3 Before 9: March 9, 2026

Caitlin Kalinowski, who led hardware and robotics at OpenAI, announced her resignation on Saturday, directly citing the company's agreement to deploy its models on the Pentagon's classified cloud networks. In a post on X she said the decision to proceed without clearly defined guardrails around domestic surveillance and lethal autonomy was rushed. "These are too important for deals or announcements to be rushed," she wrote. Kalinowski joined OpenAI in late 2024 from Meta, where she had led development of the company's AR glasses. OpenAI confirmed the departure and reiterated its red lines: no domestic surveillance, no autonomous weapons. CEO Sam Altman had already acknowledged the deal announcement looked sloppy. The resignation is a visible crack in internal consensus at a company that moved fast to fill the gap left by Anthropic's Pentagon standoff.

A new Anthropic research paper released last week introduces a metric called "observed exposure" - the gap between what AI can theoretically do to a job and what it is actually doing today. Analysing around one million real Claude conversations against 800 occupations, researchers found that while large language models could theoretically handle 94% of tasks in computer and maths roles, actual observed coverage sits at around 33%. Unemployment rates for the most AI-exposed workers have not risen significantly since ChatGPT launched in 2022. The one clear early signal: job-finding rates for workers aged 22 to 25 entering high-exposure fields have dropped by approximately 14% compared to pre-AI baselines, suggesting companies are using AI to absorb entry-level workloads rather than hire.

DeepSeek V4, the Hangzhou startup's first major release since R1 launched in January 2025, has missed three consecutive release windows: mid-February, Lunar New Year, and the opening of China's National People's Congress in early March. Multiple credible sources including the Financial Times and Reuters confirmed the model exists and is in final testing. It is expected to be a trillion-parameter multimodal model with a 1 million token context window, natively processing text, images, and video. It was reportedly optimised for Huawei Ascend chips rather than Nvidia hardware, in a deliberate break from industry norms. Community consensus on developer forums now puts the release in the second week of March. DeepSeek has confirmed nothing.