quantumgarbage

Proprietary reasoning can be recovered from its encrypted traces. Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models. We take a trace produced by a frontier model, replay it into a weaker sibling, jailbreak the weaker model, and recover the stronger model’s hidden reasoning in plaintext, without ever attacking the stronger model directly or triggering its anti-distillation safeguards.

"it is a shot across the bow of the entire software industry"

Loved it, understanding and working through your vulnerabilities is super hard but ultra rewarding. Bold but very altruistic move to publish this

I think the point was not about truck culture, but anti-social behaviours

A result proving that the capacity conjecture of Ben-Sasson, Carmon, Ishai, Kopparty and Saraf (J. ACM '23) is false. This conjecture is being used by quite a few zkVM projects, to reduce their proof size and verifier time.

nice, my kids are going to have so many fond memories of their childhood thanks to your app

We are questioning the assumptions themselves, not the act of making them.

(I don't feel like anyone is upset?)

Sure, that does not mean all models are equal.

Economics both describe and prescribe, so those models should be evaluated on their predictive power and on whether they have actually done any good.

That isn't a vague strawman, that's a great point. Economists work from assumptions, which can be flawed in two ways: they can be blatantly wrong (see the work Kahneman and Tversky did on the rational individual hypothesis) and they can be unfalsifiable (you can't always gather data about the assumption itself). There is a good essay on this from Tirole (yet a very mainstream economist) here: https://www.tse-fr.eu/sites/default/files/TSE/documents/doc/...

Sure, show us the numbers you got from your "further readings".

Plenty of peeps are being much more factual below, compared to the gvt linguo that you are just rehashing rn

Not only tether is a "thing", its actually in the top 10 US bonds buyer. So this "thing" isn't a business anymore, but an actual, proper, geopolitical actor.

This is not an attack on hash functions in general. In this paper, the authors build a statement (or "circuit") which should, by construction, not have any accepting claim. Yet, they show that when using GKR along the FS transform, you can still get accepting proofs.

This has to do with "how an hash function behaves" in the sense that, in the context of a specific protocol (GKR), it is possible to bake in the circuit the ability to predict the randomness obtained from hashing the statement itself and the public values satisfying it.

By "easy" I mean straightforward.

Previous examples which showed how instantiating Fiat Shamir leads to an unsound protocol were so contrived that people use to think that they were a testament to how unlikely breaking FS would be [1].

In "How to Prove False Statements", you can actually build what they show.

[1]: e.g. see https://eprint.iacr.org/1998/011.pdf

The protocol is in charge of producing the proof. Fiat Shamir is a heuristic, some kind of rule of thumb, which consists in using a hash function as a source of randomness.

Cryptographic protocols often feature coin tosses. The idea is that if we replace a hash function in place of the protocol coin tosses, it should still be hard for a malicious prover to craft a false statement with an accepting proof - making the protocol unsound.

Roughly, the meat of the attack consists in baking in the statement being proven the ability for the prover to predict upfront how the hash function is going to behave - hereby breaking the Fiat Shamir heuristic and making the prover able to craft a false statement with an accepting proof.

That’s it, this is “How to Prove False Statements“!

Yes, what you are missing is that attacks on Fiat Shamir were very contrived up to now.

However the paper shows that there in fact exists a pretty simple way to break the Fiat Shamir heuristic, for a protocol operating in the RO model. And such kind of efficient attacks are rather concerning in cryptography land.

So this isn't about the attack per se, rather it's about the existence of such an easy one.

> A modern sparse Transformer, for instance, is not "conscious," but it is an excellent engineering approximation of two core brain functions: the Global Workspace (via self-attention) and Dynamic Sparsity (via MoE).

Could you suggest some literature supporting this claim? Went through your blog post but couldn't find any.

I think s/he meant swans instead (in ref. to Popperian epistemology).

Not sure though, the point s/he is making isn't really clear to me as well

Look for the "Fiat Shamir heuristic" to understand the non interactive part.

It basically consists in the prover getting its random challenges from hashing public inputs, rather than from the verifier's coin tosses.

Ofc, since approx the 80s

Ah so I was right to scroll down to find a sane take

Like anyone, I'm always impressed by geniuses displaying hollywood cliché traits. But that's even more impressive when they are great humans too.

I think you are way past the argument the writer is making.

Yes, this list reads like one a Midwestern high schooler would go through to impress his failed literature teacher, who will write him a nice recommendation letter for the ultra-conformist university of his dreams, dooming him to 25 years of debt and a miserable life working as a consultant

Some states have "alcohol-exclusion laws" which make it possible for insurers to deny coverage if you had a car accident while intoxicated.

I guess that might be one explanation?

yes agree, kinda clickbaity

title: "we found insurance fraud in our crash data"

end of post: "Does this prove fraud? Absolutely not."

lmao

On this topic, I can't recommend enough the movie "The life of Others" (2006). Depicts surveillance in Eastern Germany and the state of sheer fear and paranoia its citizens had to live in.

Again, gini coefficients or GDP growth measures are, at best, proxies to understand the conditions the bottom decile of your country lives in.

Looking at housing costs, life expectancy, food insecurity or poverty rates do a much better job at capturing this.

Switzerland and Afghanistan have an almost equal Gini coefficient.

My point is: the Gini coefficient might indicate what your country's income distribution looks like, it however does not tell anything about actual life conditions.

Only 30% of americans voted for him. Yes, that's quite a share of people ready to vote for a convicted criminal, but not to the point where I would say that the U.S. population showed actual agency in this election - and that might be the crux of the problem.

A seamless, drop-in replacement, for all python users is actually a pretty good feat tbh