11,504 karma · joined July 6, 2012
There's kind of a bitter lesson variant with AI coding where the work to make a simple, consistent, flexible, DRY platform or library isn't really worth it when anyone can press a button and get a variant customized to their needs in minutes.
People playing the game don't care about the elegance of the emulator, and if there is a bug in the slop, they can just have their AI fix it immediately and keep playing.
If not now, then when will these companies be AI leaders?
Even Google, with its staggering advantages in cash, compute, real estate, training data, and having basically invented the field only manages to briefly claim a 1-2 week lead once or twice a year.
> An individual who, with no resources or special standing or institutional support, simply goes online, does his research, and starts writing,
could apply to a huge range of bloggers, thought leaders, substackers, and straight up shitposters, many of whom get even more attention than Ed. In fact, this type of personality is much more influential and popular these days than they have ever been.
If this person like's Ed's writings, and believes that his predictions could still pan out some day, I'm not going to take that away from them. But coming up with this strange criteria of being a "last great blogger" that Ed doesn't even meet that well seems like a sleight of hand.
> Zero-shot vs. Fine-tuning: Out-of-the-box base models score ~0.35 on the typed-decisions benchmark (near random). The 0.766 score is achieved by fine-tuning on the benchmark's train split. Treat Laya as a fast foundation model to specialize, not as an omniscient zero-shot oracle.
Even though I had been well-disposed to rationalism and read HPMOR to learn more, I could not take Yudkowsky seriously in any way after I gave up halfway through.
The debate has been taking place on the actual street?