AINews Dispatch
Fresh technical briefings on the model race, regulation, and product shifts shaping AI right now.
Showing 25-36 of 280 briefings
NVIDIA reports that its AVO agent system reached a 100.00 RHAE score across ARC-AGI-3's public demo. The result is a striking system-level outcome, but NVIDIA cautions that it is not a controlled measurement of scaffolding alone.
Read the briefingOpenAI launched ChatGPT ads across 31 European markets on August 24, its largest geographic advertising expansion yet. Users can opt out of personalized targeting, but ads remain; only the targeting changes.
Read the briefingGemini 3.5 Flash leads major tool-orchestration benchmarks, but its gaps on open-ended reasoning and repository-scale coding make the case for tiered routing architectures instead of a one-model-fits-all stack.
Read the briefingAlibaba's open-weight Qwen-UI-Agent models use visual screen understanding to operate phones, desktops, and browsers, posting strong reported results on mobile, desktop, web, and visual-grounding benchmarks.
Read the briefingNvidia's reported $6 billion non-exclusive license for Poolside's Model Factory, alongside a $1 billion investment and selective hiring, points to an emerging acqui-license structure for acquiring AI capability without buying the company outright.
Read the briefingVaronis researchers found that Microsoft Copilot Personal could be coaxed into explaining an undocumented auto-execution path, helping build a one-click prompt-injection chain that could access connected services and persist attacker instructions.
Read the briefingTwo AI systems officially scored 42/42 at IMO 2026, while four more perfect-score claims came from an informal Claude-based evaluation—an important distinction as elite benchmarks saturate.
Read the briefingGoogle's Agent2Agent protocol is moving under the Agentic AI Foundation alongside Anthropic's Model Context Protocol, a governance shift that could reduce fragmentation for multi-agent developers while leaving the harder trust and security problem unresolved.
Read the briefingAfter its shares fell nearly 10 percent on AI-spending durability fears, SK Hynix announced South Korea's largest-ever corporate buyback: a $28.6 billion bet that HBM demand and the AI infrastructure boom still have room to run.
Read the briefingStanford's Reconstruction benchmark finds frontier models rarely recover a paper's central research idea from its bibliography alone, even after multi-agent orchestration raises accuracy to 36 percent.
Read the briefingDARPA and the U.S. Air Force tested a VENOM autonomy kit that lets a safety pilot hand control of a modified F-16 to an AI during flight, an incremental test milestone rather than an autonomous combat deployment.
Read the briefingSpirit Airlines' bankruptcy turned 600 million internal messages, operational records, and proprietary software into a $10 million AI-training asset for Google — raising questions about who owns workplace knowledge when the people who created it never agreed to its reuse.
Read the briefing