[foolishbard]

There's a chance that these systems can actually out perform their training data and be better than the sum of their parts. New work out Harvard talks about this idea of "transcendence" https://arxiv.org/abs/2406.11741

While this is a new area, it would be naive to write this off as just science fiction.

[majormajor]

It would be nice if authors wouldn't use a loaded-as-fuck word like "transcendence" for "the trained model can sometimes achieve better performance than all [chess] players in the dataset" because while certainly that's demonstrating an impressive internalization of the game, it's also something that many humans can also do. The machine, of course, can be scaled in breadth and performance, but... "transcendence"? Are they trying to be mis-interpreted?

[Kerb_]

It transcends the training data, I get the usage intended but it certainly is ripe for misinterpretation

[sb77]

The word for that is "generalizes" or "generalization" and it has existed for a very long time.

[Kerb_]

I've been very confidently informed that these AIs are not AGIs, which makes me wonder what the "General" in AGI is supposed to mean and whether generalization is actually the benchmark for advanced intelligence. If they're not AGI, then wouldn't another word for that level of generalization be more accurate than "generalization"? It doesn't have to be "transcendence" but it seems weird to have a defined step we claim we aren't at but also use the same word to describe a process we know it does. I don't get the nuance of the lingo entirely, I guess. I'm just here for the armchair philosophy

[lupire]

That's trivial though, conceptually. Every regression line transcends the training data. We've had that since Wisdom of Crowds.

[internet_co]

"In chess" for AI papers == "in mice" for medical papers. Against lichess levels 1, 2, 5, which use a severely dumbed down Stockfish version.

Of course it is possible that SSI has novel, unpublished ideas.

[ffhhj]

Also it's possible that human intelligence already reached the most general degree of intelligence, since we can deal with every concept that could be generated, unless there are concepts that are uncompressible and require more memory and processing than our brains could support. In such case being "superintelligent" can be achieved by adding other computational tools. Our pocket calculators make us smarter, but there is no "higher truth" a calculator could let us reach.

[O_OtheGreat]

Lichess 5 is better than the vast majority of chess players

[oblio]

I think the main point is that from a human intelligence perspective chess is easy mode. Clearly defined, etc.

Think of politics or general social interactions for actual hard mode problems.

[jbay808]

The past decade has seen a huge number of problems widely and confidently believed to be "actual hard mode problems" turn out to be solvable by AI. This makes me skeptical that the problems today's experts think are hard aren't easily solvable too.

[nradov]

Hard problems are those for which the rules aren't defined, or constantly change, or don't exist at all. And no one can even agree on the goals.