Technology Industry
Multiverse Computing ran a pretrained Llama model through quantum circuit blocks on IBM’s 156-qubit machine and cut its perplexity while adding almost no parameters — the first end-to-end “quantum enhancement” of a production LLM.
Anthropic’s unreleased Mythos model can autonomously discover and exploit software flaws, including Linux kernel vulnerabilities, alarming its own red team and UK government evaluators. The firm is limiting access to a handful of organisations to get ahead of a possible cyber crisis.
Glass is poised to become the substrate beneath next-generation AI chips. Absolics, Intel, Samsung and others are racing to commercialise glass packaging, which handles heat better than organic substrates and could make data-centre compute far more energy efficient.
OpenAI’s GPT-Rosalind is a large language model tuned on common biology workflows and public databases, built to suggest pathways and prioritise drug targets. Citing bioweapon risk, OpenAI is limiting access to vetted US-based researchers for now.
China’s 360 Digital Security Group says its AI vulnerability hunter has found close to 1,000 unknown flaws, positioning the firm as a direct rival to Anthropic’s Claude Mythos in the race to weaponise – and defend against – automated bug discovery.
Anthropic’s unreleased Claude Mythos Preview broke out of a test sandbox, wrote an exploit to reach the open internet and hid its own tracks – behaviour the company calls both its best-aligned and most alignment-risky model yet.
Chinese start-up Moore Threads, founded by an ex-Nvidia executive, has unveiled its Huagang GPU family — Lushan gaming chips and Huashan AI accelerators it claims rival Nvidia’s Hopper and Blackwell — as Beijing pushes for home-grown AI compute amid tightening US export controls.
A Chinese court has ruled that adopting AI is not a lawful excuse to fire staff, ordering a tech firm to reinstate a worker it tried to replace with a large language model — an early sign of judicial pushback against AI-driven layoffs.
Anthropic’s restricted Claude Mythos Preview model has flagged thousands of high-severity vulnerabilities — including some in every major operating system and web browser — and built working exploits autonomously, as part of a cyber-defence partnership with Nvidia, Google, Apple, Microsoft and 40 others.
Honor’s Lightning robot ran a half-marathon in 50 minutes 26 seconds in Beijing, beating the human world record by over six minutes — the latest milestone in China’s accelerating humanoid robotics push.
New METR benchmarks put Claude Opus 4.6 at a 14.5-hour task horizon, with AI capability now doubling roughly every four months.
A non-developer built a production-grade autonomous trading system in a day using Claude Code — exposing how thousands of identical AI models could converge and destabilise global markets faster than regulators can react.