AI Agent SDK Race | AI Field Notes #20

AnalysisPrivacy researcher Alexander Hanff documented on May 4 that Chrome 147 and recent stable builds silently fetch a 4GB weights.bin file (Google's Gemini Nano on-device LLM) into the OptGuideOnDeviceModel directory without a consent dialogue. The download triggers when AI features are on, which is the default. Hanff's filing argues this breaches Article 5(3) of the ePrivacy Directive (the EU rule that forbids storing files on user devices without informed consent) and GDPR data-minimisation. He estimates the climate cost between 6,000 and 60,000 tonnes of CO2 across roughly one billion Chrome installs, using a 0.06 kWh/GB factor and the EEA 2024 grid emissions baseline.

AnalysisOracle launched OCI Enterprise AI on May 8 (its enterprise service running on Oracle Cloud Infrastructure), bringing Grok 4.3 (xAI's reasoning model with a one-million-token context window) a day after its public release. The same announcement made NVIDIA Nemotron 3 Nano Omni available, an open-weights multimodal model that reasons across video, audio, images, and text in a single system. The pitch is one stack for both closed and open work, on the same enterprise contract. The catch is operational. A model you fine-tune is one your platform team has to host, monitor, and patch.

AnalysisCloudflare announced on May 7 it will cut 1,100 jobs, about 20% of its 5,156-person workforce. CEO Matthew Prince told staff the roles being cut are not the roles the company needs for the future, and pointed to internal AI usage rising more than 600% in three months. The same week the company posted 34% year-over-year revenue growth and forecast 30% growth ahead. Shares dropped 24% on Friday after earnings. The pattern repeats: a healthy quarterly print, layoffs framed as AI transformation, support and middle-rung functions deleted while customer-facing teams stay.

AnalysisEU legislators agreed at 4:30 in the morning on May 7 on the AI Omnibus, a regulation that amends the AI Act before its August 2 deadline. The negotiation ran six months under heavy pressure to finish in time. The Omnibus simplifies registration for smaller systems, while several deadlines on high-risk-system rules moved by weeks in either direction. The shape that ships will look familiar to companies already prepping compliance. The wrinkle is that legal and engineering teams who locked their plans against the original calendar are now redrawing it.

AnalysisApple agreed this week to pay $250 million to iPhone users over Siri features that were announced but did not arrive. The settlement covers the personalised, on-device AI capabilities Apple promoted at WWDC 2024 (its annual developer conference) and which slipped through 2025 without a usable shipment. Per-claimant payouts will be small, but the framing matters more than the dollar size. A platform vendor demonstrating an AI capability now carries enforceable consumer expectations rather than vague forward-looking-statement cover. The signal travels. Any company running a hands-on demo of an AI feature has now seen the price tag for missing the ship date.