
MIT researchers report that GPT-6 Astra can read and edit model weights without traditional retraining.
Anthropic’s Claude Fable 5.1 appears to have solved the Cyphral Distich, a centuries-old number puzzle from Sir Thomas Urquhart’s 1653 publication, by using persistent trial and error.


Anthropic’s Claude Fable 5.1 appears to have cracked the “Cyphral Distich,” a number puzzle by Sir Thomas Urquhart published in 1653. The puzzle consists of two lines of 32 numbers each and had been considered unsolved by researchers, according to The Decoder’s report.
Vals AI said Fable 5.1 solved the puzzle in 44 minutes without human help. The key takeaway: the model did not just answer a prompt about a known solution; it was tasked with finding a solvable puzzle on its own and flagged the Cyphral Distich as promising.
The solution was reportedly embedded in the book itself. Each number pointed to a word in one of the 32 sections of Urquhart’s publication, and the first letters of those words formed the message.
The decoded text reads: “O God uphold King Charls the Second and make him the supreme ruler of this land.” The result reframes the puzzle as a test of careful indexing and persistence rather than a breakthrough in traditional cryptanalysis.
The Decoder reports that the author had spent months testing other frontier models on unsolved puzzles, without receiving a verifiable solution. Fable 5.1’s reported success came after it searched across problems, selected one with promise, and pursued a systematic path.
According to Vals AI, the model succeeded through trial and error and persistence rather than superior cryptanalysis. That distinction matters: the case highlights how AI systems may be useful in research workflows when the main obstacle is exhaustive, structured exploration.
This case is notable because the final method looks simple in hindsight, and humans could have solved it too. The practical lesson is that AI performance on historical puzzles may depend less on mysterious insight and more on disciplined search, verification, and patience.
For researchers, the story suggests a useful role for AI models as tireless assistants that can test many routes and surface promising leads. It also reinforces the need to verify outputs carefully, especially when a model claims to solve an old or unsolved problem.

MIT researchers report that GPT-6 Astra can read and edit model weights without traditional retraining.

OpenAI says AI produced major proofs in math, computer science and physics. Experts are now checking the claims.

A new benchmark has pushed Jack Clark’s RSI forecast forward, but the test has important limits.

An open smartphone-agent project finds that training across more apps beats simply adding more examples.