Comment by tel | Hacker News Reader

Comment by tel | Hacker News Reader

Parent Full thread

I was curious if GPT could handle this. I couldn't get GPT-3.5 to figure it out, but GPT-4 was very quick to propose an interpretation. It's far more confident than I could imagine being, even if I could speak German, but this sort of thing feels like an application that a large language model might excel in.

---

Given that the message is decrypted from German, it seems to be a mix of letters and numbers that are likely military terms, geographical locations, and orders. Here's the translated meaning:

"Vier fünf null eins Y nach R. Stützpunkt Lübeck. Chef Vier. Flotte Mit Vier fünf null zwei und Vier fünf null drei zur Flender Werft Lübeck gehen. Dort folgt Weiteres."

In English, this roughly translates to:

"Four five zero one Y to R. Base Lübeck. Chief Four. Fleet with Four five zero two and Four five zero three go to Flender shipyard Lübeck. More to follow there."

This appears to be a set of instructions for a fleet of ships, possibly submarines (given the context of WWII and German naval operations), directing them to a particular location (the Flender shipyard in Lübeck) and notifying that further instructions will be provided upon arrival. "Four five zero one", "Four five zero two", and "Four five zero three" are likely codes for specific ships or units in the fleet. "Y nach R" could be coordinates or specific route instructions.

Just because it's confident, doesn't mean it's right. But it probably contains training data from this or similar sites.

https://enigma.hoerenberg.com/index.php?cat=The%20U534%20mes...

> Interpretation (preliminary):

> [An] U-4701, nachrichtlich [an] U-Stützpunkt Lübeck von Chef 4. U-Flottille: Mit U-4702 und U-4703 zur Flender Werft Lübeck gehen. Von dort folgt Weiteres.

> Translation (preliminary):

> [To] U-4701, for information [to] Submarine Base Lübeck from Chief of 4th Submarine Flotilla: With U-4702 and U-4703 go to Flender Dockyard at Lübeck. From there more follows.

There are about 2^67 different Enigma machine initial settings. The inverse probability of the appearance of a real seven letter German word (LUEBECK) twice in a random string of similar length to this message is a number that's pretty close to 2^67. So if you decrypted one ciphertext message with all the different incorrect settings you might expect to see one purported plaintext which isn't correct but which has two appearances of LUEBECK, or a similarly misleading occurence. Since there's also one correct plaintext, seeing LUEBECK twice already puts you at roughly 50/50 that it's the real message versus the most convincing wrong plaintext (if you had no prior knowledge of what the settings might be). The additional presence of even a few of the other recognizable German words (or common abbreviations such as triple letters and the shortened names for the numbers) makes it overwhelmingly likely that this is the correct plaintext. LUEBECK + LUEBECK + STUETZPUNKT in one message make the chance that it's not the real message of the order of winning a jackpot in state lottery two weeks running, even if the rest of the message was gibberish. In practice, much shorter pieces of plaintext than the double LUEBECK (like the presence of a single triple U, one spelled-out number, or highly abbreviated weather info) were used to validate guessed settings with a high degree of confidence.

That would be true if it actually did the decryption, but my point was, an LLM doesn't decrypt. It just has the encrypted string followed by the decrypted string in its training data and so it outputs something that's almost correct. (the numbers being wrong 4501, 4502, 4503 instead of 4701, 4702, 4703 - maybe some bugged training data, maybe hallucination).

Sure, and I said it was confident, not correct.

I find it interesting that it pulled some kind of interpretation from the string. Far more than I would have. I asked it to translate the English data back into a similarly plaintext string and then asked a second instance to decode it and it came back with a similar, slightly distorted response.

The point is more to say that a language model is exactly the sort of thing that would be used to determine whether a given potentially decoded plaintext string is actually decoded, and given various anachronisms and shorthands our personal language models may not be adequate.

But a giant one that's been fed all sorts of data including examples of text of similar usage sounds actually like it might be exactly the tool for this problem.

Seems mostly reasonable, except for consistently translating SIBEN to fünf (five) rather than sieben (seven).

Reply on news.ycombinator.com