GPT-6 Cracks a WWI Cipher, While OpenAI Builds Chips with Its Own Models

A blogger named Prinz documented GPT-6 Astra — OpenAI's current flagship model — successfully decoding a World War One German radio cipher this week, the kind of operational military code that Allied signals intelligence teams spent days or weeks cracking in the field between 1914 and 1918. The author describes the model's approach as methodical rather than lucky, working through pattern recognition and substitution logic in a single session.

The Hacker News thread was appropriately cautious about a central unresolved question: whether this represents genuine deductive reasoning on a novel problem or whether similar cipher patterns existed somewhere in the model's training corpus. Cracking an unknown cipher variant is not something that can be memorized — it requires reasoning through structure — but the distinction between recall and reasoning in large language models remains difficult to verify from the outside.

The AI capability story with larger long-term implications may be the IEEE Spectrum report on OpenAI's custom silicon project. The chip, called Jalapeño, is a custom AI accelerator designed to reduce the company's dependency on Nvidia hardware, and the IEEE piece focuses on the design process: OpenAI reportedly used its own language models to help write and optimize the register-transfer level code — the hardware description layer that specifies how data flows between chip registers on each clock cycle — a domain historically requiring deep specialist expertise.

Hacker News commenters flagged the competitive dynamic this creates: companies with the most advanced models can use those models to design better hardware, which enables training of still-better models, producing a compounding loop that is difficult for anyone without frontier model access to enter. A counterpoint emerged in the form of Cactus Needle 3, a Show HN entry claiming that automation models in the eight-to-twenty-nine megabyte range can match DeepSeek V4 Flash on certain benchmarks — a comparison that, if it survives scrutiny, would suggest the efficiency frontier for narrow-task automation is compressing dramatically, with models small enough to run on embedded systems and microcontrollers where cloud API calls are impossible.

The week's most-discussed AI story by comment count was a piece from sockpuppet.org on how to write with a language model, which earned more than five hundred points and three hundred thirty-nine comments. The author's core argument is that writers extracting the most value from these tools maintain authorial control of structure and argument while using the model for fluency and iteration — rather than prompting for finished prose and accepting the output. Hundreds of comments treated the piece as a referendum on craft and identity, with programmers drawing parallels to how IDEs and code completion changed software development without eliminating the need for skill.