100x less data, so ~4 TB, which fits in a laptop. also, that data is just for knowledge. strong reasoning should not depend on that much data

@VictorTaelin Sorry Taelin I was wrong, it was about 75% marketing :(

@VictorTaelin For mythos on a laptop you still need data, it’s probably not just a clever algorithm

@wweb_aristocRAT @DaleCloudman I don't know, I've just seen the output. I didn't use it myself

@VictorTaelin What made the mythos output so stunning?

@VictorTaelin 4.8 on max effort is very self conscious on the scope and where it stands no idea why you had a different impression of 5.5 at any moment of time

@VictorTaelin @DaleCloudman How long did it take?

@VictorTaelin No denial, I just don't trust Anthropic or anything they say, and I thought you had reached acceptance about that.

@VictorTaelin how are people getting access i'm mad

@Ceoz_1 but I never said Mythos was solid because I trusted their word

I said so because all the evidence pointed to it

@VictorTaelin that’s a bold goal. training something that big on a laptop sounds like a nice dream, though.

@VictorTaelin hvm5 at 5x is solid work, buried under an AGI redefinition in the same post. now I can't tell which claim I'm supposed to be skeptical of first.

@VictorTaelin 4.8 is mostly on autopilot here grinding through the task list https://github.com/GrigoryEvko/FX/commits/master/

@VictorTaelin Maybe you are right, I like to believe that. I was thinking about knowledge, but maybe that’s indeed distillable in few TB

@VictorTaelin One type of strength and wisdom is the ability to admit things even if we wish it weren't so. (the earlier the better)

@VictorTaelin My feeling lately. Going back to Claude and might as well resign if 5.6 sucks compared to mythos