Given that all the technical details boil down to "I don't know, an LLM wrote it for me", the article can be loslessly be summarized to a single word: "yes".
On the one hand, I'm glad the author is upfront about it. But on the other, it sucks the joy out of it for me. Of course, they aren't writing for me, so what I feel doesn't matter – that's fine.
I can best describe it like hearing about a dream someone had last night, versus hearing about how their day was.
Well, not completely; it does provide e.g. the split that moving to a weak Pentium-crunched neural-network evaluation does nothing, and the later direct gain from a better-taught NN is larger than from speedups when you start using it for pruning. So, the impact and the interdependencies of the algorithmic improvements are given.
Interesting! It's easy to get the perception that "new algorithms are better but too complex for older devices which is why they weren't used", but it turns out that sometimes new algorithms are just always better haha
And sometimes, amusingly, the opposite. There are at least a few examples of "surprisingly advanced" algorithms that were invented in the 1960s and then forgotten because nobody had the entire megabytes of RAM that would have been needed to put them into production. Then they stayed forgotten until the 2000s or so when someone said "well what if I just precompute this... and I can optimize by dynamic-programming that..." and a little work later it turns out that yes, that idea does work but someone had it 50 years prior :)
hjvt | 8 hours ago
Given that all the technical details boil down to "I don't know, an LLM wrote it for me", the article can be loslessly be summarized to a single word: "yes".
gspr | 8 hours ago
On the one hand, I'm glad the author is upfront about it. But on the other, it sucks the joy out of it for me. Of course, they aren't writing for me, so what I feel doesn't matter – that's fine.
I can best describe it like hearing about a dream someone had last night, versus hearing about how their day was.
k749gtnc9l3w | 4 hours ago
Well, not completely; it does provide e.g. the split that moving to a weak Pentium-crunched neural-network evaluation does nothing, and the later direct gain from a better-taught NN is larger than from speedups when you start using it for pruning. So, the impact and the interdependencies of the algorithmic improvements are given.
hobbified | 47 minutes ago
This is clearly untrue. The author knows what they're talking about, and knows how to write a chess engine.
bo0tzz | 10 hours ago
This is awesome, and it has me wondering what the most minimal system that could defeat a GM would be.
gunduzc | 10 hours ago
This is amazing. I would pay money to see Hikaru or Magnus go against a Pentium.
nytpu | 8 hours ago
Interesting! It's easy to get the perception that "new algorithms are better but too complex for older devices which is why they weren't used", but it turns out that sometimes new algorithms are just always better haha
hobbified | 43 minutes ago
And sometimes, amusingly, the opposite. There are at least a few examples of "surprisingly advanced" algorithms that were invented in the 1960s and then forgotten because nobody had the entire megabytes of RAM that would have been needed to put them into production. Then they stayed forgotten until the 2000s or so when someone said "well what if I just precompute this... and I can optimize by dynamic-programming that..." and a little work later it turns out that yes, that idea does work but someone had it 50 years prior :)
JulianSildenLanglo | 13 hours ago
Cool stuff, and I like the demo. It definitely defeated me, though I'm admittedly 1-2 kELO under Kasparov: