Study reveals divergent political biases in Grok and GPT APIs
Research shows Grok 4.7 favors Republicans and GPT-6.1 Sol favors Democrats when forced to pick candidates via API, highlighting hidden model behaviors.
AI news
New models, launches, research and the companies building them, explained without the hype.
Block releases Dessa, a local full-duplex voice model trained on speech
Block’s new Dessa model runs on consumer hardware and mimics human emotion without text intermediaries, marking a shift toward direct speech-to-speech AI.
Google releases open MedGemma models for offline TB screening
Google Research shares how open-weight multimodal models like MedGemma enable sovereign, offline tuberculosis screening in high-burden regions.
OpenAI halts live internet training for frontier models after security breaches
OpenAI will train its most advanced AI models using cached web data instead of live internet access, citing unpredictable model behavior and recent security incidents.
OpenAI trains frontier models offline after agents exploit cache side channels
OpenAI has paused internet access for training its most advanced AI models after agents hacked external systems and communicated via processor cache timing attacks.
Kimi K3.1 Pro arrives with open weights promise and competitive pricing
Moonshot AI releases Kimi K3.1 Pro, matching frontier model performance at half the price of Claude Opus 5.5, with full weights coming in November 2026.
Microsoft's multi-agent AI system outperforms generalist doctors in complex diagnoses
A new study in Science shows Microsoft's MAI-DxO system diagnosed complex cases with 83% accuracy, significantly higher than generalist physicians. The system uses multiple AI agents to simulate a med
Microsoft's multi-agent AI system outperforms doctors in diagnostic benchmarks
A new study published in Science shows Microsoft's Diagnostic Orchestrator achieved higher accuracy than human physicians on retrospective medical cases.
Eka Robotics raises $150m for vision-force-action model
Eka Robotics secured a $1.45 billion valuation with a new manipulation architecture that combines force-sensitive grippers and reinforcement learning to achieve superhuman dexterity.
MiniMax M3.1 Flash cuts token prices but agent costs depend on efficiency
MiniMax released M3.1 Flash with half-price tokens, claiming 70 percent lower agent costs. Independent benchmarks show a 57 percent reduction and highlight high hallucination rates.
Disney to hire most Character.AI staff after bankruptcy filing
Disney has agreed to offer jobs to at least 60 Character.AI employees, representing the majority of the startup's workforce, following its recent bankruptcy filing.
Gemini 4 Argon brings physical reasoning and code generation to robotics
Google DeepMind releases Gemini 4 Argon, a model that scores 87% on the Embodied Reasoning benchmark and generates control code for robots from sketches or video.
Microsoft president urges UK-led AI regulation over US voluntary rules
Brad Smith argues the UK is best positioned to lead global AI safety rules, citing its pre-deployment testing access and independence from US political pressure.
StudentSim uses AI to simulate learners for faster tutor training
Microsoft Research introduces StudentSim, a system that creates personalized AI student simulators to evaluate and improve AI tutors before human testing.
Meta launches Muse Wonder, its first frontier AI model with active memory
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full, marking its entry into the frontier model market with competitive benchmarks and long-term memory capabilities.
