AI

TL;DR: On 2026-03-31, AI coverage centered on agents/tooling (Claude Code, OpenCode, browser/host tooling), model releases (Cohere Transcribe, Google TimesFM), and the infrastructure bottlenecks powering demand (foundry capacity, hardware constraints).

Agents & tooling: leaks, limits, and safer execution

• Anthropic’s Claude Code faced a source code leak claim tied to an npm source map; related writeups also describe built-in anti-distillation/“undercover mode” concepts.
• Users reported Claude Code usage limits running faster than expected, with Anthropic investigating quota/token consumption behavior.
• Practical agent engineering themes appeared across projects:
  • Nango/OpenCode: building many API integrations with permissioning, post-checks, and traceability.
  • Pardus Browser: a Rust, Chromium-free, semantic web-page representation for agent consumption.
  • Coasts: containerized isolated dev environments for agents.
• A red-team style report (“Agents of Chaos”) described governance/security failures in real-world autonomous agents with persistent capabilities.

Models, systems, and infrastructure

• Cohere released Transcribe, an open-source ASR model (conformer-based) aimed at low word error rate and local/prod deployment.
• Google Research released TimesFM (time-series foundation model), including an updated checkpoint with 16k context and continuous quantile forecasting.
• Hardware/infrastructure signals tied to AI demand:
  • TSMC reportedly booked through 2028, potentially constraining leading-edge GPU/CPU supply.
  • Raspberry Pi profits rose on AI-driven demand.
  • Architecture-focused discussion highlighted KV cache costs and design approaches to reduce per-token memory.

Policy/market signals

• The U.S. President’s new science council was criticized as heavily industry/billionaire-led, with AI named among focus areas.
• Legal/finance risk analysis warned that AI data center buildouts could trigger litigation tied to capital-stack and GPU collateral complexities.