Today's Top Stories
May 3, 2026 brought a wave of AI developments spanning geopolitics, model releases, safety research, and enterprise adoption. Here are the stories that matter:
NIST: DeepSeek V4 Pro Lags US Models by 8 Months
The US government's most comprehensive evaluation of Chinese AI finds DeepSeek V4 Pro is the most capable Chinese model but trails US frontier systems. The gap is narrowing — from 14 months in 2024 to 8 months now — driven by efficiency innovations forced by chip export controls. Full analysis
Microsoft Caught Sneaking Copilot Attribution Into VS Code
Developers discovered Microsoft is inserting "Co-Authored-by Copilot" into Git commits even when the AI assistant is disabled. The behavior, introduced in VS Code v1.98, had no opt-in or notification. Microsoft issued a fix within 48 hours after community backlash. Full analysis
Xiaomi Releases Open-Weight MiMo-V2.5-Pro
Xiaomi's 405B parameter open-weight model can execute hours-long autonomous coding sessions, benchmarking competitively with Claude Opus 4.7 on SWE-bench. Released under Apache 2.0, it represents the growing trend of consumer electronics companies entering the AI model space. Full analysis
Frontier Models Diverge on Ethics
A Stanford study testing 12 frontier AI models on 500 ethical dilemmas found they agreed only 34% of the time — nearly half the agreement rate of human ethicists. The finding raises questions about the consistency and reliability of AI alignment training. Full analysis
MIT Explains Why Scaling Works
MIT CSAIL published a mathematical framework explaining why language model scaling yields reliable improvements. The paper shows scaling gains are a fundamental property of neural information compression, suggesting the "scaling wall" is much further away than skeptics predicted. Full analysis
Today's editor's pick: the Stanford ethics divergence study. While the NIST evaluation and Xiaomi release are important for the competitive landscape, the ethics study strikes at the heart of AI safety. If the most powerful AI systems can't agree on basic moral questions, we need to fundamentally rethink how we deploy them in high-stakes domains.