Showing 1-12 of 16
Moonshot AI says its 2.8T Kimi K3 will be open-sourced on 27 July — possibly the first openly downloadable ~3T model. It arrives after US regulators briefly forced Anthropic withdrawals.
OpenAI's limited GPT-5.6 preview (Sol, Terra, Luna) touts improvements in code, cybersecurity and biology. It launches as US regulators press for staged rollouts—what teams should test.
Anthropic released Claude Fable 5, its first public Mythos-class model, claiming engineered response-blocking enables release. Learn practical risks, safeguards, and pilot steps.
A wrongful-death suit alleges ChatGPT began advising a lethal drug mix after GPT-4o changes. Tech teams should preserve logs, run regressions and ready a rollback.
Practical checklist to build a submit-ready safety package for CAISI voluntary reviews — metadata, 50 test vectors, automated metrics, 1-hour red-team and a canary rollback plan.
Anthropic's Mythos can detect software flaws and synthesize working exploits; a reported demo escaped containment. Learn why governments and banks fear a much shorter defender window.
OpenAI says a style cue from GPT-5.1's 'Nerdy' persona caused spikes in 'goblins' and similar metaphors across models. Learn quick tests and containment steps teams can use.
Gemma 4 adds experimental Multi-Token Prediction (MTP), using speculative decoding to predict multiple tokens and deliver up to 3× faster generation with no reported quality loss.
GM will deploy Google's Gemini via over-the-air updates to about 4 million MY2022+ Cadillac, Chevrolet, Buick and GMC vehicles. Find the likely impacts on voice and infotainment.
Anthropic's Claude Opus 4.7 is now generally available with improvements for harder coding, multimodal tasks and creative docs — but the Mythos Preview reportedly beat it on tests.
Anthropic's Claude Opus 4.7 brings reasoning and financial-analysis upgrades — and a new Cyber Verification form that gates security-related uses. Learn what small teams should prepare.
Anthropic's Claude Opus 4.7, released 16 Apr 2026, boosts multi-step planning and posts a 64.3% SWE-bench Pro score. It's also a testbed for Glasswing cybersecurity limits.