snewman

(I'm the author of the blog post)

My thinking here is that Waymo has logged hundreds of millions of miles at this point (and even more in sim), and there are a lot of nines in their safety record. So even the rare edge cases should have come up. You make a good point that the deep water scenario may be not only rare, but also under-represented in the training data. On the other hand, you'd think they would have thought to explicitly train for it; it's not that obscure of a possibility.

Writing a simulator requires understanding the game well enough to spec the simulator. Acquiring said understanding - within the action limits of the benchmark – seems like the heart of the challenge, so this doesn't strike me as "cheating" at all.

Correct.

In essentially all cases where a Waymo and a human-driven car have collided, the human driver has clearly been at fault. This seems definitive and not susceptible to cherry picking.

Water on the moon is limited and difficult to collect, it wouldn't make sense to use it for industrial purposes. It's a very challenging thermal environment (baking during the day, freezing at night). But perhaps worst of all, every month there's a 14-day period with no solar power. Overall seems worse than low-earth orbit.

+1 to this. It's also visually confusing, the gate looks like it's covering two cells.

Great game! Feature request: add a button that shows my submitted solution. I'd like to be able to compare it with the optimal solution (so it'd be nice if a single tap could toggle between my submission and the optimal).

No, because there are lots of things people can do that it still can't do.

I presume you're referring to the recent METR study. One aspect of the study population, which seems like an important causal factor in the results, is that they were working in large, mature codebases with specific standards for code style, which libraries to use, etc. LLMs are much better at producing "generic" results than matching a very specific and idiosyncratic set of requirements. The study involved the latter (specific) situation; helping people learn mainstream material seems more like the former (generic) situation.

(Qualifications: I was a reviewer on the METR study.)

A few questions / thoughts:

1. I didn't see it stated explicitly, but I presume the neural net is on the far end of a radio link somewhere, not running on hardware physically mounted on the drone?

2. After viewing the FPV video on the linked page: how the hell do human pilots even come close to this pace? Insane (even assuming that the video they're seeing is higher quality than what's shown on YouTube – is it?)

3. The control software has access to an IMU. This seems to represent some degree of unfair advantage? I presume the human pilots don't have that – unless the IMU data is somehow overlaid onto their FPV view (but even then, I can't imagine how much practice would be needed to learn to make use of that in realtime).