Just read an article from an #ai influencer musing on the Innovative power of AI.
(Edit 2: while linked articles a current / May 2025 they refer to an event in 2017 🤪 )

Tasked with breaking the Enigma code, an AI system trained to recognise German using Grimm’s fairytales, utilizing 2,000 virtual servers, cracked a coded message in 13 minutes.

Let's pause for a second to let it sink.

And let's think for a second

Alan Turing “Bombes” could decipher two messages every minute.

😱 Suddenly the AI result isn't all that impressive any more.
AI cuts out all the research, knowledge gain, and insight. With all the resources available today, it still performs worse than a solution from 70 years ago (to be precise 26 times).
And this is seen as an impressive innovation 🤡🤯

"Sources":
Influencer post https://mastodon.social/@Caramba1/114470245795906227
Guardian article
https://www.theguardian.com/science/2025/may/07/todays-ai-can-crack-second-world-war-enigma-code-in-short-order-experts-say

@realn2s even less impressive when you read about OctaPi running py-enigma on an 8 node cluster of raspberry pi's (32 cores) which can decrypt an Enigma message in just under 20 seconds.

@realn2s Exactly my thoughts when I read that headline. Hell, my watch can decrypt Enigma!

@realn2s I saw someone pointing out that AES was broken by AI, in the sense that there was a paper describing a model that could guess the topic of encrypted payloads.

It was AES in ECB mode 🤦

@realn2s Did the ai also have access to an enigma and “just” needed to find they key, or did it decode the message from scratch?

@Kierkegaanks
As far as i see it was more or less brute force
The AI was only used to determine of the result was valid German text

@realn2s are you confusing the time to crack the code with the time to decrypt a message? Shouldn't we compare months with 13 minutes?

@dsw
Yes and no. AFAIK the 13 minutes are really the cracking of a single encrypted message. They don't include developing the model, training it, setting up the cluster, ...

@realn2s given that ai can’t consistently multiply two numbers together, cracking anything non trivial would be quite the achievement.

@realn2s That's very dishonest reporting. The article is from 2017, when GPT-1 and BERT did not even exist. There were no LLMs as understood today back then.

And how long did Turing take to create the Bombe, with help from Polish cryptographers who reverse-engineered it 8 years before Turing broke the code?

https://www.techradar.com/news/we-watched-an-ai-crack-the-enigma-code-in-just-over-ten-minutes

https://en.wikipedia.org/wiki/Marian_Rejewski

@lnicola
GPT-1 and BERT are both from 2018. Precursors of them reach back into the 90th. There is no indication that LLM like technology was used or that GPT, BERT, or LLMs in general a suitable to solve the problem (apart from reusing existing solutions).

How long did it take to develop the AI?
What of the cracking method reused knowledge and understanding gained by Alan Turing or follow up analysis?

AFAI understand the AI was just used to "detect" German and therefor plausible cleartext. All the rest sound like brute force to me (based on Enigma emulations which again are based on Alan Turings work)

As state in edit 2 the whole story is getting more and more abstruse,

While the Guardian article is from 2025-05-07 and is titled "Today’s AI can crack second world war Enigma code ‘in short order’, experts say".

The event it is based on, happen somewhen in 2017.
See https://www.techradar.com/news/we-watched-an-ai-crack-the-enigma-code-in-just-over-ten-minutes

An seems to be a marketing stunt from a (now defunct) AI and Machine Learning startup Enigma Pattern and DigitalOcean.

According to https://www.digitalocean.com/blog/how-2000-droplets-broke-the-enigma-code-in-13-minutes they recreate the Nazi navy’s version of the machine in Python. Initially, they tried to teach their AI to decode the Enigma code itself, but it didn’t work. Neither did Lambda functions from Amazon.

As a next step they trained an algorithm to recognize German by using Grimms Fairy Tales (not AI mentioned here).

Then they ran the following on 2000 server:
Distribute combinations of "passwords" to run on the ciphertext. Check if the decoded message sounds like German, and if so, do a more detailed inspection by the AI.

The 2,000 virtual servers ran through 41 million combinations per second. Leading to roughly 40 billion combinations they tried.

So, the "AI" part was only used to check if the decoded message was plausible German text.
The message was “German is a beautiful language”. It would have been interesting to see what would have happened of the message had contained words which weren't in the training corpus (like "Panzer", "U-Boot", "Hitler").

So basically, it was brute force with a bit of language detection sprinkled in

Thanks to @magjo, @jaseg, @lnicola
and others for pointing that out

@realn2s Hi,
I used the Guardian article (linked) as a starting point for my thoughts.
In short: what once took years can now be done in minutes – not by a specialized system, but by a general-purpose AI.
In my article, I followed this development toward modern encryption and highlighted how quantum computing could render it obsolete in the near future.

At no point did I claim that the mere speed was the impressive part.

@Caramba1
My point is that the AI part on the whole story is minimal and not at all impressive.

Nevertheless, the main culprit here is the Guardian writing about a 2017 event with the clickbait title using "Today’s AI" and the weasel words "experts say".

I wouldn't even call the language detection "AI", it's a statistical method and doesn't require any continuous "learning".