GPT-6 Astra Solves a WWI German Radio Cipher
A WWI-era German radio message recently decrypted with OpenAI’s GPT‑6 “Astra” prompts debate over how impressive the feat really is, given that the cipher type and key list were already known and the model mainly selected an existing key used on an unexpected date. Commenters argue over whether this counts as genuine cryptanalysis or just automating grunt work that no human had bothered to finish, while noting that such tools can still unlock many low‑priority historical problems. The thread widens into questions about AI capabilities, consciousness, and “goalpost moving,” as well as concerns about hype versus the practical but limited strengths of current large language models.
Nature of the cipher and solution
- Cipher is ADFGVX: a fractionating substitution via Polybius square followed by keyed columnar transposition.
- All daily keys are historically known; this message remained unsolved because:
- The operator apparently used a key from the “wrong” date (TRUPPENVERSCHIEBUNG, documented as starting later).
- There is likely a typo in the ciphertext (e.g., “S4STEN” probably meant “24STEN”).
- Astra reportedly applied a known key from a historical source and corrected an error, then cross‑checked the plaintext against ship logs of HMS Canterbury, which matched.
Brute force vs reasoning
- Disagreement on whether this counts as “brute forcing”:
- Some say iterating through a list of candidate keys is brute force.
- Others argue the keyspace was drastically narrowed by historical knowledge and common-sense reasoning, so the main work wasn’t raw search.
- Consensus that a dedicated human cryptanalyst with the same references could have done this but hadn’t, mostly due to effort vs reward.
Originality, training data, and “cheating”
- Some wonder if the model could have:
- Found a prior (unknown) human solution in its training data.
- Or effectively “fit” an invented key to historical events.
- Others consider this extremely unlikely given:
- The use of a documented key already known for other messages.
- The specific alignment with historical logs.
- One commenter notes Astra could not directly access or quote the cited book pages, raising questions about how it knew that exact key usage.
Significance and “nothingburger” debate
- Some see this as modest: applying a published key to a small search space, comparable to a straightforward dictionary-style attack.
- Others emphasize the practical value: AI makes previously uneconomical problems (tedious, low-priority historical ciphers) tractable with a prompt.
AI capability, AGI, and consciousness
- Ongoing meta‑discussion:
- Whether such feats indicate progress toward AGI or are just fast pattern-matching.
- Repeated “moving goalposts” as AI accomplishes tasks previously considered to require intelligence.
- Arguments that capability does not require consciousness; alignment and optimization risks are orthogonal to subjective experience.
- Ethical stance from some: in doubt, treat putative minds (including future AIs) cautiously to avoid harm.