Login
You're viewing the discuss.systems public feed.
  • Sophie Schmiegsophieschmieg@infosec.exchange
    Oct 8, 2026, 8:41 PM

    I am an incredibly minor public figure. One thing that, even at my incredibly narrow status, I learned, is to not speculate on main, when you have absolutely zero evidence to back up any of your speculation.

    Unfortunately not everyone in this position has learned that lesson.

    In case you've seen some wild cryptography related speculation today, my status, at least at this point in time, October 8th, 13:40 PT is that absolutely no one who has the expertise to evaluate it has seen any result that would be in the least concerning for the state of public key cryptography. That might change, of course, but that is always the case.

    💬 4🔄 35⭐ 39

Replies

  • Sophie Schmiegsophieschmieg@infosec.exchange
    Oct 11, 2026, 7:14 PM

    Oh God, this is getting even worse: apparently the "source" of the claim that AI might have broken classical cryptography is the fact that some numbers are skipped in the OpenAI math dump. You know, that same dump that doesn't even have formal verification on more than half of the results and some of the formally verified results have been shown to be incorrectly formalized?

    I dunno, but isn't the far, far more likely explanation that the skipped numbers simply contained obviously false statements and were removed because they were, well, false?
    Brother Occam, we're going to need a bigger razor!

    (Also it is now Sunday, the 11th, 12:15 PT and I still have zero evidence that anyone with the qualifications to review it has seen any claim whatsoever that would put public key cryptography at risk that had not been known already)

    💬 3🔄 7⭐ 15
  • Sophie Schmiegsophieschmieg@infosec.exchange
    Oct 11, 2026, 7:27 PM

    "but AI is good at math now!"
    Yeah, you know who else is good, arguably even a lot better at math? Humans. Volume in proved statements isn't going to get you the type of complexity theory shattering results that you'd need to break public key cryptography.

    Yes AI is advancing at fast speeds. This statement might look silly at some point in the future. But at least as of now, there is more risk of a major breakthrough coming from some underpaid PhD student than there is from AI. If you don't consider the risks of paying PhD students more might have for cryptography, you should not worry about AI.

    Also pay PhD students more, please.

    💬 2🔄 6⭐ 11
  • 💬 2🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • Justin DerrickJustinDerrick@infosec.exchange
    Oct 9, 2026, 2:48 AM

    @sophieschmieg @starchturrets I totally respect this way of dealing with it.

    Same as folks announcing the names of fake websites — don’t name the fake one, name the REAL one. Don’t give oxygen to dubious sites (and/or claims).

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 3⭐ 3
  • 💬 2🔄 0⭐ 0
  • 💬 0🔄 0⭐ 8
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • ohirohir@vivaldi.net
    Oct 11, 2026, 7:55 PM

    @sophieschmieg
    > Brother Occam, we're going to need a bigger razor!

    The current razor is long enough IMO.

    While what they published is lacking merit, the publishing itself has it – yet another 'investor' will part with their real money ;)

    💬 0🔄 0⭐ 0
  • Anthonyaps@mathstodon.xyz
    Oct 12, 2026, 10:33 PM

    @sophieschmieg That reflect poorly on OpenAI too, in my opinion: they could have at least used consecutive numbering if they didn't plan to give a full account of their tests. For that matter, they've actively made it difficult to contact the people involved to ask for clarification (according to <mathoverflow.net/questions/515>). So much for "scientific transparency and openness"!

    💬 1🔄 0⭐ 0
  • Sophie Schmiegsophieschmieg@infosec.exchange
    Oct 13, 2026, 12:16 AM

    @aps ugh, so that's even more allegations of plagiarism and bad scientific conduct.

    It really starts to look like LLMs are great at math when you give them the opportunity to look up the answer.

    Oh and I learned today that apparently they also have duplicate numbers in the dump.

    Maybe the LLM should study Peano a bit more deeply

    💬 1🔄 0⭐ 0
  • MissConstrueMissConstrue@mefi.social
    Oct 16, 2026, 9:23 PM

    @sophieschmieg @aps

    I am not a maths theory kinda person, please forgive what may be an illogical question.

    LLMs are engines built by extracting human data and incorporating it into a data set. (simple version). They are predictive. They *guess* what the next token should be. (I realize decision models operate differently but those are new and untested and not the models used in any of these use cases.)

    If I am entering data into, say a spreadsheet or accounting software, and I sum a list, it is not guessing, it is doing predefined calculations on the data provided. Excel is unlikely to come up with a new theory of cost analysis just because of my creative data entries.

    How could any output from a predictive engine be trusted to be true? It is not calculating anything. It is guessing. Yes? Or have I failed to understand something fundamental? That happens. :)

    💬 1🔄 0⭐ 0
  • Sophie Schmiegsophieschmieg@infosec.exchange
    Oct 16, 2026, 9:55 PM

    @MissConstrue @aps I mean yeah, that's definitely part of the problem. The newer results overcome this by using formal verification to check whether a given statement is true, which is definitely a much better approach.

    💬 1🔄 0⭐ 0
  • MissConstrueMissConstrue@mefi.social
    Oct 16, 2026, 10:09 PM

    @sophieschmieg @aps

    As I understand decision models (I don't), they are significantly better at verifying within given parameters, so I could see value in providing a hypothesis, and verifying veracity...and perhaps I have an overly romantic view of how mathematicians think, but inspiration is a uniquely human trait, I would argue, and that all breakthroughs are a result of an "Ah Ha!" moment when a human brain brings together a seemingly disparate set of information to create a unified concept.

    So, I guess I tend to view any "breakthrough" announcements by frontier labs that their LLM has done something no human could do, to be unlikely at best, made up from whole cloth at worst. (AGI claims, I'm looking at you.)

    In any case, pay PhD students more, and give them a career path. I have more faith that an exhausted kid will see the brownian motion in her coffee cup and discover new theory than I do that a resource furnace will do it.

    💬 0🔄 0⭐ 0