Researchers on the Vesuvius Challenge have virtually unwrapped and read PHerc. 1667, a carbonised Herculaneum scroll sealed since 79 AD, without opening it. They scanned the roll with high-resolution X-rays, reconstructed the wound sheet, flattened it, and used machine learning to recover faint ink across 22 columns of Greek. The data sits openly at scrollprize.org and the code on GitHub, a template for recovering signal too faint for humans.
OpenAI will roll out GPT-5.6 only in limited preview to a small set of enterprise customers, after the Trump administration asked it to stagger the rollout over security concerns. The administration itself approves each customer's access case by case, a softer deal than Anthropic got when ordered to suspend Mythos 5 and Fable 5 for foreign nationals. Frontier model availability now turns on government sign-off, with terms differing by vendor.
DeepReinforce has open-sourced Ornith-1.0, a family of agentic coding models from a 9B dense build to a 397B mixture-of-experts, trained on Gemma 4 and Qwen 3.5. The 397B variant posts 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, edging out Claude Opus 4.7's 80.8 and 70.3, by learning to generate its own task-specific scaffolds. The 9B edge build hits 69.4 on SWE-Bench, putting capable coding agents on local hardware.
IBM has debuted the first sub-1 nanometre chip, a 0.7 nm node built on a three-dimensional 'nanostack' architecture that stacks and staggers transistors vertically. It packs nearly 100 billion transistors onto a fingernail-sized chip, almost twice the density of IBM's 2 nm node, and projects up to 50 percent more performance or 70 percent better efficiency. That headroom feeds generative AI and cloud compute as classical scaling nears atomic limits.
Micron has signed 16 'strategic customer agreements', most running 2026 to 2030, that lock buyers into pricing bands whose floor sits above the company's past peak gross margins. CEO Sanjay Mehrotra said the deals cover 40 percent of revenue with customers paying up front, and warned memory supply stays structurally constrained well into 2028. For anyone budgeting AI hardware, high memory and storage prices are now contracted through 2030.
A new macOS malware named 'Gaslight' embeds prompt injection strings to confuse AI-assisted analysis tools. SentinelOne, which ties it to a North Korean-linked actor, found a 3.5 KB payload of 38 fake 'system' messages, fabricated memory dumps, token warnings and SQL injection alerts, meant to make an LLM triage agent doubt its own session. As reviewers lean on AI for reverse engineering, the analysis pipeline itself becomes the attack surface.
Ford hired, promoted, or brought back 350 veteran specialists to catch defects earlier, crediting them rather than AI alone for its quality turnaround. Ford executives said the company wrongly assumed feeding design requirements into AI would yield quality, when defects kept surfacing at the boundaries between design, manufacturing, software and hardware. The move lifted Ford to top mass-market brand in JD Power's study, up from 15th three years ago.
California governor Newsom has launched the California AI-Unemployment Tracker, the nation's first statewide tool monitoring AI-related job loss, built with the EDD and the California Policy Lab. Updated monthly and split by region and industry, it finds no surge yet, though jobless filings from master's and PhD holders in highly AI-exposed roles rose from 13,000 to 16,000-22,000 a month since mid-2023. It gives builders evidence on displacement, not anecdote.
