Daily briefings

AI Briefing: Desktop Copilot, Claude training and private learning

Three recent announcements explain how AI products are moving into everyday work: desktop controls, deployment training and a different approach to private model training.

Product

Copilot adds desktop controls in public preview

GitHub has introduced computer use in Copilot CLI and the Copilot app for macOS and Windows. It can read app content and interact with controls, including in software without an API. The feature is a public preview. Copilot asks for approval before controlling an app, and organization settings can disable it.

Why it matters

Before trying a desktop workflow, check which apps it can control and what approval is required. Preview availability does not establish reliability for your task.

Source: GitHub · Source published:

Skills

Anthropic launches a nominated engineer residency

Anthropic announced a $100 million commitment to Claude Frontier Academy, aiming to train 10,000 engineers by the end of 2027. Its first program combines in-person instruction with a 12-week project at the participant’s organization. Participation is by organizational nomination; this is not an open course. The training target has not yet been achieved.

Why it matters

For teams adopting Claude, the announcement describes one route to deployment skills. Organizations need to confirm eligibility with their Anthropic account or partner manager.

Source: Anthropic · Source published:

Privacy

Google describes auditable server-side federated learning

Google Research described a federated-learning system that uploads encrypted training examples and processes them in trusted execution environments. Published access policies and remotely verifiable code help auditors check permitted workloads. Google says Gboard has adopted the system. The design still depends on current trusted-hardware limitations; it does not keep every training example on the device.

Why it matters

When checking an AI product’s privacy claims, ask what leaves the device, which workloads can access it and whether those controls can be independently inspected.

Source: Google Research · Source published: