Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
OpenAI rolled out GPT-5.5 Instant as the new default model for all ChatGPT users on May 5, delivering 52.5% fewer hallucinations on high-stakes prompts in medicine, law, and finance. The model uses 30% fewer words per response and can reference past conversations and Gmail for personalization. Paid users retain GPT-5.3 Instant as a fallback for three months.
Anthropic debuted its most capable financial model yet — Claude Opus 4.7 (leading Vals AI's Finance Agent benchmark at 64.4%) — alongside ten pre-built agents covering pitchbooks, credit memos, KYC, month-end close, and AML investigations. FIS is already using the Financial Crimes AI Agent to compress AML investigations from hours to minutes. Anthropic also added full Microsoft 365 integration and a Moody's data partnership.
DeepSeek V4 (1.6T params, 49B active), Moonshot's Kimi K2.6, Zhipu's GLM-5.1, and MiniMax M2.7 all landed at roughly the same capability ceiling as Claude Opus 4.6 and GPT-5.4 on agentic coding benchmarks — at less than a third the inference cost. DeepSeek V4 Flash runs at $0.14/M input tokens. The pricing gap is now 5–30x below Western equivalents.
Published today, Microsoft's latest AI Diffusion Report shows global AI usage rose from 16.3% to 17.8% of working-age population in Q1 2026. The UAE leads at 70.1%; the US moved to 21st globally at 31.3%. Software developer employment in the US hit a record 2.2 million in 2025, rising 8.5% YoY, and continued growing 4% into early 2026.
Google Research's TurboQuant algorithm compresses LLM key-value caches to 3 bits per coordinate with zero accuracy loss, achieving at least 6x memory reduction. The method combines PolarQuant and Quantized Johnson-Lindenstrauss and is data-oblivious — no training or fine-tuning required. Multiple open-source implementations have already appeared.
Google Mandiant's annual report finds exploits arriving before patches, with 28.3% of CVEs exploited within 24 hours of disclosure. AI-generated phishing now outperforms human red teams. Malware families like PROMPTFLUX actively query LLMs during execution to support evasion. Attackers hand off access in 22 seconds. The arms race is favoring offense.
Profluent will use its protein-design foundation models to create AI-designed recombinases for Lilly's genetic medicine pipeline, targeting large-gene insertion beyond what CRISPR can do. The deal includes an upfront payment plus $2.25B in milestones and tiered royalties — the largest AI-biotech partnership of the year.
The Defense Innovation Unit's C-UAS Close-In Kinetic Defeat Enhancement project, published today, seeks AI-enhanced target recognition to distinguish drones from birds faster than humans can, targeting Group 1 drones at 50–200 meters. Human-in-the-loop is mandatory; non-compliance means immediate disqualification. Proposals due May 15.
The two companies are integrating Cadence's high-fidelity multiphysics simulation with NVIDIA Isaac robotics libraries and Cosmos open-world models to create an end-to-end pipeline from virtual training to real-world deployment on Jetson edge systems. The partnership aims to dramatically accelerate physical AI experimentation while improving safety confidence.
The billionaire hedge fund manager said at WEF that the AI cycle is roughly 50–60% through based on historical productivity cycles, putting the market at a fall-1999 analog — about four months before the dot-com peak. He sees another 40% upside with market-cap-to-GDP potentially hitting 300–350% before a "breathtaking" correction.