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.)

Obligatory xkcd: https://xkcd.com/3106/

For people who prefer reading to watching videos, I wrote a detailed account of my process for solving one of last year's IMO problems, along with thoughts on how this relates to AI:

https://secondthoughts.ai/p/solving-math-olympiad-problems

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).

The numbering jump is because there was "Claude 3.5" and then "Claude 3.5 (new)" and they decided to retroactively stop the madness and rename the later to 3.6 (which is what everyone was calling it anyway).

I'm not sure exactly what you mean by "dynamic text box" but it was just a contenteditable div. There have been at least two complete rewrites since I was involved, nowadays I believe it's a canvas with all of the editing, formatting, layout, and rendering done "by hand" in JavaScript.

This one line was like 90% of the original implementation of Writely (the startup that became Google Docs; source: I was one of the founders).

The other 90% was all the backend code we had to write to properly synchronize edits across different browsers, each with their own bizarre suite of bugs in their contenteditable implementations :-)

The article clearly states that providing enough power to run the heaters was one of the challenges that led to the death of the probe. Satellites are rarely in the shade for an extended period.

Two different people I know with pro subscriptions report not having access yet.

There isn't a uniform temperature across the entire exchanger. There's a smooth gradient extending from one end to the other. If the outside is hotter, then the inbound air gradually cools as it gives up heat to the outbound air which is gradually warming.

NDAs don't generally have an expiration date. (As opposed to non-competes and non-solicitation agreements, which generally do.) An NDA typically ends only if the information in question becomes public, and then only for that particular information.

Note that very often, the solution process for a more difficult puzzle involves going through nontrivial logic to prove that a particular square must be water. It's important to be able to record non-obvious "must be water" squares.

Often the connection between the load balancer and app backend also uses TLS. I've operated a large / complex service on AWS and all internal communications at each level were encrypted.

Of course, in principle, a cloud provider could tap in anywhere you're using their services – ELB (load balancer), S3, etc. I presume they could even provide backdoors into EC2 instances if they were willing to take the reputational risk. But even if you assume the NSA or whoever is able to tap into internal network links within a data center, that alone wouldn't necessarily accomplish much (depending on the target).

I'm saying they weren't solved until the problem composer (created and) solved them. They're not, in general, problems for which solutions have been lying around. So "these are problems that are already solved" isn't introducing anything interesting or useful into the discussion. The post I was replying to was trying to draw a contrast with chess moves, presumably on the grounds that (after the opening) each position in a chess game is novel, but IMO problems are equally novel.

It's true that IMO problems are vetted as being solvable, but that still doesn't really shed any information on how the difficulty of an IMO problem compares to the difficulty of chess play.

Why do you say these are problems that are already solved? Sure, they're often variations on existing themes, but the same is true for chess positions and, honestly, almost everything else in any field of human endeavor.

Agreed that the absolute upper tier of chess players have trained longer and harder than most or all IMO contestants. Though I do wonder which (top-tier chess or the IMO) draws on a larger talent pool. To my understanding, a significant fraction of all high school students on Earth take some form of qualifying exam which can channel them into an IMO training program.

And as far as the being amenable to brute force (relative difficulty for humans vs. computers): it seems that chess was comparatively easier for computers, IMO problems are comparatively easier for humans, and the game of Go is somewhere in between.

Contestants get 4.5 hours for each of the two days of competition. They have to solve three problems in that time, so on average you can spend 1.5 hours per problem (if you're aiming to finish all three).

That said, the gap from "can't do it at all" to "can do it in 60 hours" is probably quite a bit larger than the gap from 60 hours to 1.5 hours.

When tackling IMO problems, the hard part is coming up with a good approach to the proof. Verifying your proof (and rejecting your false attempts) is much easier. You'll know which one to submit.

(Source: I am a two-time IMO silver medalist.)

In many ways it doesn't cancel out. On Mars, things weigh less but their mass is the same. So, for example, if you are walking on Mars at a normal Earth walking speed, and you bump your shin on something, the inertia of your leg as it hits the obstacle is just as great as the inertia of your leg on Earth. In general, a lot of the little physical interactions that you rely on bone strength to get through scale according to mass / inertia (and muscle strength), not weight.

> What Boeing really needs is a complete change in management culture, as that was the real root cause for the MAX disasters, but that is impossible to enforce, you can't even really verify that it has happened.

Serious question: why not? Why couldn't we require Boeing to change its culture, and then verify and enforce that change? Verification might involve periodically interviewing employees across all levels of the organization, performing spot checks and audits to make sure that procedures are being followed (no more failing to enter a work item into the system), and so forth.

(I don't mean to imply that this would be feasible under current regulatory law – I have no idea whether that is the case or not. I'm just saying that you could imagine a world where this could be done.)

Presumably Anthropic / Claude, OpenAI / GPT, Google DeepMind / Gemini.