Notices by Graham Sutherland / Polynomial (gsuberland@chaos.social), page 2

@thomasfuchs @whitequark and... huh? it literally never covers my feed?

@whitequark I'm not sold on it because it appears to cover the feed while composing and I do a *lot* of cross-referencing the thread when writing, so that would be really annoying. (on tusky my average post gets put into my clipboard at least once so I can close the posting, re-read something further up, then continue)

finally, and at long last, it is raining

@mcc to cut it to a soundbite: both systems are producing imprecise translations, but one of them is more honest in its representation.

@mcc almost certainly. they'll be feeding the from/to languages into the prompt, which will skew the model output.

@mcc it's sad that an org as large and well-resourced as google couldn't envision this exact failure mode, given that it's one of the first things you figure out after building a mental model of how ML models (classical and LLM alike) represent the state space.

@mcc precisely. it's the wrong tool for the job. LLM translation produces text that is more intelligible to the reader, but with more frequent and more significant deviations in accuracy of translation compared to traditional approaches. that's truly the worst outcome, because the vast majority of users cannot possibly vet the translation *and* come away feeling more confident in the translation's accuracy because it reads clean. the wonkiness in an imprecise translation is a critical signal.

funny thing in rimworld yesterday night. got a quest to look after 8 people for 14 days, in exchange for a fair chunk of silver and a stack of glitterworld medicine. and not only that but they'll work and fight like regular colonists while they're there. neat! my colony's ideologion despises long range weapons, but these guys love 'em, so it's a perfect match. they turned up in shit gear (ragged clothing and sticks) so I got them running some .338 magnum rifles and flak gear. the days go by.

I have been getting this on and off for weeks now. the heck is going on over at LCSC?

@whitequark oh I definitely don't see those as bloat

@whitequark and my complaint isn't even bloat when I think about it. it's mostly new immature feature sets replacing established capabilities, messing things up for people, not because there was a strong need for a replacement but because it's what the devs felt like tinkering with

@whitequark redesign for redesign's sake and such

or feature churn

the reason bsd rules is 'cos it's maintainable and understandable and doesn't catch fire every two fucking seconds because someone's pet redesign project was just sooooooo important to be in mainline

now that y'all are ditching linux cos of the ai stuff you better not ruin freebsd by bringing all the godawful governance problems with you

the attack is a single xor instruction to flip one bit at the config address.

doing useful stuff with it involved using z3 to solve the mapping delta between the two scrambler modes, but that's far from complicated. with that result you can figure out the physical address translations (i.e. what address moves where after the scrambler mode change) and just access whatever the hell you like.

scrambler settings are often exposed in the BIOS UI, but being able to change them directly in a running system from ring 0 with an MMIO poke is mind-blowing. I would've expected it to be handled by the SMM, or a one-shot ring 0 wrmsr at boot.

a few people have asked, so:

if you're an end-user, you almost certainly don't need to worry about this.

probably the only widely applicable security impact from this is against laptops with AMD CPUs from 2014 or earlier, with no discrete TPM, running Windows with BitLocker in fTPM-only mode (no password entry at boot; login password only), where evil-maid style attacks are in the threat model. so still rather narrow.

it also impacts things like TEE and remote attestation, but if you're working with that stuff then you don't need me to tell you that.

this bug is much more interesting from a platform security research perspective, given how easy it is to exploit when you own the hardware and you're actively trying to dump firmware stuff and processor internals you can't usually access or interrogate.

it also very likely has some utility in the PS4 and Xbox One homebrew scene, since they use the same family of processors as the ones discussed in this research. but I am not a console hacking guy so I can't tell you anything more about that.

@dalias you and I fundamentally disagree on this so I don't think it's worth us getting into it yet again.