A sufficiently capable AI would be able to plug itself in properly too.
One more reason to be wary of pushing for better capabilities.
380 karma · joined March 5, 2025
One more reason to be wary of pushing for better capabilities.
"Self-distillation" refers to distilling from a model into a copy of itself. Which is of limited use - unless you can steer the teacher, and want the student to internalize that steering.
The reason for doing self-distillation here is that we have both access to a richer representation (logit stream), and want to capture a richer behavior - not the answers themselves, but better reasoning techniques that are downstream from better prompts.
Solving an audio-only CAPTCHA with AI is typically way easier than solving some of the more advanced visual challenges. So CAPTCHA designers are discouraged from leaving any accessibility options.
It's a very old service, active since 00s. Somewhat affiliated with cybercrime - much like a lot of "residential proxies" and "sink registration SMS" services that serve similar purposes. What they're doing isn't illegal, but they know not to ask questions.
They used to run entirely on human labor - third world is cheap. Now, they have a lot of AI tech in the mix - designed to beat specific popular captchas and simple generic captchas.
Isn't "vitality of bees" that this method claims to improve actually supposed to be desirable to beekeepers themselves?
If existing practices are somehow radically worse, I would expect the first entity to adopt better practices to obtain a significant advantage - and the competition to copy them eventually.
I'm incredibly skeptical of any "everyone is doing X completely wrong and you should listen to ME and BUY MY BOOK instead".
Which is the kind of thing you would expect it to do.
Humans are not reliable. For every "no human would make this kind of mistake", you can find dozens to hundreds of thousands of instances of humans making this kind of mistake.
They are also known to operate on high level abstracts and concepts - unlike systems operating strictly on formal logic, and very much like humans.
This generation of AI doesn't yet have the knowledge depth of a seasoned university professor. It's the kind of teacher that you should, eventually, surpass.
A lot of digital copies are also DRM'd to shit - to obtain raw text usable for AI training, you'd have to break DRM. Which isn't that hard, on a technical level - but DMCA exists.
DMCA is a shit law that should have been dismantled two decades ago - but as long as it's around, bypassing DRM on things you own can be illegal. Scanning sidesteps that.
This is exactly the kind of attack that's used to extract DRM keys, which are normally made completely inaccessible to the user by malicious device vendors.
The name "TrustZone" is rather ironic. It's most commonly used to run DRM code the user should never ever trust.
It is, and always was a flimsy excuse to the strip user of control over his own device.
"Secure Boot" isn't actually there to protect the device from an attacker. It's there to "protect" the device from its own user. It's used to "secure" DRM schemes and App Store revenue streams.
Jared Isaacman is out, and what we're seeing actually happen now is the opposite of that. All the pork barrels are getting funded, and the brunt of budget cuts seems to be stated to be born by science missions like Roman Space Telescope.
SLS. Orion. Gateway. Ambitionless Artemis. JPL's disaster of an MSR proposal. NASA reeks of rot and decay. It's not in a good place, and hasn't been in a long time now.
If you "just continue operating", it's only going to get worse.
The issue is, if you do small targeted cuts, you'll spare the very people you want to cut. Because they're the best at playing office politics and finding ways to justify why they shouldn't be cut.
If you can't find a way to bypass that, your options are few. One of those options is the sledgehammer approach. Axe entire agencies, fire everyone and never hire of the fired people back. Rebuild an organization from the ground up, with new people and less rot.
It's what was done in ex-USSR countries after the fall of USSR. It wasn't pretty. It worked.
If you want to go beyond planting a flag, you need to be thinking of how to land hundreds of tons of equipment and industrial infrastructure on the Moon.
Saturn V isn't a very good fit for that. But SLS is much worse.
"Slicing at random" could actually outperform most other methods, as long as it's truly random. You can weasel your way out of a firing based on vibes or performance reviews - but you can't convince an RNG that its roll was wrong.
Having "skin in the game" doesn't somehow make a human surgeon more capable. It makes the human use more of the capabilities he already has.
Or less of the capabilities he has - because more of the human's effort ends up being spent on "cover your ass" measures! Which leaves less effort to be spent on actually ensuring the best outcomes for the patient.
A well designed AI system doesn't give a shit. It just uses all the capabilities it has at all times. You don't have to threaten it with "consequences" or "accountability" to make it perform better.