Well, Debian has signed updates. I assume most other distros do. I'm not clear what the benefit of this is. Who are you going to trust to do the signing? Does the entire system need to be resigned after every update? At first glance this only works if your entire system is in one "file", and you sign that file with a unique signature. But who's going to know what the proper signature is?

Signing individual packages seems much more reasonable. Then two people can compare the signatures on a particular piece of software to determine that they are the same. (But for that, why rely on signatures when things like sha256 exist...unless that what they mean by signature, in which case the prior caveats still apply.)

*You* may know that, but that doesn't make it true. It *may* be true. At this stage I expect most attempts to use AI to fail. The same thing happened when computers came out. And again when personal computers came out.

Power new tools have a learning curve. AI is definitely a powerful new tool, but it's currently changing so rapidly that we haven't really started learning how it should not be used.

Well, the first computer game I found really attractive was Orbquest, a text game on a Unix system. No graphics at all.

Obsessing over graphic is at right angles to the quality of the game.

Well, for things headed out, power supplies are the crucial problem. For things in the inner solar system, solar storms are irregular, but *will* happen. So I'm not really convinced by this test for anything but LEO (as those are temporary anyway).

That said, as long as it's intended for "special missions" it's reasonable. But AI is power hungry compared to fixed electronics, so that's going to be a continuing problem.

The problem with testing radiation hardness that way is that solar storms are quite erratic and severe ones are infrequent...but definitely not absent.

This test may work for things intended for low Earth orbit, as those are temporary anyway, but I don't think it's going to validate other uses. A better choice would be to send something to orbit Jupiter, where harsh radiation challenges are predictably present.

While true, Idi Amin, e.g., was not white.

OK. That puts them in the same basket as Flock.

What does it matter whether or not they're white? I've seen examples from Asia and Africa that were worse.

I'd put him as one of the top 3 or 4, or at least one of the top 10, because of the power of his position, not because he's intrinsically the worst.

Well, if all cars do the same thing, you have a good point. I haven't looked in the last decade, but the last time I did there were lots of alternatives. I admit many posts over the months have been saying that this isn't true any longer...or that this is why they only drive (some model) used car.

But it's still not as bad as Flock (et al.), which doesn't even ask for permission.

But I believe that Lexis is only selling the data of those that have agreed to have their data sold, even if they didn't realize it.

It may not stop tracking, but it lets you do the search for something without losing your place.

Claiming that "Most psychology experiments are not replicated" is both true and irrelevant. Most psychology experiments are done to satisfy arbitrary rules. When I took a psych class, I had to sign up for, I think, one experiment per week. None of them were either replicated or significant.

There *are* psychology experiments that are replicated. Some survive replication, some don't. The ones that aren't replicated aren't important...and there are lots of experiments in other sciences that aren't replicated unless there's a good reason. Many are so expensive that replication is done only if it appears to be needed. (The experiment was done and it agrees with theory. OK.)

Actually those two stances are NOT incompatible. I'm not sure they're both true, but they might well be.

Finding a code exploit is a task with an easily validated result. That's where AI is best.
Writing code that doesn't have exploits doesn't have a easy check. So AI is not as good there.

You could use a harness to check for exploits after you wrote the code, but that increases both time and cost, and isn't guaranteed to find all the exploits. (People and AIs think differently, and don't recognize all the same patterns.)

This will probably change. AIs have "jagged competence", but the areas that they are weak have been getting stronger. For now, however (and as I understand it) there are still many areas where humans are more competent than publicly available AIs.

It's quite plausible that sometime next year Yang's reported claims will be appropriate. OTOH, that would imply massive other changes, also. And it's also possible that the claims will still be exaggerated.

I was referring specifically to an AI agent interface, but that does drop my estimate of how long until it appears. Now I'd say it may appear within months.