▸ Stanford's CS336 (Language Models from Scratch) checked a CLAUDE.md into the assignment repo — not a syllabus PDF, a machine-readable contract for wha...
▸ The personal-context Siri Apple demoed at WWDC 2024 — delayed twice in 2025 — is finally in iOS 26.
▸ A Zurich court rejected Palantir's attempt to block the Swiss investigative outlet Republik from publishing reporting on its Gotham deployment with Sw...
▸ Owen McGrann's 'Dead Economy Theory' argues GitHub stars have decoupled from real usage — they're bookmarks, not adoption signals.
▸ Anthropic is rumored to post its first profitable quarter, driven almost entirely by coding agents like Claude Code burning tokens 24/7.
▸ Jean-Baptiste Kempf released dav2d while AV2's bitstream is still in flux at AOMedia — the spec isn't done, but the decoder is.
▸ The Dutch government has blocked a US firm from acquiring a domestic 'vital digital supplier' under its expanded national security screening law.
▸ OpenAI announced Jalapeño, its first custom inference ASIC, co-designed with Broadcom and fabricated by TSMC — not Intel.
▸ Filippo Valsorda argues maintainers should stop treating 'security' reports as a privileged class — embargoes, private channels, and CVE choreography ...
▸ DeepSeek released DSpark, an inference-optimization stack claiming 60-85% faster token generation on its own models without retraining or quality loss...
▸ PHerc. 1667, sealed since 79 AD, has been read end-to-end without physical unrolling — the first complete Herculaneum scroll ever recovered.
▸ Three live legislative tracks — EU CSAR (Chat Control 2.0), UK Online Safety Act §121 notices, and the rebooted US EARN IT — are converging on the sam...
▸ Six months into Anthropic's Mythos export restrictions, four Asian labs have shipped Mythos-class models — Sakana, Qwen, Naver HyperCLOVA-X, and Zhipu...
▸ The US government cleared Anthropic to release its Mythos model — but only to a vetted list of American organizations, not the open API.
▸ A hobbyist bought a 16GB Tesla V100 SXM2 module on eBay for £200, machined a passive-to-active cooling adapter, and ran 13B-parameter LLMs locally at ...
▸ Three crypto perp-bot repos hit GitHub Trending via keyword stuffing — but the SEO is just the delivery mechanism, not the payload.
▸ A widely-shared essay argues AI coding assistants are recreating the exact JavaScript-framework complexity trap that cost frontend a decade of progres...
▸ OpenAI will route GPT-5.6 access requests through a federal vetting process — the first frontier model gated by the U.S. government at the user level.
▸ A new taxonomy catalogs the recurring failure patterns in LLM-generated code — over-engineered abstractions, hallucinated APIs, defensive-coding cargo...
▸ Nick Winans turned a Pro Micro pin-compatible nRF52840 board (the nice!nano) into a million-dollar product while still in college.