AI News Briefing — June 20, 2026

Nobel laureate John Jumper, the AlphaFold lead, is leaving Google DeepMind for Anthropic after nearly nine years, following Noam Shazeer's departure to OpenAI and AlphaGo researcher David Silver leaving to found a company. The exodus marks a striking loss of top talent from DeepMind in a short window.

Miami startup Subquadratic claims its new SubQ model solves a long-standing mathematical bottleneck, running faster, cheaper, and more energy-efficiently while processing up to 12x more text at once. An independent evaluation reportedly lends some credibility to the bold claims.

OpenAI researchers report that reinforcement learning on small amounts of desired behavioral traits like truthfulness and corrigibility generalizes broadly, making models safer and harder to manipulate. Training on health data even improved deception detection, with gains on 44 of 53 benchmarks, in contrast to Anthropic's constitution-based approach.

US officials suggest ASML's most advanced chipmaking tool may have reached China, though ASML disputes this, citing the commercial risk to its export license. The dispute touches the core of semiconductor export controls shaping AI compute.

A podcast segment examines whether the US government's forced withdrawal of Anthropic's Fable 5 and Mythos 5 models is inadvertently boosting the brand. Security researchers signed an open letter calling the move dangerous, and Anthropic noted the same jailbreaks exist in rival models.

A new benchmark for realistic knowledge work shows even the best AI models fully solve just 3% of tasks, exposing significant gaps between hype and capability. The results challenge claims that agents are ready for complex professional work.

NVIDIA Research released SpatialClaw, a training-free agent that treats code as the action interface for spatial reasoning in vision-language models. It reaches 59.9% average accuracy across 20 benchmarks, outperforming the prior SpaceTools agent by 11.2 points without retraining.

Reliance is integrating AI across telecom services reaching more than 500 million users, per Mukesh Ambani's plans to embed AI in every call, app, and home. The move represents one of the largest-scale consumer AI deployments globally.

Sina Weibo researchers released VibeThinker-3B, a 3B-parameter dense reasoning model built on Qwen2.5-Coder-3B using a spectrum-to-signal post-training pipeline. Under an MIT open-source license, it reportedly matches far larger models on math, coding, and STEM tasks.

Google is appealing a Munich court ruling that held it directly liable for inaccurate AI search overviews after the system falsely linked two publishers to fraud schemes. Google characterizes the outputs as minor errors, but the court treated them as actionable harm.

Analysis argues that three decades of cyber-related export controls have failed to stop the spread of sensitive software, casting doubt on restrictions targeting Anthropic's Mythos cybersecurity model. The piece frames the Mythos controls as historically unlikely to be effective.

Barret Zoph, OpenAI's head of enterprise AI sales, has departed just five months after rejoining from Mira Murati's Thinking Machines Lab. His exit affects OpenAI's enterprise revenue push ahead of a planned IPO.