OpenAI's GPT-6 Astra has successfully decrypted a German Army Enigma message (MVUEH) from July 10, 1941, that had resisted all decryption attempts since 2005. The AI independently developed an Enigma simulator and Bombe, used a crib based on the repeated place name ROSENOW, and corrected transcription errors in the original ciphertext. The breakthrough, verified by cryptanalyst Frode Weierud, demonstrates GPT-6 Astra's advanced autonomous research and cryptanalysis capabilities, though experts caution against overgeneralizing these results.

In a surprising experiment, developer Nathan Barry demonstrates that gzip, the ubiquitous compression tool, can function as a basic language model. By leveraging the compression-prediction equivalence and beam search, his pure-Python tool 'gzipt' generates Shakespeare-inspired text without neural networks or learned parameters. This highlights the thin boundary between compression and prediction, raising fundamental questions about what constitutes a language model.

Jared Palmer's Kev project delivers a family of small decision models (0.8B, 4B, 9B) built on Qwen3.5, replicating TypeSafe's Jev architecture. With Apache-2.0 licensing, public weights, and a TypeSafe-compatible API, Kev lets developers run fast, structured classification and scoring locally. Tests show strong accuracy on new sources, though calibration gaps remain, and Apple Silicon performance lags behind CUDA.

Mini-AGI is a new open-source byte-level language model that trains from scratch on an 8GB VRAM GPU using continual learning. It pages weights between disk and VRAM, assembles its own architecture via adaptive depth and dynamic expert routing, and retains knowledge without catastrophic forgetting. The project demonstrates that personalized, self-improving AI is achievable on consumer hardware, with a reported 99.84% retention against catastrophic forgetting.

Spain's Intellectual Property Commission ordered ISPs to block Archive.today and its mirrors (archive.ph, archive.li) without a court ruling, following a complaint from a rights holder. Users are redirected to a government warning page accusing them of facilitating piracy. The move raises serious concerns about due process, censorship, and the chilling effect on web archiving and research in the EU.

Independent AI researcher Nandakishor Mukkunnoth claims TypeSafe AI's Jev model replicates his non-autoregressive decision architecture, which he released a year prior. He introduces Laya, an open-source System 1 decision engine with sub-35ms latency, multilingual routing, and calibrated probabilities, benchmarking it favorably against Jev. The article covers technical details, performance comparisons, and honest limitations.

AI-generated posters often share a generic, clichéd aesthetic, but that's not inevitable. By specifying distinct design styles—from Bauhaus to Japanese minimalism—users can produce visually appealing, standout posters. This article explores practical prompting techniques, critiques from industry experts, and the importance of human oversight in AI-assisted design.

Stanford Medicine researchers have discovered that the human brain is not a single organ but two distinct organs that evolved independently over hundreds of millions of years. Published in Nature Neuroscience, the study reveals that the forebrain and hindbrain arise from completely different progenitor cells during embryonic development, overturning decades of biological consensus. This breakthrough explains why scientists struggled to grow hindbrain neurons in the lab and opens new avenues for studying and treating brain stem diseases like ALS and spinal muscular atrophy.

A practical guide to using LLMs for writing without losing your voice, plus how AI is reshaping malware creation, FPGA design, and more. Learn the two rules that prevent 'LLM-creep' and how to use models as copyeditors, not ghostwriters.