> [The model's authors] can do this kind of work with Claude, but if you or I try to do it, we will get blocked.

If this seems concerning, wait until the models get effective enough (gets work done properly at a lower energy need than humans have) that there's no reason for them to release anything. Why offer your weights for download or grant any third party usage access when you can just do the actual job yourself and collect all that turnover?

Maybe, but that's what the cyber program is for. Nobody knew if the seed was fine or a backdoor. If it was a backdoor, and Claude helped you find it as a normal person, Anthropic would get bad PR. If you're a verified security researcher, actually that's fantastic and Anthropic would get good PR and everybody claps.

Anthropic wants to IPO, they don't want another Fable shutdown situation.

To be clear I would prefer the cyber safeguards to be less restrictive, I think they're currently too much. I often go to open models because of it. But there are likelier reasons to restrict cyber than world domination.

Anthropic gets zero of my money because their favorite word is no. I cannot figure out how to as an individual or a small startup request elevated permissions for the purposes of anything that they think might potentially be dangerous for someone who is not me to access. I have a background in cybersecurity. I have a background in life sciences. Can I put those skills to use to benefit humanity? Who knows.

I worked on this specific task myself, using Claude, and I was not blocked (although I did not succeed!)

I feel like history is repeating itself - to many people AOL, Compuserve, MSN, et al seemed superior to "the web" in the early 1990s because they dumped money into making it slicker, licensed content, provided guardrails, made it easy for non-tech people... They had slick ads, were cultural phenoms, and people would take sides with the companies and think their arguments were strong and even obvious!

I'm skeptical that this particular task would get blocked, though I could be wrong. But I agree that the moat is a problem in the general case.

As much as I dislike the guardrails, they are absolutely irrelevant to this task, and you will not get blocked.

Why would you be blocked? I cannot imagine why you would be for this specific task unless the supposed backdoor turned out to be real.

I know Anthropic models have a lot of insane, paranoid filters, but I don't remember them blocking cryptography work last I checked (some 3 months ago now).

Anyone can do this kind of work with a model as simple as GLM 5.3 or Kimi K3

How do you know this to be true? I am not disagreeing; I genially need to understand why or if others are seeing blockers that I don't see or if I am seeing blockers that they do not see.

Are their guardrail benchmarks or anything of the sort that we can use to see where the bifurcation of permission/capability occurs?

> moat widening between the model providers, their government partners, and the public.

Part of me wonders whether the NSA/other TLA has either (a) an unrestricted model from a lab with less scruples like SpaceX, or (b) an in-house model, either abliterated or with data they've bought themselves.

Compute has always been gated. This is nothing new. If we look at the HPC space, many of the largest clusters were restricted access and used by governments organizations. Even the ones with "public" access, not just anyone could get access to it.

Open models are just as quick to refuse to do stuff, in my experience.

JFYI, this is not true. I do stuff like this all the time.

Whenever I say that the rich have access to much more capable and unrestricted models I get downvotes. Yet here we have another example.

> While this confirms the method of production it doesn't say anything about their selection, especially since they were, apparently, generated over a period of months. Specifically you cannot see those candidates that were generated and discarded.

Well, at least now we have a way to perform that analysis! Generate a bunch of curves by varying (along any degree of freedom) the input parameters to the seed generating function and see if there's any difference between the curves that are generated.

God bless you for keeping this alive.

If you keep posting like this you're gonna get microwaved from space ;)

"However, you would only need to do that if you were going to disclose the timestamp as some evidence that the seed was innocuously generated."

And the disclosure happened now through Anthropic that is 100% not going to anger the NSA before the IPO and will have permission.

The two people were Bob Reiter and Jerry Solinas. Jerry wrote in a 2015 email to Daniel Bernstein that he thought it was a hash of the passphrase. I corroborated that story with 4 people, including two ex-NSA folks off the record.

Everyone was wrong!

I wasn't able to reach Bob. He was working in private industry as of 2018, but never responded.

Someone just posted this:

https://www.youtube.com/watch?v=eOQfNT6Jdkg

This is also why one of the Zodiac ciphers remained unsolved for so long. The dude made a mistake in his transcription.

But the friends we made along the way!...

> 13 years of tinfoil hats for srand48(time(NULL)

Probably not quite this given the timestamps span over almost a year (but have some clustering in August 1997).

Given that the spread suggests they weren't just now() timestamps it would actually be interesting to know what the significance of the timestamps is, if any. Though it could be something boring like randomly sampling some local file creation times.

I see it as, they're trying hard problems that aren't in the training data, to gauge if their models are improving. The evidence is everything they've released in the last few weeks, being somewhat broad in topic.

It’s perplexing that NIST chose to burn so much trust and goodwill with the independent cryptography community instead of just explaining empirically that they really had nothing up their sleeve.

Did that clanker call us dumb? Scorch the skies, this means war.

Extremely unlikely - these are very sparse domains.

This happens in academia all the time. You read a ton of papers, chat with tons of people about tons of ideas.

Later you come up some idea, you do the experiment, write the paper, search for relevant citations when you write the intro.

Honestly I think this is actually desirable. For a mind to work well and do things creatively and efficiently it can't be a simple search engine, it needs to blend and meld and reform the information.

And that means the output can't be cleanly deterministically linked to the input. Unless you just keep a log then cite everything you've ever read in your life (and all the conversations and talks too).

Sounds entirely reasonable to me. They tried to give some credit. The main contribution is the solution that Claude came up with.

I have seen nothing that would make me believe there is a "[gag] by the NSA" preventing frontier labs from publishing cryptographic results.

We published HAWK and McEliece attacks no problem.

The conspiracy theory going around the Internet is that they (the AI labs) broke cryptography in some way a la last season of HBO's Silicon Valley, and they're keeping the results under wraps until they can steal Bitcoin out of Satoshi's dead wallets, and then use that to pay for GPUs et al.

Yep.

NIST closed it: https://www.muckrock.com/foi/united-states-of-america-10/ori...

NSA ignored it: https://www.muckrock.com/foi/united-states-of-america-10/ori...

I think part of it is I don't know how to write these FOIA requests in the right way they can comply with it. I think you need to be more specific.

probably something like:

"read this: https://words.filippo.io/seeds-bounty/ then find the solution to the bounty"

> the "ancient prng + date" angle

I remember hearing this in a cryptanalysis lecture at some point, so it was not a completely novel theory.

Admittedly I am not finding written evidence of this anywhere.

> 2. Whats with all the "hmm..." in the internal reasoning process log? Do models really do that? And if so, why? Some weird emergent language thing, or just mimicry?

Assuming that is actually the real reasoning trace (open weight model reasoning traces usually do things like switch languages or talk in non-standard English because of model drift, so it's a bit surprising that Claude wouldn't do the same) there was a study a few years ago that found base models produce better answers if you put "take a deep breath" in the prompt[1]. Maybe in the RL training stage stuff like that gets reinforced and because reasoning traces are hidden in RLHF it doesn't get selected out? My money is on it not being the real reasoning trace though.

[1]: https://arxiv.org/abs/2309.03409

From the video posted upthread, the vibe I get is that they prompted Claude about the bounty and to approach it from first principles rather than building another clone of John the Ripper.

Nobody conversant with elliptic curve cryptography research believed there was a backdoor based on these seeds; for a number of reasons, best articulated by Koblitz and Menezes (who took on the question directly).

But it's nice to be able to shut the whole unproductive debate down directly.

If there’s one thing everyone knows about conspiracy theorists, it’s that they’re willing to accept concrete disproof of their theories. /s

The process to generate these was opaque, so there was some concern that the values might have been backdoored somehow like Dual_EC_DRBG was; we didn't even have confidence that they were generated from a PRNG rather than some other process. Knowing that the values were generated this way makes it much less plausible that there is such a backdoor.

This is discussed in the post - since the NSA never mentioned how these seeds were created, it seems unlikely that they would develop a cover story that would not be shared and only opened to scrutiny many years later

Yes. I did speak to a cryptographer.

>Uh… did you talk to any cryptographers

google some of the names under the "Who we are" heading