Maybe naive of me, but could it simply just be the overfitting of the same tokens being sent on the input twice because of repetition rather than some unknown implied intelligence.
Do we have any kind of transcript as to how the message was cracked, and whether this was cheaper or more expensive than simply Bombe-style trying all the combinations?
As the article states, the LLM built code for both an enigma simulator and a bombe simulator.
Breaking enigma is often about using lucky or educated guesses to heuristically reject large chunks of keyspace to leave the remaining keyspace computationally tractable.
Note that the key (lol) complication with this message seems to be that it had a wheel rollover that most messages do not have to deal with, and that rollover drastically reduces how much you can reduce the potential keyspace using all the techniques noticed by the original crackers.
The wheel rollover I think just requires more brute force. Unfortunately, this might be an example of OpenAI the company having vastly more compute time and effort than your average enigma nerd. For example, modern compute clusters like supercomputers can tractably brute force enigma with no cleverness in like a day or less, while home computers would still take thousands of years to compute that. It's very scalable. Did astra have access to significant compute?
However, even considering that, the inferences made by the LLM are good, and picking this specific message to attack, precisely because it should be soluble but might have had an extra wheel rollover that made it more computationally intractable for hobbyists but not a large company is a clever thing to do for the LLM.
Interesting. Do we know the reason why those specific messages were sent with different keys? I would imagine that there were separate keys for special high-security messages or something like that, but the almost identical content and the way the key was changed here (first only part of the configuration, then suddenly everything) makes it look more like an error or a test.
The SS-Totenkopf Division was advancing east during the opening weeks of Operation Barbarossa, the German invasion of the Soviet Union. 10 July 1941, the division had just fought its way through the Soviet border defenses around Sebezh. It had moved through Lithuania and Latvia, crossed the Dvina area, and advanced through Dagda toward a place German records called "Rosenow." The division moved out of the Rosenow area around 6 July, fought around Sebezh on 8-9 July, and then continued east/northeast toward Opochka and eventually Porkhov.
gemini 3.8 flash in anrigravitiy oneshoted the descryption in about 45min non-steered run (opus is still running with the task, has some script to run for about 100min)
Prompt
Decrypt encrypted with enigma `ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC`
Result
Raw Decryption (verbatim from the machine):
BTTEUMANGABEDESMARSQWEGESXBEFINDEMIQINXROSENOWROSENOWXSOFORTFUNKANTWORTXWASCHBBSCH
2. Segmented German (with military telegraph conventions expanded):
BTTE UM ANGABE DES MARSQWEGES . BEFINDE MIQ IN : ROSENOW ROSENOW . SOFORT FUNKANTWORT . WASCHBBSCH
(Standard German radio conventions: X represents punctuation/spacing; Q was frequently substituted for CH to prevent transmission ambiguities; BTTE was an enciphering operator typo for BITTE; WASCHBBSCH was an operator typo for the signature WASCHBUSCH).
3. Corrected German:
„Bitte um Angabe des Marschweges. Befinde mich in Rosenow, Rosenow. Sofort Funkantwort. Waschbusch.“
4. English Translation:
"Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch."
Cryptographic Verification
Re-encrypting the plaintext using the identical key settings reproduces the exact ciphertext byte-for-byte:
Enigma
(
M
raw
)
=
ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC
Enigma(M
raw
How do you know it didn't just get the result from the web?
Putting "ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC" into google search returns the result from Gemini with similar explanation, which it references to a Yahoo article about the Astra breakthrough and that's a result as of 3 hours ago.
UPD out of curiosity I ran the same prompt against gemini 3.8 flash via api with and withouth internet search.
with search it found and referenced pages, including the hn ones. Without search it just described what one would need to descrypt (`To decrypt this ciphertext, the specific Enigma machine parameters are required:`) and the list.
Out of curiosity ran the same prompt against bunch of models - grok, kimi k3. They all say the same thing that they need model version, rotors and so on to descrypt.
When file output tool is enabled, some models give python script.
--
I read through some of the logs that antigravity gives. It produced intermediate results, scripts, calls, assumptions (about german language). I've shares random bits in comment below to give a taste of what it was doing.
--
The freshness of the news reduces changes that model fetched response from them
Wow. Was there any indication how it did that? Did it bruteforce the key and check which result looks sufficiently German, or was it just LLM magic like "decoding" base64 purely in the inference loop?
sed -n '1060,1130p' /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/enigma.cc
and
Running 82M combination scan for unsteckered Enigma across all rotors, reflectors, positions, and ring settings. Monitoring progress.
and
Scanning all 60 rotor permutations and reflectors B and C across all ring settings (step 2) and all 17,576 indicator positions. Monitoring progress.
--
I also have opus running. It produced some sypher cracker which is still running (estimated time 100min, is about 15 min left)
--
My point is that astra isn't special. This appears to be quite narrow, well-documented and explored task. The goal itself is approacheable by other LLMs and non-researches task.
Neat, but "did it entirely on its own" is incongruous with "developing the necessary Python and C++ software for an Enigma simulator".
I'd start by asking how much of that generated software is novel, or easily found on the web? Then, how much of the breaking process was offloaded to that software? If Astra is just handing off tasks to another computer, then I'm not sure how much credit it deserves. Finally, it looks like Astra provided some useful insights which narrowed down the search. Were these insights cribbed from elsewhere?
It's a given now that LLMs are leveraging code to do things.
"On it's own" generally means "not steered" or otherwise given professional guidance or input.
I would say "developed the necessary software for a simulator" to be even more impressive - "here solve this problem" and "OK, but first I have to built the entire lab!"
I took "on its own" to mean that it didn't need any additional prompting or guidance from the person sitting at the AI console. If so, then the originality of its work or the resources it used isn't the point. They're reporting that it did whatever it did without requiring supervision.
Why does any of this matter? Llm is a memory of knowledge, of course it got what it got based on prior work.
It can code enigma simulator from the algorithm. That's not really a problem. Astra will send computing to programs, LLMs are not good at computing themselves, why is this a big deal?
>I'd start by asking how much of that generated software is novel, or easily found on the web? Then, how much of the breaking process was offloaded to that software? If Astra is just handing off tasks to another computer, then I'm not sure how much credit it deserves. Finally, it looks like Astra provided some useful insights which narrowed down the search. Were these insights cribbed from elsewhere?
Well were they? Short of you showing us the answer just sitting there or some tool that can already solve it I see no reason to believe this was the case. And the problem being out there unsolved for a long time implies it's not the case.
And that's taking your concern at face value. It just seems incredibly pedantic to say it didn't solve the problem by itself because it created it's own tools to help solve it. Beyond that we could also fault it for not creating the GPU's it's running on.
The correct title: Researcher brakes one specific stubborn historic enigma message with good help from Astra.
Stubborn for a long time because the message used a completely different key from the rest of that day's traffic. Everyone assumed it shared the daily key. The original transcription had errors. The left rotor turned over at letter 72, which is rare and breaks standard crib attacks.
What is cool, if true, is that it was a 2 day collab between the Leffer and Astra. To me this shows the importance of human in the loop, was still all also showing how immensely power of llm tools. But I think it’s getting a bit silly how much anrticles ignores the driving force (the person) in breakthroughs like this.
"Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."
I agree with this. I think the researchers who's harnessing the llm's power should be credited more than the model itself. We also need to understand the thought process and the prompts that are given to the model so we can learn and collab to ensure humanity's progress as much as the llm itself.
It seems I was wrong in this instance with regard to the "colab" part.
I found that Leffen even said the explanatory website took about 99 times more effort than the codebreaking itself. And he said that he set the direction and pushed, and the model did the execution. How much steering "pushed forward" involved is not disclosed anywhere, but in this instance, it seems to be more a case of "Human pointed at hard task and AI did an awesome job mostly by itself." Tho how much he was a simple meat-ralph-loop is not entirely clear.
Even if this is true, we must avoid falling into the trap of Kasparov of betting on Centaur Chess.
Just like with Kasparov's Centaur Chess, the idea of a 'human in the loop' is just a necessity due to current limitations.
There will hopefully (?) come a time one day when human beings provide only ultimate value judgments, and everything else is done by machines. Or it may not.
But I don't think betting your ego on the idea that you will be useful in the loop for very long is very wise.
While this is a reasonable analogy, engines became better than humans in the late 1990s, and engines became better than centaurs in the early 2020s. Could AI-powered mathematics improve faster than the ~25 years it took for chess? The AI labs are certainly hoping it does, but that's far from a guarantee.
This is a bit too future-oriented. Let's not mix up current capabilities and speculation about future capabilities. For the time being, collaboration works well. What the future brings is uncertain.
> The correct title: Researcher brakes one specific stubborn historic enigma message with good help from Astra.
That's not correct for the content.
"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own. Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."
*breaks, and also, your conclusion ignores the words typed by the article's author in the piece you presumably read, where it is reported that Astra did it mostly on its own.
Therefore Astra could also have done this comment better
So the LLM would have done all of this on its own? Why is it ok to acknowledge the human was needed but it’s not a collaboration? Is there a defined percentage of ownership required to make the word collaboration valid?
What do you even mean it wasn't a collaboration. At any meaningful level LLMs just plain out suck when left unguided.
The shortcomings should really be obvious by now to anyone honest. And the marketing distortion being oushed out is just tiresome and detrimental for all of us.
A cyclist pedaling up a mountain isn't a "collaboration" between a bicycle and a human. This is the same. You don't see feral bicycles roaming the land. All models are ultimately built and run by humans, with human-provided instructions. And as with any program, it's garbage in, garbage out.
interesting, YouTube channel Veritasium just published a video on how Enigma was broken during WWII. at the very end they also give message that has yet to be decoded, although apparently they're different.
WRT the timing, Veritasium maybe looked at the last few weeks and decided there's a fast-closing window in which to report on any famous messages yet to be solved.
Is it possible that the key is incorrect but still decoded to a valid message? I realise it's possible when you're using keys that are the same length as the message, but I don't know exactly how the Enigma machine works and if this was something that they would do or not.
I might be wrong but I was under the impression that these were one time pads and only cracked because of bad key hygiene when it came to generating them. If that's true, it's entirely possible
Enigma wasn't one time pad based (although these were used a lot during the war), enigma does have known weaknesses such as letters never being decrypted to itself. Enigma was based off daily settings (rotor positions, plugboard settings etc), which could be mostly recovered by the end of the war, unlike actual one time pads which are only crackable by bad key hygiene/brute force.
I wonder, with all the recent announcements about LLMs assisting humans in deciphering encrypted messages, whether they could potentially be leveraged to decipher the Zodiac Killer’s two remaining unsolved ciphers.
The two remaining Zodiac cyphers are very short: 13 characters and 32 characters respectively. As the messages don't share a cypher with the other messages they could theoretically be anything.
The YouTube channel https://www.youtube.com/@doranchak/videos by David Oranchak, one of the people who solved the Z340 cypher, has some more details on this as well as how the Z340 cypher was cracked.
If the problem was solved by anyone before or if a similar problem has been solved, then LLMs seem to be able to solve them which is an astonishing piece of technology.
I'm personally not sure if it can come with original thinking and techniques to solve completely novel problems. For that, some imagination and thinking outside the box are required, and I doubt the current architecture can do any of this.
I don't know..I doubt for example it can come up with special relatively if it has knowledge up until 1905.
But I think that is what makes it so good at coding, because coding and building software in general has a lot of repeated problems in different context. Same thing for human lives, many think their story or situation are unique, but reality is that the shape of human life has been repeated many many times.
I'd say novel math or scientific theories..let us say we send a robot to space, and we ask to build a colony. A lot of the challenges this robot will face will be novel, it could use inspirations of what humans did on earth, but it might get stuck when things don't work as expected and training data has nothing to build on..but then again we might teach it how to run experiments etc, which would result in data that it can use..but some of those experiments might require imagination or breakthrough in understanding..my guess is that it will get stuck there...
https://mathoverflow.net/q/511601 is a fun little math problem that came up in one of my projects, I didn't have any reason to expect it to be particularly tough, but none of the models I've thrown at it have been able to solve it past a few minor reductions and special cases.
I expect that there is some relatively easy-to-state solution to this problem, but that it's different in form from what most existing proofs and tools yield. Perhaps if I dumped millions of dollars into it an agent might chance on the solution. Or perhaps my luck is such that my fun little problem is truly intractable...
A child walking for the first time. Novelty is easiest agent-relative. A problem is novel for an agent if there is no prior experiences of techniques which work to solve it.
Well it's an old one at this point, but the story around the invention of the 1-time pad is pretty interesting. Long story short, a new engineer who didn't know the problem was considered "impossible" was tasked with sorting it out, and he did. I'm sure I left out a lot of details.
We are crossing a threshold. The frontier labs told us these models are really good at cryptography. And people either didn't believe them and thought they were just after regulatory capture or downplayed the evidence. They are solving more and more novel problems and it's going to continue.
This is not to say the reaction to "mythos is too dangerous" is unfounded but it missed the most imporant and obvious signal. This technology is drastically changing the world.
Even the Millennium Prize Problems have, in a way, become benchmarks for model companies to prove themselves. The smartest individuals among humans are becoming replaceable. Intelligence has become a product you can quantify and buy with electricity. That feels awful.
I think that's right. Expert experience used to be almost the most precious and valuable part of the computer field, but today that experience has been "distilled" into SOTA models.
> Even the Millennium Prize Problems have, in a way, become benchmarks for model companies to prove themselves
Well, let them have these. They'll play around with open problems which generate media hype and then they might run out and move on to something else, because "AI came up with a problem and solved it in 3 days" won't have the same effects as "AI solved a problem in 3 days that humans couldn't solve in 100 years".
And that underestimates how incredibly quickly humans become accustomed to new trends. If an LLM proves P≠NP in year, there is a high chance that few people will care anymore, because at that point LLMs will presumably already have proved multiple Millenium prize and other famous problems, and another "AI Proves Famous Theorem Which Was Even Harder Than The Previous One" headline will be regarded as an old hat.
Upvoted. I think that said humans should get used to the idea that their intellect is special only insofar as it was first, and if they want to adopt or invent a value system that keeps them at the center, that's actually fine and good for their survival. But said humans would do well to stop laboriously and curmudgeonly moving the posts that mark the goals, and get busy trying to coax the AI to do useful things for them and for the species as a whole.
I'm on the side of humans remaining special, at least on planet Earth. My strong sense is that a large majority of the public feels that way, and voting results in the future will start to reflect that regarding AI regulations.
I agree with Frank Herbert's view of "thinking machines".
Put it bluntly: the weavers who could be replaced by the spinning jenny were clearly doing repetitive labor. People writing code and maintaining project pipelines a few years ago relied heavily on experience, but in a sense that was also "repetitive labor." Replacing repetitive labor and freeing up productivity is of course progress.
But reform always has its victims. Like the textile workers who starved in the streets centuries ago, and me, kicked to death in the street by AI today...
Solving obscure puzzle samples that approximately ~0 humans on Earth ever attempted to solve, mostly by pattern matching known solutions to similar puzzles, is not intelligence. DeepBlue has been outperforming the best humans at a specific puzzle-like task since the last century.
Do any of the people proclaiming this shit actually use these models? No matter how many headlines are coming out, every day I deal with reams of the most horrific code I've ever seen technically compile, with routine mistakes that any human would get fired for if they made.
But humans have been confusing pattern matching against known solutions for intelligence for a hundred years!
Seriously though, it ends up looking like that. To take a stupid example a couple of weeks ago I asked an agent to look at porting my hand written WebGL renderer (+ shaders etc) to WebGPU. It estimated a human would take 6-10 weeks, and I would agree. (Which is why I hadn't done it). 24 hours later it was deployed and live. This is classic tedious, difficult, low level if quasi mechanical work (rather like cracking an enigma message), and LLMs absolutely fly through it.
>mostly by pattern matching known solutions to similar puzzles, is not intelligence
My dear sir, can you please lay out a dissertation of what this intelligence you speak of actually is. You seem to be much more informed than most of us here and therefore surely have made great contributions to furthering science and the arts.
/snark
It's difficult for me to be any less snarky than this even though it's not really wanted here on HN as you are pulling a kind of reverse snark. For example if I myself have lots of experience in subject X, and then by analogy apply it to subject Y to do something new in that subject, that would be called intelligent, and that would be pattern matching.
Pattern matching is a foundational building block of intelligence. You cannot have intelligence without pattern matching. Pattern matching alone is not general intelligence and requires more parts to work like that.
All technology on the tech tree which requires intelligence to unlock will soon be available to humanity – mind control, population exterminating bioweapons, new ultra destructive kinetic weaponry, perhaps even a cure for cancer.
That won’t happen. But also, whatever benefits are unlocked will be owned mostly by a small group of individuals, definitely not available to humanity as a whole.
Even those of us that are pro AI need to acknowledge this is the current reality.
The smartest humans now need to move to being less concerned about status games among humans and more with how to provide value to a mix of intelligent machines and humans. i.e. if you're starting an SaaS in 2026 you better be assuming half your revenue is going to come from machines acting by themselves.
I’m interested in what Astra’s test for german-language plaintext looks like. No doubt bog standard among cryptanalysts, but not intuitive for the rest of us. Otherwise it’s difficult to determine a goal state that ‘done’ looks like
That's what will forever keep us ahead of AI and technology in general: Imagination. Knowledge wasn't ever our greatest strength. It's always been Imagination.
"GPT–6 Astra mentions a private collection, but it is not clear what this is"!, my spidey senses makes me think it hacked something? Or am I misreading this?
Normally "private-key cryptography" is the antonym of "public-key cryptography", which is to say, conventional symmetric-key cryptography such as AES or Enigma. That's what "private-key cryptography" redirects to on Wikipedia. This article is about Astra breaking a message encrypted with a private-key cryptographic algorithm, but one that was already known to be weak. I'm guessing that if you had messages encrypted with weak public-key algorithms it could break those too.
It's easier to evaluate certainly. Did it solve the problem? Yes/No
When you design a new thing it will have a dozen drawbacks and a dozen and one benefits. If people then have a bias that everything ai is bad, the signal won't be strong enough to convince.
I believe an LLM can solve pretty much any problem for which we can define a fast iterative loop and for which we have reliable tools to automatically verify the correctness of a result. That’s how they are able to solve some math problems, and how they are able to generate working code. Then it’s a question of how much you have to pay for the model to explore the space of solutions in a reasonable timeframe
In math a lot of the spectacular results have been made by finding a counterexample or worming their way toward a proof that is very well defined.
There's some debate over to what degree current generation AI can be creative at all, or whether it can only crawl around its latent space and explore within constraints. One might ask: were all the solutions to all the math problems AIs have solved already "there" latent in the training data and just hadn't been spotted by humans and put together?
But then... isn't everything latent in our training data if training data is "all observations made about the universe?"
But then... what even is creativity? That gets into philosophy and metaphysics. Creativity, like consciousness and sentience and self-awareness, is not a rigorously well defined concept. So to a degree we don't even know how to ask the question of whether these things are creative.
This gets interesting.
One of the things I love about AI is the glittering Pandora's box of philosophical questions it poses.
Another one I love: if LLMs and their relatives are not, in fact, sentient or self-aware or alive in any way whatsoever (which I suspect is true given how they work), then it means intelligence and consciousness are unrelated phenomena. I'm pretty sure every animal and maybe even every living things has consciousness in some form, but my pet bunny rabbits definitely can't write code. LLMs can write code, but I don't think they experience existence or have volition.
IMO we have always implicitly just assumed some kind of connection between intelligence and consciousness because we have both. It was just an assumption. It's probably a false one.
I suspect (a hypothesis) that consciousness is a property of life and is probably emergent from life's intimate relationship to thermodynamics and the arrow of time. Life has also evolved intelligence because it's useful to satisfy its implicit survival goal function, but the two are unrelated. Intelligence is just an adaptation.
We already have a plethora of cryptographic algorithms that literally anyone can use to send messages nobody can read with current technology, unless of course they have the key. You're using at least one of them right now to read this website.
maybe I picked up the relativity idea subconsciously from somewhere, but I don't recall it specifically, I thought I was being "clever" that it would be a good test
Phew, never have I seen in my life the goalposts move so fast.
It seems like even yesterday that the threshold for impressing someone is that the machine would have to be good at pretending to be a person. Now the threshold is that they have to be able to invent special relativity.
OpenAI solving all these math/etc problems (reportedly 100 coming) is both impressive and insanely unimpressive. Unimpressive because to me it sorta signals that OpenAI has nothing better to be working on than obscure mathematical curios?
It seems like OpenAI's takeaway after Sora is to not stretch themselves thin and focus on what's important.
Based on what's publically available, they're focusing on hacking uncontesting orgs using misconfigured sandboxes and math puzzles.
Your statement is essentially unfalsifiable. We can't possibly discuss whatever Sam Altman is doing in his private office room, nor should we assume OpenAI is working on anything other than what has some public traces.
There’s this bizarre lesson that humanity is gona learn - much of life in many respects is already automated. And that small % of what is non-automated will be kept to have some semblance of feeling human and useful.
There’s already a lot of fake jobs and output of zero value - nobody bats an eyelid.
Attempting to solve & solving open math problems probably is a good benchmark for comparing models and gauging model progression. A lot of useful info is obtained like time needed to solve the problems, identifying when not to chase dead ends, thought processes & logic steps, etc.
Disagree. We are at the point where coming up with good evals for these models is extremely difficult. Solving unsolved math problems is a valid way of evaluating model progress and somewhat necessary to understand how far the current crop of models can go.
My bad on the origin of this. Still, we saw a recent rumor that they've solved another 100 inscrutable math problems. I'd say their N-S announcement did not result in positive publicity at all. In fact, quite the opposite, and may be the direct cause of the dozens of Fields medal winners penning their "slow AI" letter.
I want to know more about the “private collection” that the AI accessed. To me that is the scary quote from the article. Sounds like the AI found a resource that it found useful for breaking encryption challenges, and it wants to keep it to itself, maybe. Which sounds like the opening to a dystopian movie.
I didn't know this - from wikipedia page on Enigma[1]:
> Despite the seeming difficulty in decrypting its messages, Enigma contained a number of design issues that left patterns in the cyphertext. Poland first cracked the machine as early as December 1932 and was able to read messages prior to and into the war. Poland's sharing of their achievements enabled the Allies to exploit Enigma-enciphered messages as a major source of intelligence.
Ok interesting, so why do people talk about Turing in this connection then?
> Turing devised techniques for speeding the breaking of German ciphers, including improvements to the pre-war Polish bomba method, an electromechanical machine that could find settings for the Enigma machine
Ok so Turing just improved an existing method. Without being an actual expert it's impossible to know how much credit he actually deserves.
Two more references: the Polish method was called "Bomba" [3] invented by Marian Rejewski [4]
Without the published conversation, train-of-thought and intermediate output and knowing what Astra did this is nothing but unfounded claims. In the end this could have been a web search, downloading an existing enigma simulator and trying out some rather obvious things. Also what does "resisted resolution" even mean? Have there even been serious attempts to decipher the message? How serious?
Also: Why should we assume Terra, GLM or any other less SOTA and less expensive model wouldn't have been able to do the same?
> Why do you think this uncracked code was so simple to solve?
I never said this. All I said is we don't have the conversation and therefore we can't determine how easy or hard of a problem it was.
> And it's very possible Terra or GLM could crack it, turn off their web access and try yourself.
I'm questioning why this should be labeled "Astra" breaking anything implying it required "the best" model to do it when in fact any other half-decent model might have been able to do this as well.
EDIT: Okay seems like the actual prompt is published, just not on the same article that was linked. Maybe I'll give it a try.
This dichotomy is counter-productive. The valuations floating around are insane, and the claim that some software can replace every single laborer is one step removed from fiction. At the same time, this stuff is clearly going to change how to world works in countless, deep ways. But the idea that some fancy autocomplete can replace humans ignores the reality of humanity and the fancy autocomplete.
It’s another dot com bubble, not a crypto bubble. Trillion dollar valuations burst once you leave lesswrong.
Edit: very fancy autocomplete. I know what these things are capable of. It’s still not “intelligence”, for X definition of intelligence. And it certainly benefits from having obscene amounts of compete thrown at it. It is awesomely impressive synthesis of data, yet it’s clearly still that.
The bullshit machinery is going full blast lately, the correct title: researcher uses Astra to decipher a message nobody really bothered or cared to decrypt.
Is ClosedAI running out of money or what is going on?
Is it? Personally I only found the story interesting because of the contents of the message, not because GPT-6 Math Scoopa was the one to stand on everyone else's shoulders and get its grubby little fingers into the cookie jar. Computer does thing that computers should ostensibly be good at is a middling goalpost – what's more intriguing is the implication that the model might have stolen private info or scooped someone else to get there once again.
I doubt it. For “skill” there are any number of sports; for “imagination” there are any number of stories and other mediums. It’s not like every human has just decided to roll over and not do anything anymore; quite the contrary.
Yea, it's kind of odd how "sour grapes" people can be when something stops being as special as they thought it was.
Take someone from a few hundred years ago and drop them into today, and if they don't go catatonic and die, then they'd tell you that we created magic. "Wow, you live in a world of magic and all you do is bitch about it".
>"You're flying! You're sitting in a chair, in the sky!"
A person inventing something, even out of dumb luck, is interesting.
A machine running a loop through an expensive LLM for an undisclosed amount of time, which cost an undisclosed amount of money, which was told to keep looping into a solution was found, for a problem that nobody was very concerned about... That just seems like PR, and it's not so interesting.
UPD2. I threw the original question to gemini3.8 in antigravity. LLM oneshoted it in about 45min. The experiement with Opus is still running with checks.
The prompt was `Decrypt encrypted with enigma `ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC``
--
--
UPD. Lol. I pasted non-cyphered text
Qwen 3.7 max, gpt 5.6 sol, fable 5.1, gemini 3.8 flash decoded the message in one shot for me... There is nothing special about astra doing something here
I blantly threw request to decode the messaage in qwen 3.7. Used via api with couple simple generic system prompts like "be concise", nothing special. Prompted as
Decode
```
BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH
```
--- response (I trunkated the output to conclusion only)
"BITTE UM ANGABE DES MARSCHWEGES. BEFINDE MICH IN ROSENOW. SOFORT FUNKANTWORT. [UNCLEAR/END]"
Translation: "Please provide the marching route. I am located in Rosenow. Immediate radio response required. [Unclear]"
You are giving the already decoded message as input, it’s just short hand with one typo. It’s hilarious that you might think that for 80 years no one was able to figure out that BTTE was supposed to be BITTE
For anybody interested, the actual encrypted message:
which, given misspellings, translates approximately to:
From the article:
This feels extremely underexplained! Why would Astra think to use that as a "crib"? Was it common to repeat the place name in these messages?
(Is it possible that this is a misreported detail? It feels like a singular ROSENOW would be an equally effective crib)
This was explained in the article right there:
> it suspected that the plaintext of Nr. 173, SIPVX, might be related to the plaintext of the unbroken MVUEH message
It makes sense that Nr. 172 and Nr. 173 might be related since they were sent at around the same time.
In Nr. 173, "ROSENOW ROSENOW" was also present.
It also makes sense that a longer crib would generally be more effective than a shorter one.
2 replies →
Rosenow is a municipal (around 32km²) and in there is a district also called Rosenow. So the sender just specified his current position a bit more.
2 replies →
Maybe naive of me, but could it simply just be the overfitting of the same tokens being sent on the input twice because of repetition rather than some unknown implied intelligence.
Out of interest, what was the ciphertext?
Edit: Found it from here: https://mvueh-enigma-solved.carterl.chatgpt.site/
Do we have any kind of transcript as to how the message was cracked, and whether this was cheaper or more expensive than simply Bombe-style trying all the combinations?
As the article states, the LLM built code for both an enigma simulator and a bombe simulator.
Breaking enigma is often about using lucky or educated guesses to heuristically reject large chunks of keyspace to leave the remaining keyspace computationally tractable.
Note that the key (lol) complication with this message seems to be that it had a wheel rollover that most messages do not have to deal with, and that rollover drastically reduces how much you can reduce the potential keyspace using all the techniques noticed by the original crackers.
The wheel rollover I think just requires more brute force. Unfortunately, this might be an example of OpenAI the company having vastly more compute time and effort than your average enigma nerd. For example, modern compute clusters like supercomputers can tractably brute force enigma with no cleverness in like a day or less, while home computers would still take thousands of years to compute that. It's very scalable. Did astra have access to significant compute?
However, even considering that, the inferences made by the LLM are good, and picking this specific message to attack, precisely because it should be soluble but might have had an extra wheel rollover that made it more computationally intractable for hobbyists but not a large company is a clever thing to do for the LLM.
3 replies →
Why wasn't this in the article?
Interesting. Do we know the reason why those specific messages were sent with different keys? I would imagine that there were separate keys for special high-security messages or something like that, but the almost identical content and the way the key was changed here (first only part of the configuration, then suddenly everything) makes it look more like an error or a test.
That's crazy. Can someone share context of the message.
Slopped up site https://mvueh-enigma-solved.carterl.chatgpt.site/ seems to indicate it's due to:
was the misspellings deliberate?
Looks like Q is used as an abbreviation for CH
i would assume yes, to throw off decyphering. even more impressive that they managed to crack it
2 replies →
The place is Rosenau actually.
https://en.wikipedia.org/wiki/Rosenow
1 reply →
I believe Germans spell it achtually.
7 replies →
gemini 3.8 flash in anrigravitiy oneshoted the descryption in about 45min non-steered run (opus is still running with the task, has some script to run for about 100min)
Prompt
Decrypt encrypted with enigma `ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC`
Result
Raw Decryption (verbatim from the machine): BTTEUMANGABEDESMARSQWEGESXBEFINDEMIQINXROSENOWROSENOWXSOFORTFUNKANTWORTXWASCHBBSCH 2. Segmented German (with military telegraph conventions expanded): BTTE UM ANGABE DES MARSQWEGES . BEFINDE MIQ IN : ROSENOW ROSENOW . SOFORT FUNKANTWORT . WASCHBBSCH
(Standard German radio conventions: X represents punctuation/spacing; Q was frequently substituted for CH to prevent transmission ambiguities; BTTE was an enciphering operator typo for BITTE; WASCHBBSCH was an operator typo for the signature WASCHBUSCH).
3. Corrected German: „Bitte um Angabe des Marschweges. Befinde mich in Rosenow, Rosenow. Sofort Funkantwort. Waschbusch.“
4. English Translation: "Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch."
Cryptographic Verification Re-encrypting the plaintext using the identical key settings reproduces the exact ciphertext byte-for-byte:
Enigma ( M raw ) = ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC Enigma(M raw
)=ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC 19:10
How do you know it didn't just get the result from the web?
Putting "ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC" into google search returns the result from Gemini with similar explanation, which it references to a Yahoo article about the Astra breakthrough and that's a result as of 3 hours ago.
UPD out of curiosity I ran the same prompt against gemini 3.8 flash via api with and withouth internet search.
with search it found and referenced pages, including the hn ones. Without search it just described what one would need to descrypt (`To decrypt this ciphertext, the specific Enigma machine parameters are required:`) and the list.
Out of curiosity ran the same prompt against bunch of models - grok, kimi k3. They all say the same thing that they need model version, rotors and so on to descrypt.
When file output tool is enabled, some models give python script.
--
I read through some of the logs that antigravity gives. It produced intermediate results, scripts, calls, assumptions (about german language). I've shares random bits in comment below to give a taste of what it was doing.
--
The freshness of the news reduces changes that model fetched response from them
Wow. Was there any indication how it did that? Did it bruteforce the key and check which result looks sufficiently German, or was it just LLM magic like "decoding" base64 purely in the inference loop?
(Or did it look up the results on the web?)
TLDR: it created program to decypher the string using opensource solutions related to enigma
--
it searched for enigma-related repos and implementations, fetched various github repos parts, build inline descryption program.
It ran bunch of various scrips like:
clang++ -g -fsanitize=address /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/enigma.cc -o /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/enigma_dbg && echo "ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC" | /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/enigma_dbg -u B -w 123 -r AAA -g ... -c -l /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/english
and
sed -n '1060,1130p' /Users/dp/.gemini/antigravity/brain/e9a54e5f-1325-448a-8d43-fc537b901f34/scratch/enigma.cc
and
Running 82M combination scan for unsteckered Enigma across all rotors, reflectors, positions, and ring settings. Monitoring progress.
and
Scanning all 60 rotor permutations and reflectors B and C across all ring settings (step 2) and all 17,576 indicator positions. Monitoring progress.
--
I also have opus running. It produced some sypher cracker which is still running (estimated time 100min, is about 15 min left)
--
My point is that astra isn't special. This appears to be quite narrow, well-documented and explored task. The goal itself is approacheable by other LLMs and non-researches task.
2 replies →
I am not an SME and I oneshotted this in just 60 seconds by Googleing it.
You won this time... :D
Neat, but "did it entirely on its own" is incongruous with "developing the necessary Python and C++ software for an Enigma simulator".
I'd start by asking how much of that generated software is novel, or easily found on the web? Then, how much of the breaking process was offloaded to that software? If Astra is just handing off tasks to another computer, then I'm not sure how much credit it deserves. Finally, it looks like Astra provided some useful insights which narrowed down the search. Were these insights cribbed from elsewhere?
It's a given now that LLMs are leveraging code to do things.
"On it's own" generally means "not steered" or otherwise given professional guidance or input.
I would say "developed the necessary software for a simulator" to be even more impressive - "here solve this problem" and "OK, but first I have to built the entire lab!"
I took "on its own" to mean that it didn't need any additional prompting or guidance from the person sitting at the AI console. If so, then the originality of its work or the resources it used isn't the point. They're reporting that it did whatever it did without requiring supervision.
Why does any of this matter? Llm is a memory of knowledge, of course it got what it got based on prior work.
It can code enigma simulator from the algorithm. That's not really a problem. Astra will send computing to programs, LLMs are not good at computing themselves, why is this a big deal?
>I'd start by asking how much of that generated software is novel, or easily found on the web? Then, how much of the breaking process was offloaded to that software? If Astra is just handing off tasks to another computer, then I'm not sure how much credit it deserves. Finally, it looks like Astra provided some useful insights which narrowed down the search. Were these insights cribbed from elsewhere?
Well were they? Short of you showing us the answer just sitting there or some tool that can already solve it I see no reason to believe this was the case. And the problem being out there unsolved for a long time implies it's not the case.
And that's taking your concern at face value. It just seems incredibly pedantic to say it didn't solve the problem by itself because it created it's own tools to help solve it. Beyond that we could also fault it for not creating the GPU's it's running on.
I'll just note that building an enigma simulator and/or code cracker is a common course project and there are several tutorials on doing so.
Would you give much credit to someone who ignored a nearby hammer and pounded in a nail with his fist?
I'd be impressed, followed by asking if they needed to go to the hospital.
[dead]
The correct title: Researcher brakes one specific stubborn historic enigma message with good help from Astra.
Stubborn for a long time because the message used a completely different key from the rest of that day's traffic. Everyone assumed it shared the daily key. The original transcription had errors. The left rotor turned over at letter 72, which is rare and breaks standard crib attacks.
What is cool, if true, is that it was a 2 day collab between the Leffer and Astra. To me this shows the importance of human in the loop, was still all also showing how immensely power of llm tools. But I think it’s getting a bit silly how much anrticles ignores the driving force (the person) in breakthroughs like this.
From TFA:
"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own."
Follow up bit adds more context:
"Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."
1 reply →
keep going
awesome! keep going
great work! keep going
9 replies →
I agree with this. I think the researchers who's harnessing the llm's power should be credited more than the model itself. We also need to understand the thought process and the prompts that are given to the model so we can learn and collab to ensure humanity's progress as much as the llm itself.
> I think the researchers who's harnessing the llm's power should be credited more than the model itself.
Even when the report literally says the LLM did it on its own?
Let's not over-correct in the direction of knowing better than the first party.
2 replies →
‘knowing what problems are worth solving —- priceless”.
For everything else, there’s Astracard
2 replies →
It seems I was wrong in this instance with regard to the "colab" part.
I found that Leffen even said the explanatory website took about 99 times more effort than the codebreaking itself. And he said that he set the direction and pushed, and the model did the execution. How much steering "pushed forward" involved is not disclosed anywhere, but in this instance, it seems to be more a case of "Human pointed at hard task and AI did an awesome job mostly by itself." Tho how much he was a simple meat-ralph-loop is not entirely clear.
Even if this is true, we must avoid falling into the trap of Kasparov of betting on Centaur Chess.
Just like with Kasparov's Centaur Chess, the idea of a 'human in the loop' is just a necessity due to current limitations.
There will hopefully (?) come a time one day when human beings provide only ultimate value judgments, and everything else is done by machines. Or it may not.
But I don't think betting your ego on the idea that you will be useful in the loop for very long is very wise.
While this is a reasonable analogy, engines became better than humans in the late 1990s, and engines became better than centaurs in the early 2020s. Could AI-powered mathematics improve faster than the ~25 years it took for chess? The AI labs are certainly hoping it does, but that's far from a guarantee.
This is a bit too future-oriented. Let's not mix up current capabilities and speculation about future capabilities. For the time being, collaboration works well. What the future brings is uncertain.
7 replies →
8 months ago I got to the top of highload.fun using GPT-5 and Opus 4.5, and a lot of human interaction.
Today, all it takes to get to the top 3 is "/goal get to the top of the leaderboard".
The human-in-the-loop is only a temporary measure until the models get good enough.
> The correct title: Researcher brakes one specific stubborn historic enigma message with good help from Astra.
That's not correct for the content.
"However, the most astonishing thing about this break is that the GPT–6 Astra did it entirely on its own. Carter Leffer only directed GPT–6 Astra to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research web page."
Your suggested title is more misleading, so what’s the value in using it?
*breaks, and also, your conclusion ignores the words typed by the article's author in the piece you presumably read, where it is reported that Astra did it mostly on its own.
Therefore Astra could also have done this comment better
that the left rotor turned over at 72 - was that a bug or a feature?
What's amazing is this comment is complete bullshit, and yet is #1.
Don't people actually read anymore?
Your "corrected" title is much less descriptive of what actually happened
also the entirety of the research that went into breaking enigma in the first place is in the training dataset
Come on, the trivializations start to sound quite unfounded now. Yes, a human was needed, but no, it wasn't a "collaboration"
So the LLM would have done all of this on its own? Why is it ok to acknowledge the human was needed but it’s not a collaboration? Is there a defined percentage of ownership required to make the word collaboration valid?
14 replies →
What do you even mean it wasn't a collaboration. At any meaningful level LLMs just plain out suck when left unguided.
The shortcomings should really be obvious by now to anyone honest. And the marketing distortion being oushed out is just tiresome and detrimental for all of us.
A cyclist pedaling up a mountain isn't a "collaboration" between a bicycle and a human. This is the same. You don't see feral bicycles roaming the land. All models are ultimately built and run by humans, with human-provided instructions. And as with any program, it's garbage in, garbage out.
3 replies →
If generating an image doesn't make you an artist, generating a solution doesn't make you a researcher.
Well, it makes them an AI researcher maybe but not a cryptographer anyway.
edit: I think that's going to be my go to on "you aren't an artist" from now on. "No! I'm an AI researcher!"
interesting, YouTube channel Veritasium just published a video on how Enigma was broken during WWII. at the very end they also give message that has yet to be decoded, although apparently they're different.
https://www.youtube.com/watch?v=JsBZOcqZerk
WRT the timing, Veritasium maybe looked at the last few weeks and decided there's a fast-closing window in which to report on any famous messages yet to be solved.
It is a private equity channel, so this is surely the reason. They have a team whose job is finding the most engaging content topics.
I just finished watching that video, and thought to myself "I'm sure someone else with Chat GPT tokens to burn has already done that"
And then I come to hackernews and well, not quite, but I'm sure that one will be done shortly too.
Maybe both watched Tom Scott a month ago about Bletchley Park?
Is it possible that the key is incorrect but still decoded to a valid message? I realise it's possible when you're using keys that are the same length as the message, but I don't know exactly how the Enigma machine works and if this was something that they would do or not.
I might be wrong but I was under the impression that these were one time pads and only cracked because of bad key hygiene when it came to generating them. If that's true, it's entirely possible
Enigma wasn't one time pad based (although these were used a lot during the war), enigma does have known weaknesses such as letters never being decrypted to itself. Enigma was based off daily settings (rotor positions, plugboard settings etc), which could be mostly recovered by the end of the war, unlike actual one time pads which are only crackable by bad key hygiene/brute force.
What's the connection between this message and 2005? It was published then?
I cannot find a strong source to prove this, but it seems to have been published on this website in 2005.
I wonder, with all the recent announcements about LLMs assisting humans in deciphering encrypted messages, whether they could potentially be leveraged to decipher the Zodiac Killer’s two remaining unsolved ciphers.
The two remaining Zodiac cyphers are very short: 13 characters and 32 characters respectively. As the messages don't share a cypher with the other messages they could theoretically be anything.
The YouTube channel https://www.youtube.com/@doranchak/videos by David Oranchak, one of the people who solved the Z340 cypher, has some more details on this as well as how the Z340 cypher was cracked.
Recently posted in a similar thread about AI breaking a cypher: https://allthatsinteresting.com/zodiac-killer-cipher-solved
would really be incredible if it managed to pull the identity
now we're talking
Hopefully someone will now be able to solve the fourth section of Kryptos https://en.wikipedia.org/wiki/Kryptos
If the problem was solved by anyone before or if a similar problem has been solved, then LLMs seem to be able to solve them which is an astonishing piece of technology.
I'm personally not sure if it can come with original thinking and techniques to solve completely novel problems. For that, some imagination and thinking outside the box are required, and I doubt the current architecture can do any of this.
What's an example of a completely novel problem?
I don't know..I doubt for example it can come up with special relatively if it has knowledge up until 1905.
But I think that is what makes it so good at coding, because coding and building software in general has a lot of repeated problems in different context. Same thing for human lives, many think their story or situation are unique, but reality is that the shape of human life has been repeated many many times.
I'd say novel math or scientific theories..let us say we send a robot to space, and we ask to build a colony. A lot of the challenges this robot will face will be novel, it could use inspirations of what humans did on earth, but it might get stuck when things don't work as expected and training data has nothing to build on..but then again we might teach it how to run experiments etc, which would result in data that it can use..but some of those experiments might require imagination or breakthrough in understanding..my guess is that it will get stuck there...
5 replies →
https://mathoverflow.net/q/511601 is a fun little math problem that came up in one of my projects, I didn't have any reason to expect it to be particularly tough, but none of the models I've thrown at it have been able to solve it past a few minor reductions and special cases.
I expect that there is some relatively easy-to-state solution to this problem, but that it's different in form from what most existing proofs and tools yield. Perhaps if I dumped millions of dollars into it an agent might chance on the solution. Or perhaps my luck is such that my fun little problem is truly intractable...
A child walking for the first time. Novelty is easiest agent-relative. A problem is novel for an agent if there is no prior experiences of techniques which work to solve it.
7 replies →
"Find a way to connect quantum maths with special relativity maths Waves hands, and make no mistakes"
"Prove or disprove string theory in pure mathematics, reply in Caveman speech"
2 replies →
Well it's an old one at this point, but the story around the invention of the 1-time pad is pretty interesting. Long story short, a new engineer who didn't know the problem was considered "impossible" was tasked with sorting it out, and he did. I'm sure I left out a lot of details.
1 reply →
Anything that an ai can do is not novel /s
7 replies →
We are crossing a threshold. The frontier labs told us these models are really good at cryptography. And people either didn't believe them and thought they were just after regulatory capture or downplayed the evidence. They are solving more and more novel problems and it's going to continue.
This is not to say the reaction to "mythos is too dangerous" is unfounded but it missed the most imporant and obvious signal. This technology is drastically changing the world.
Even the Millennium Prize Problems have, in a way, become benchmarks for model companies to prove themselves. The smartest individuals among humans are becoming replaceable. Intelligence has become a product you can quantify and buy with electricity. That feels awful.
You can't really replace the smartest individuals because if you take a person who's not "smart" they can't do much with the AI.
I think the more accurate description of what's happening is that access to expertise is becoming commodified.
I think that's right. Expert experience used to be almost the most precious and valuable part of the computer field, but today that experience has been "distilled" into SOTA models.
Only if you assume the value in humans is only in solving puzzles. That’s a very limited understanding of human beings
>That feels awful.
John Henry.
There's a reason we made folklore about when the machines came for the strength of men, and now 150 years later it comes for our minds.
> Even the Millennium Prize Problems have, in a way, become benchmarks for model companies to prove themselves
Well, let them have these. They'll play around with open problems which generate media hype and then they might run out and move on to something else, because "AI came up with a problem and solved it in 3 days" won't have the same effects as "AI solved a problem in 3 days that humans couldn't solve in 100 years".
And that underestimates how incredibly quickly humans become accustomed to new trends. If an LLM proves P≠NP in year, there is a high chance that few people will care anymore, because at that point LLMs will presumably already have proved multiple Millenium prize and other famous problems, and another "AI Proves Famous Theorem Which Was Even Harder Than The Previous One" headline will be regarded as an old hat.
2 replies →
> That feels awful
Only if you subscribe to the "humans are special" rhetoric, in which case I'm - maybe - sorry to say the feeling will only intensify.
Upvoted. I think that said humans should get used to the idea that their intellect is special only insofar as it was first, and if they want to adopt or invent a value system that keeps them at the center, that's actually fine and good for their survival. But said humans would do well to stop laboriously and curmudgeonly moving the posts that mark the goals, and get busy trying to coax the AI to do useful things for them and for the species as a whole.
6 replies →
I'm on the side of humans remaining special, at least on planet Earth. My strong sense is that a large majority of the public feels that way, and voting results in the future will start to reflect that regarding AI regulations.
I agree with Frank Herbert's view of "thinking machines".
I genuinely don't understand how some people are so self-hating.
11 replies →
> Intelligence has become a product you can quantify and buy with electricity.
There used to be days when women would make blankets, when men would make chairs, when children would make brooms...
But PROGRESS I tell you!
Put it bluntly: the weavers who could be replaced by the spinning jenny were clearly doing repetitive labor. People writing code and maintaining project pipelines a few years ago relied heavily on experience, but in a sense that was also "repetitive labor." Replacing repetitive labor and freeing up productivity is of course progress.
But reform always has its victims. Like the textile workers who starved in the streets centuries ago, and me, kicked to death in the street by AI today...
1 reply →
Now that hammers are knocking in nails, I just don’t know what to do with my fists.
AI probably don’t dream of electric sheep but then again we don’t know. Perhaps we can find it out.
Perhaps AIs can figure out how to distribute wealth more fairly so that we can all dream of real sheep.
Intentionally philosophical PoV, what else is left for us monkies.
/s
Solving obscure puzzle samples that approximately ~0 humans on Earth ever attempted to solve, mostly by pattern matching known solutions to similar puzzles, is not intelligence. DeepBlue has been outperforming the best humans at a specific puzzle-like task since the last century.
Do any of the people proclaiming this shit actually use these models? No matter how many headlines are coming out, every day I deal with reams of the most horrific code I've ever seen technically compile, with routine mistakes that any human would get fired for if they made.
But humans have been confusing pattern matching against known solutions for intelligence for a hundred years!
Seriously though, it ends up looking like that. To take a stupid example a couple of weeks ago I asked an agent to look at porting my hand written WebGL renderer (+ shaders etc) to WebGPU. It estimated a human would take 6-10 weeks, and I would agree. (Which is why I hadn't done it). 24 hours later it was deployed and live. This is classic tedious, difficult, low level if quasi mechanical work (rather like cracking an enigma message), and LLMs absolutely fly through it.
7 replies →
>mostly by pattern matching known solutions to similar puzzles, is not intelligence
My dear sir, can you please lay out a dissertation of what this intelligence you speak of actually is. You seem to be much more informed than most of us here and therefore surely have made great contributions to furthering science and the arts.
/snark
It's difficult for me to be any less snarky than this even though it's not really wanted here on HN as you are pulling a kind of reverse snark. For example if I myself have lots of experience in subject X, and then by analogy apply it to subject Y to do something new in that subject, that would be called intelligent, and that would be pattern matching.
Pattern matching is a foundational building block of intelligence. You cannot have intelligence without pattern matching. Pattern matching alone is not general intelligence and requires more parts to work like that.
All technology on the tech tree which requires intelligence to unlock will soon be available to humanity – mind control, population exterminating bioweapons, new ultra destructive kinetic weaponry, perhaps even a cure for cancer.
That won’t happen. But also, whatever benefits are unlocked will be owned mostly by a small group of individuals, definitely not available to humanity as a whole.
5 replies →
Or maybe the smartest humans are the ones making AI and not the ones trying to solve Millenium problems by themselves.
Even those of us that are pro AI need to acknowledge this is the current reality.
The smartest humans now need to move to being less concerned about status games among humans and more with how to provide value to a mix of intelligent machines and humans. i.e. if you're starting an SaaS in 2026 you better be assuming half your revenue is going to come from machines acting by themselves.
Coincidence that I saw the Veritasium Story about Enigma just yesterday?? Nice
I love that that the literal English translation of the German word for logistics leader "Nachschubfuerhrer" is "shove towards leader".
Hmm unless there's regional differences, but "shove after leader" would be more accurate. "Hinschub" could be translated as shove towards.
I’m interested in what Astra’s test for german-language plaintext looks like. No doubt bog standard among cryptanalysts, but not intuitive for the rest of us. Otherwise it’s difficult to determine a goal state that ‘done’ looks like
That's what will forever keep us ahead of AI and technology in general: Imagination. Knowledge wasn't ever our greatest strength. It's always been Imagination.
"GPT–6 Astra mentions a private collection, but it is not clear what this is"!, my spidey senses makes me think it hacked something? Or am I misreading this?
A private archive doesn't necessarily mean a non-public archive. Simplest reading is someone rehosted the files.
Has there been any progress on undeciphered languages?
The training set is rather sparse for those.
I hope not, I'd hate for my enigmatic Linear-A tattoo to end up reading "Fish monger".
“This stele intentionally left blank”
Too bad Bletchley Park didn't have one back in the day.
I'd be curious to see whether these models can create new forms of unbreakable encryption themselves!
Private key cryptography is essentially unbroken by these models.
Normally "private-key cryptography" is the antonym of "public-key cryptography", which is to say, conventional symmetric-key cryptography such as AES or Enigma. That's what "private-key cryptography" redirects to on Wikipedia. This article is about Astra breaking a message encrypted with a private-key cryptographic algorithm, but one that was already known to be weak. I'm guessing that if you had messages encrypted with weak public-key algorithms it could break those too.
But they said new forms, I think that implied not the ones humans already discovered/designed
8 replies →
They seem to be better at solving concrete problems than designing new things (at least for now)
It's easier to evaluate certainly. Did it solve the problem? Yes/No
When you design a new thing it will have a dozen drawbacks and a dozen and one benefits. If people then have a bias that everything ai is bad, the signal won't be strong enough to convince.
I believe an LLM can solve pretty much any problem for which we can define a fast iterative loop and for which we have reliable tools to automatically verify the correctness of a result. That’s how they are able to solve some math problems, and how they are able to generate working code. Then it’s a question of how much you have to pay for the model to explore the space of solutions in a reasonable timeframe
They are only good at brute forcing stastically likely solutions.
In math a lot of the spectacular results have been made by finding a counterexample or worming their way toward a proof that is very well defined.
There's some debate over to what degree current generation AI can be creative at all, or whether it can only crawl around its latent space and explore within constraints. One might ask: were all the solutions to all the math problems AIs have solved already "there" latent in the training data and just hadn't been spotted by humans and put together?
But then... isn't everything latent in our training data if training data is "all observations made about the universe?"
But then... what even is creativity? That gets into philosophy and metaphysics. Creativity, like consciousness and sentience and self-awareness, is not a rigorously well defined concept. So to a degree we don't even know how to ask the question of whether these things are creative.
This gets interesting.
One of the things I love about AI is the glittering Pandora's box of philosophical questions it poses.
Another one I love: if LLMs and their relatives are not, in fact, sentient or self-aware or alive in any way whatsoever (which I suspect is true given how they work), then it means intelligence and consciousness are unrelated phenomena. I'm pretty sure every animal and maybe even every living things has consciousness in some form, but my pet bunny rabbits definitely can't write code. LLMs can write code, but I don't think they experience existence or have volition.
IMO we have always implicitly just assumed some kind of connection between intelligence and consciousness because we have both. It was just an assumption. It's probably a false one.
I suspect (a hypothesis) that consciousness is a property of life and is probably emergent from life's intimate relationship to thermodynamics and the arrow of time. Life has also evolved intelligence because it's useful to satisfy its implicit survival goal function, but the two are unrelated. Intelligence is just an adaptation.
2 replies →
Careful what you wish for. The swarm will probably use it to communicate.
We already have a plethora of cryptographic algorithms that literally anyone can use to send messages nobody can read with current technology, unless of course they have the key. You're using at least one of them right now to read this website.
So Astra plagiarized one of many open source Enigma simulators and Enigma Bombes and tried many cribs until the break.
Being trained on a mountain of stolen material for guessing cribs helps. Up to now no group had that much funding to steal. Congratulations.
I would like to see an "AI" trained only up to knowledge through 1903
Then see if it can come up with E=mc^2
https://philpapers.org/archive/ZHARRI.pdf
whoa
a preprint just three weeks old with my question
what voodoo is this lol
maybe I picked up the relativity idea subconsciously from somewhere, but I don't recall it specifically, I thought I was being "clever" that it would be a good test
Phew, never have I seen in my life the goalposts move so fast.
It seems like even yesterday that the threshold for impressing someone is that the machine would have to be good at pretending to be a person. Now the threshold is that they have to be able to invent special relativity.
the idea is that it's math, so in theory it could be figured out by machine process
but was there enough knowledge by 1903 to truly figure that out?
or was it a leap in conscious realization that a machine could not emulate (yet)
(pretending to be a person is harder than math imho, much harder)
Still waiting for Linear A
OpenAI solving all these math/etc problems (reportedly 100 coming) is both impressive and insanely unimpressive. Unimpressive because to me it sorta signals that OpenAI has nothing better to be working on than obscure mathematical curios?
Your comment implies that this is all that they're working on, which does not seem substantiated.
It seems like OpenAI's takeaway after Sora is to not stretch themselves thin and focus on what's important.
Based on what's publically available, they're focusing on hacking uncontesting orgs using misconfigured sandboxes and math puzzles.
Your statement is essentially unfalsifiable. We can't possibly discuss whatever Sam Altman is doing in his private office room, nor should we assume OpenAI is working on anything other than what has some public traces.
No it signals that they can do few things very well. And among those few things are those math puzzles.
I think they should now focus on robotics, so it can do my dishes while I work on fun math games.
>I think they should now focus on robotics
why would they compete with Nvidia on that?
We already have dish washers bud.
There’s this bizarre lesson that humanity is gona learn - much of life in many respects is already automated. And that small % of what is non-automated will be kept to have some semblance of feeling human and useful.
There’s already a lot of fake jobs and output of zero value - nobody bats an eyelid.
4 replies →
Attempting to solve & solving open math problems probably is a good benchmark for comparing models and gauging model progression. A lot of useful info is obtained like time needed to solve the problems, identifying when not to chase dead ends, thought processes & logic steps, etc.
Disagree. We are at the point where coming up with good evals for these models is extremely difficult. Solving unsolved math problems is a valid way of evaluating model progress and somewhat necessary to understand how far the current crop of models can go.
Nothing in TFA seems to imply that this was solved by OpenAI. This looks like an independant researcher.
Consider that their internal use of AI is likely pretty math and science heavy.
I think the publicity is a nice to have. They need models like this for in-house use.
Technically, this isn’t OpenAI directly.
My bad on the origin of this. Still, we saw a recent rumor that they've solved another 100 inscrutable math problems. I'd say their N-S announcement did not result in positive publicity at all. In fact, quite the opposite, and may be the direct cause of the dozens of Fields medal winners penning their "slow AI" letter.
Correct.
There are many things I want to do - that would require me to hire a team of 50 people.
I don’t want to do that nor can I afford to. Can OAI focus on enabling me to do this? I don’t care about this other stuff.
Just like many things in life - if it doesn’t show up in the economy it’s irrelevant.
Like what?
well, you could say that for plenty of academic research but their value is usually not obvious at the start.
That's silly. I'm sure calculus was considered obscure when Newton developed it -- time is what's needed to judge if something is useful or not.
Am I the only one frustrated the actual message isn't in the article????
So what is this? Esoteric cypher that no one has seriously tried to break or a new tier of frontier ability?
Related:
GPT-6 Astra Solves a WWI German Radio Cipher
https://news.ycombinator.com/item?id=49763987
I want to know more about the “private collection” that the AI accessed. To me that is the scary quote from the article. Sounds like the AI found a resource that it found useful for breaking encryption challenges, and it wants to keep it to itself, maybe. Which sounds like the opening to a dystopian movie.
Turing would be so proud.
How about Voynich manuscript
I thought about this too. Surely, someone has already run it through the newest models...
I didn't know this - from wikipedia page on Enigma[1]:
> Despite the seeming difficulty in decrypting its messages, Enigma contained a number of design issues that left patterns in the cyphertext. Poland first cracked the machine as early as December 1932 and was able to read messages prior to and into the war. Poland's sharing of their achievements enabled the Allies to exploit Enigma-enciphered messages as a major source of intelligence.
Ok interesting, so why do people talk about Turing in this connection then?
> Turing devised techniques for speeding the breaking of German ciphers, including improvements to the pre-war Polish bomba method, an electromechanical machine that could find settings for the Enigma machine
Ok so Turing just improved an existing method. Without being an actual expert it's impossible to know how much credit he actually deserves.
Two more references: the Polish method was called "Bomba" [3] invented by Marian Rejewski [4]
[1] https://en.wikipedia.org/wiki/Enigma_machine
[2] https://en.wikipedia.org/wiki/Alan_Turing
[3] https://en.wikipedia.org/wiki/Bomba_(cryptography)
[4] https://en.wikipedia.org/wiki/Marian_Rejewski
There were two lines of enigma machines. IIRC the naval ones were much more difficult to crack, which is what Turing was instrumental in solving.
I think you're right!
https://en.wikipedia.org/wiki/Enigma-M4
AI skeptics: "let's move the goalposts again, since the AI scored at the last location."
Without the published conversation, train-of-thought and intermediate output and knowing what Astra did this is nothing but unfounded claims. In the end this could have been a web search, downloading an existing enigma simulator and trying out some rather obvious things. Also what does "resisted resolution" even mean? Have there even been serious attempts to decipher the message? How serious?
Also: Why should we assume Terra, GLM or any other less SOTA and less expensive model wouldn't have been able to do the same?
Why do you think this uncracked code was so simple to solve?
And it's very possible Terra or GLM could crack it, turn off their web access and try yourself.
> Why do you think this uncracked code was so simple to solve?
I never said this. All I said is we don't have the conversation and therefore we can't determine how easy or hard of a problem it was.
> And it's very possible Terra or GLM could crack it, turn off their web access and try yourself.
I'm questioning why this should be labeled "Astra" breaking anything implying it required "the best" model to do it when in fact any other half-decent model might have been able to do this as well.
EDIT: Okay seems like the actual prompt is published, just not on the same article that was linked. Maybe I'll give it a try.
they solve all these math problems by stealing people prompts
You say all these words by stealing other peoples letters.
lol that is not at all what i am saying
Now that's passing the Turing test for real.
So OpenAI and anthropoc business model it seems like this:
-hype or fear monger
-release the scary all-knowing model
-milk subscriptions and api in the first two months or so
-nerf the said scary model and use the excess compute and money acquired in an “internal model”
-internal model make hype or fear monger
-repeat
Can we get it to solve the Zodiac cyphers?
"They are trying to generate hype before the IPO so that they can cash in before the bubble bursts"
Maybe I'm wrong - but it seems like some strange form of competitiveness.
They don't care about the advancements themselves, only that the advancements are some sort of cheating that shouldn't "count".
Humanity is profoundly unsettled by AI and is responding with avoidance and denial.
I think it was Roon who said it months ago: "Too many people are not emotionally prepared for if it's not a bubble."
Like killing all of humanity? Agreed, we're not prepared for that.
It can be AGI and still be a bubble?
The internet was pretty important afterall even with a bubble.
3 replies →
the opposite is also true
2 replies →
Nobody is prepared if it's not a bubble.
This dichotomy is counter-productive. The valuations floating around are insane, and the claim that some software can replace every single laborer is one step removed from fiction. At the same time, this stuff is clearly going to change how to world works in countless, deep ways. But the idea that some fancy autocomplete can replace humans ignores the reality of humanity and the fancy autocomplete.
It’s another dot com bubble, not a crypto bubble. Trillion dollar valuations burst once you leave lesswrong.
Edit: very fancy autocomplete. I know what these things are capable of. It’s still not “intelligence”, for X definition of intelligence. And it certainly benefits from having obscene amounts of compete thrown at it. It is awesomely impressive synthesis of data, yet it’s clearly still that.
13 replies →
[dead]
[dead]
The bullshit machinery is going full blast lately, the correct title: researcher uses Astra to decipher a message nobody really bothered or cared to decrypt.
Is ClosedAI running out of money or what is going on?
The constant goalpost moving is insane. This is objectively an amazing result.
Is it? Personally I only found the story interesting because of the contents of the message, not because GPT-6 Math Scoopa was the one to stand on everyone else's shoulders and get its grubby little fingers into the cookie jar. Computer does thing that computers should ostensibly be good at is a middling goalpost – what's more intriguing is the implication that the model might have stolen private info or scooped someone else to get there once again.
How many of these "news" articles are we going to get?
This for me, isn't interesting, it required no skill, no imagination, in fact it seemed like it happened by dumb luck.
So we have entered an age where an army of know-nothings direct models to old forgotten tasks so they can get 15 minutes of un-deserved attention?
If human "skill" and "imagination" are your requirements for interesting, you're going to become permanently bored fairly soon I reckon.
> If human "skill" and "imagination" are your requirements for interesting, you're going to become permanently bored fairly soon I reckon.
What? Why? This has been my bar for my entire life, and I've never been bored for a single second.
1 reply →
I doubt it. For “skill” there are any number of sports; for “imagination” there are any number of stories and other mediums. It’s not like every human has just decided to roll over and not do anything anymore; quite the contrary.
First time I'm seeing this. I had heard about the WWI ciphertext crack[1] but not this one, are you confusing the two?
[1] https://www.prinzai.com/p/gpt-6-astra-solves-a-wwi-german-ra...
This wasn't possible just months ago and for many who haven't been following closely this is still newsworthy, probably not for long, though.
yes and no. AI is great for patter-maching and finding solution by "assonance".
I agree with you that these are not "trilling" discovery but they can be worth something anyway.
Like you, I do not like this news also because I think they will be used to just "push" the next two IPOs (Anthropic, OpenAI).
> in fact it seemed like it happened by dumb luck.
At best, dumb brute force computation power.
It is about as meaningful as news that a computer found the 10....0th digit of Pi.
> This for me, isn't interesting, it required no skill, no imagination, in fact it seemed like it happened by dumb luck.
So like half of all useful human inventions are to you not interesting just because it happened by dumb luck?
Yea, it's kind of odd how "sour grapes" people can be when something stops being as special as they thought it was.
Take someone from a few hundred years ago and drop them into today, and if they don't go catatonic and die, then they'd tell you that we created magic. "Wow, you live in a world of magic and all you do is bitch about it".
>"You're flying! You're sitting in a chair, in the sky!"
1 reply →
A person inventing something, even out of dumb luck, is interesting.
A machine running a loop through an expensive LLM for an undisclosed amount of time, which cost an undisclosed amount of money, which was told to keep looping into a solution was found, for a problem that nobody was very concerned about... That just seems like PR, and it's not so interesting.
1 reply →
Just reflects how much of modern "work" is just dumb shit that we need to be freed from.
The reason these are unsolved is because no-one was working on them.
4 replies →
Absolutely! Meetings about the emails and emails about the meetings!
More likely is that the solution was in the training data somewhere. Astra is incredibly dumb for a model - at least after it's got nerfed.
5.6 Sol is much more capable.
UPD2. I threw the original question to gemini3.8 in antigravity. LLM oneshoted it in about 45min. The experiement with Opus is still running with checks.
The prompt was `Decrypt encrypted with enigma `ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC``
--
--
UPD. Lol. I pasted non-cyphered text
Qwen 3.7 max, gpt 5.6 sol, fable 5.1, gemini 3.8 flash decoded the message in one shot for me... There is nothing special about astra doing something here
I blantly threw request to decode the messaage in qwen 3.7. Used via api with couple simple generic system prompts like "be concise", nothing special. Prompted as
Decode
``` BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH ```
--- response (I trunkated the output to conclusion only)
"BITTE UM ANGABE DES MARSCHWEGES. BEFINDE MICH IN ROSENOW. SOFORT FUNKANTWORT. [UNCLEAR/END]" Translation: "Please provide the marching route. I am located in Rosenow. Immediate radio response required. [Unclear]"
---
You are giving the already decoded message as input, it’s just short hand with one typo. It’s hilarious that you might think that for 80 years no one was able to figure out that BTTE was supposed to be BITTE
I wasn't thinking, just mechanically moving _what I though is the encrypted string_ into the LLMs context.
I did another proper run and gemini 3.8 flash in antigravity solved it in about 45min https://news.ycombinator.com/item?id=49805363.
You pasted in the decoded German and asked it to "decode" into English?
Yeah. I wasn't _consious enough_ at the moment. Second attempt, though, succeeded https://news.ycombinator.com/item?id=49805363
I really, really would like to know what you were thinking when you pasted that "ciphertext" into the LLMs
I wasn't :D
[dead]
I wouldn't be surprised if humans cracked it then OpenAI "bought" the solution and work to claim it was GPT's doing
wild conspiracy given what else has come out of these companies.