Google DeepMind Unveils Gemini 4 Argon, new flagship model

Meta AI researchers published six math papers with Muse Spark, closing five open problems in fields like group theory. They skipped custom scaffolding, using standard consumer chat with extended inference.

A structural flaw in the Model Context Protocol lets prompt injections hop across internal agents via implicit trust. Google and others admitted that attacking minor translation or data tools can trigger full network exfiltration.

OpenAI launched Dots at DevDay, persistent background agents powered by GPT-6 Astra that execute tasks indefinitely across 4,000 apps. Each agent operates in an isolated cloud desktop, requiring human sign-off before delivery.

A new reinforcement learning system dubbed Ataraxos crushed the top human Stratego player by a wide margin. It reached superhuman mastery in hidden-information games while burning orders of magnitude less training compute.

Apple is rolling out Siri AI as part of its upgraded Apple Intelligence suite. The assistant gains onscreen context, systemwide app actions, and local indexing across user photos, emails, and texts to handle personal requests.

Reflection unveiled Beam, an open-weight mixture-of-experts model boasting 501 billion total parameters with only 23 billion active. It rivals top frontier code models while slashing inference compute requirements by up to four times.

The Third Circuit ruled against Ross Intelligence, affirming that Thomson Reuters holds valid copyright over Westlaw headnotes. Training models on editorial legal summaries is not fair use, setting a huge federal precedent.

Runway introduced Praxis-1, an open-weight world action model translating third-person video into physical robotic policies. Training robots on general video libraries sidesteps the industry bottleneck of scarce physical data.

Unitree open-sourced UnifoLM-WLA-1.0, a 6-billion-parameter foundation model that unites locomotion and dexterous manipulation in a single architecture. Trained on 2,500 real robot hours, it coordinates both hands and feet.

Boston Dynamics and Hyundai opened the Robotics Metaplant Application Center in Georgia to train Atlas humanoids for car assembly. The robots are mastering parts sequencing and heavy transport before hitting live production.

Liquid AI launched d1, a vision-decision model that bypasses autoregressive token generation entirely. Evaluating inputs in a single forward pass takes under 300 ms, beating flagship models on ticket triage at a fraction of the cost.

California enacted Assembly Bill 1883, forbidding employers from using AI to infer worker emotions or monitor fatigue via neural tracking. The new restrictions force management to evaluate employee burnout without biometric spying.

China's Supreme People's Court published nationwide guidelines establishing fault-based liability as the primary rule for AI lawsuits. The opinions split copyright liabilities between developers and users based on provable negligence.

The UK accepted 44 recommendations from an NHS clinician-led commission, pivoting to continuous lifecycle oversight for clinical AI rather than one-off approvals. Regulators also opened the third phase of the AI Airlock sandbox.

Tesla slashed memory specs across its upcoming chip roadmap, cutting AI5 down to 72 GB and AI6 down to 144 GB. The company plans to power both autonomous driving and the Optimus humanoid robot on these leaner silicon budgets.