@teortaxesTex why only 5 of the 13 RULER tasks? They left out the aggregation tasks, where the model has to use the whole context instead of finding a needle, exactly where you'd expect a cheap shortcut architecture to fall apart
@teortaxesTex The 10M is also kind of inflated. 10.2M tokens = 23MB, so ~2.4 bytes per token. The other models were getting 4+ on the same code. In normal tokenizer terms it's more like 5M, and every byte of code costs you twice as many tokens
@TheStalwart But what if that PHD was also a PHD in everything else there is to be a PHD in and they were also the best PHD in the world at all those PHDs and could work 24/7 1000x faster than any other PHD?
@JamesSurowiecki Why couldn't you win with Bryce Young? He's looked better than many of the top 15 on this list in most of his games this year. You don't think Bryce Young has four games in him where he can play at least as well as Sam Darnold or noodle-arm Peyton Manning?
@robbensinger@pnwsnorlax_@DevyDegaimer@schw90 If the robot can walk into my house and I can tell it that I have replacement bulbs in a drawer that it can use to replace my light bulbs in my house, and it does it
@robbensinger@pnwsnorlax_@DevyDegaimer@schw90 Can the same robot walk down my stairs, open my door, open the drawer, take the light bulbs out of the packaging, walk back upstairs, Identify the location of each of the bulbs. Reach the 8-foot ceiling, successfully replace all the bulbs, and dispose of the old ones?
@robbensinger@pnwsnorlax_@DevyDegaimer@schw90 10/ My odds on that: 0.5% by 2030, 5% by 2035, 10% by 2040, 15% by 2045. And AI doesn't need to beat us at physical tasks to kill us. I think a bioweapon is a much likelier path than robots.
@robbensinger@pnwsnorlax_@DevyDegaimer@schw90 9/ My odds that AI is vastly superhuman at all tasks, physical included: 1% by 2030, 10% by 2035, 20% by 2040, 30% by 2045. That's conditional on AI not killing all or most of us first.