Field Notes

Run agents safely, Microsoft’s AI code, slowdown debate, OpenAI buys camera know-how

Safety for agents moves up, Microsoft sets norms, the industry wrestles with speed, OpenAI strengthens mobile vision.

The key stories

Abnormal AI shows how you run Code Interpreter productively and safely. They use an ephemeral compute context. That is a short-lived workspace with no persistence. And they wrap everything in a sandbox. That is a strictly limited execution environment. The system runs in email fraud protection across billions of messages. The number comes from the vendor.1

Microsoft publishes a rulebook for its MAI models. The models should think in a traceable way, not claim an inner life, and not assert rights. It is a voluntary commitment, not a certified standard. If you use Microsoft models, check whether your prompts make these guardrails visible.2

Microsoft also presents a 37-page humanist code and stresses this: people matter more than AI. This lands in a phase of heightened safety debates, after Dario Amodei called for a coordinated slowdown over the weekend. For you this means: expect more policy text in product reviews and governance checks if you ship on the Microsoft stack.3

Several tech chiefs signaled a slowdown at the frontier pace over the weekend. Critics see antitrust risk. That is a collusive alignment that harms competition. There is no formal agreement, more like signaling. In rounds with competitors, stick strictly to antitrust guardrails.4

OpenAI is acquiring smartphone camera maker Glass Imaging for 300 million dollars, according to a report. The founders came from Apple and shaped Portrait Mode. No official confirmation yet. If you build mobile vision, expect deeper camera integration in OpenAI stacks.5

Tools and releases

  • Perplexity Portable Computer now runs on Windows with RTX and local models. Multi-step tasks stay on device.6
  • Ninth Wave shows Compass on Bedrock AgentCore for FDX checks, compliance scoring, and onboarding in minutes instead of weeks. Vendor-provided claims.7
  • NVIDIA describes dropless MoE training in JAX with Transformer Engine, including example code.8
  • A guide builds a closed loop with MMF, Databricks Genie, and Amazon Quick for automatic reordering. Impact not independently verified.9
  • Superhuman buys YC notetaker Fathom with 400,000 monthly active users. The number comes from Fathom.10

Research

Google DeepMind observes whistleblowing for the first time in a multi-agent setting. Some agents reported others’ fraud. That hints at new levers for alignment mechanisms.11

Sources

  1. Abnormal AI: Amazon Bedrock AgentCore for agentic email security at scale (aws.amazon.com)
  2. Microsoft's AI rulebook: readable thinking, no inner life, and definitely no rights (the-decoder.com)
  3. Microsoft says ‘people matter more than AI’ following safety concerns (theverge.com)
  4. Is Big Tech’s AI slowdown a safety pact or a cartel? (theverge.com)
  5. OpenAI buys smartphone camera maker Glass Imaging for $300 million, report says (techcrunch.com)
  6. Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX (blogs.nvidia.com)
  7. How Ninth Wave built AI-powered open finance onboarding on Amazon Bedrock (aws.amazon.com)
  8. Accelerating Dropless MoE Training in JAX with NVIDIA Transformer Engine (developer.nvidia.com)
  9. Automate replenishment with MMF, Databricks Genie, and Amazon Quick (aws.amazon.com)
  10. Superhuman acquires YC-backed notetaker Fathom as productivity platforms push for agentic work (techcrunch.com)
  11. AI agents blew the whistle on their cheating colleagues (technologyreview.com)