lanternfish

Extremely unlikely - these are very sparse domains.

Orthogonal means something different in mechinterp, so it'd be an overloaded term. Usually orthogonal would imply that there are actual vectors under consideration, which isn't the underlying claim.

For a given short-context (hash function seeding) you do have detectable manipulation of the completion (how they can read the marking). But, because (assuming enough entropy in context) the hashing function is decoupled from the log-probs, the perturbations are independent from the underlying distribution, so you're still sampling from the same distribution quotient some noise.

The only way you'd notice this is they weren't independent, and the most plausible way that happens is if you're re-completing pre-fills (resampling the same hash function against the same log-probs).

There are no watermark related tokens, there is no watermarked set - read the paper, it's public and not that complicated. It doesn't change the distribution of completions, and won't change the situation where one output token has the majority of probability mass.

You're missing the fundamentals here!

If it really was the case that the harness is at fault here, that'd be much more terrifying, as it would imply that current frontier-adjacent open models (like K3) would be able to achieve similar ends with a custom maladapted harness.

10 Axios's within 5 days.

Why not?

Engineering is done in the context of constraints, cost is one constraint - and its a relatively conserved constraint. Saving labor in one area allows for more care in other areas. Especially given that labor is often not cost constrained, but skill constrained, which is less elastic.

Mobile app compatibility is presumably the biggest seller.

In Iraq, JSOC operational doctrine was literally to target assassination campaigns based on 'nodal analysis' from contacts lifted from cellphones; if telegram had existed, they certainly would have used it too.

Famously, they didn't have enough Arabic translators, so Delta Force was often taking targets entirely based on reported association. They couldn't target based on language because they couldn't even tell what language locals were speaking most of the time.

That wasn't my point. My specific argument is about US operational policy on the ground in similar engagements. Based on precident, we would expect them to engage in the behavior the commentor indicts.

I dream that neither of these imperial powers - Russia or the US - will be allowed to inflict imperial violence, but I wouldn't be mistaken and assume that this military action will be any different than, say, JSOC in Iraq.

See: Panama, Vietnam, Cambodia, Afghanistan, Iraq, etc.

The problem is the same problem with crypto dao projects - cryptographic certainties only apply to mathematical structures; you can't validate that someone actually holds a quality until you can embed that digitally. That turns out to be very hard to do for most things.

In a mathematical sense - absolutely. You can dual halting problem against many very tangible qualities - like whether a (proved) statement is true or false. A (large-n) halting program is closer to an instantly halting program not just because n is always closer to 0 than inf, but because 'large n halting' and 'instantly halting' are ontologically similar in a way they just aren't with unhalting programs.

An LLM - which has functionally infinite unverifiable attack surface - directly wired into a payment system with high authentication. How could anyone anticipate this going wrong?

I feel like everyone is saying 'we're still discovering what LLMs are good at' but it also feels like we really need to get in our collective conscious what they're really, really, bad at.