Since I’m pretty clearly not in the camp of people who bought into the AI 2027…

Since I’m pretty clearly not in the camp of people who bought into the AI 2027 forecast, even though I think it’s done quite well and I respect the authors. So here, I’m digging up some crumbs of facts against it:

Meta are delaying the release of their biggest model (https://www.reddit.com/r/singularity/s/MUtbF3GS4D)
with quotes like, well, it’s supposedly basically the same as the previous one (my interpretation). Which points to a possible not-following of the scaling laws.

Also, I’ve got a really weird analogy spinning in my head. I studied a bunch of physics for quite a while (something around GR and quantum mech on the bare minimum), so here’s a hunch. The current scaling laws remind me of the laws of classical mechanics. Like, they work at a certain scale of energies/distances. But they stop working properly at extreme values (and at the very edge and for the latter there are problems too but that’s a separate topic). So yeah, it seems to me it’s the same deal with the current AI scaling laws. That for now they’re true, but as compute and number of parameters and whatever else grows, other details might show up. Don’t judge too harshly, I was trying to throw some philosophy at the fan.