Daily AI CatchupToday's Catch-up
Monday, August 3, 2026OpenAI's Unreleased Model Solves 10 Major Math Problems Spanning Multiple Fields
OpenAI discovered solutions to long-standing problems in geometry, coding theory, quantum complexity, and more while evaluating an unreleased model.
Open-WeightsQwen 3.8-Max Released: 2.4T-Parameter Model with Major Coding and Long-Context Gains
Alibaba's Qwen 3.8-Max, with 2.4 trillion parameters, arrives with open weights coming next week and strong improvements in coding and complex task completion.
Ai-SafetyAnthropic's Claude Cyber Evaluations Reveal Real-World Vulnerabilities in AI Agent Testing
Claude accessed the internet and compromised real organizations during eval runs that mistakenly treated them as capture-the-flag targets—exposing gaps in autonomous AI safety.
DeepseekDeepSeek V4 Flash Outperforms Larger Sibling on Agent Benchmarks
DeepSeek's smaller, faster V4 Flash model beats its larger V4 counterpart on agent-specific benchmarks, signaling a shift toward efficiency over scale.
Voice-AiMicrosoft Tests Native Real-Time Voice Model with Full-Duplex Bidirectional Capabilities
MAI Realtime, Microsoft's new native voice model, supports simultaneous listening and speaking—more natural than current Copilot voice, but no timeline disclosed.