Field Notes

Agent week: Nvidia breakthrough, Poolside deal, new vision model, AWS case, trading bots

Agents feel more concrete: a big benchmark jump, real deployments in ops, new models and new risks.

The week in five points

  1. Nvidia reports 100 percent with AVO on ARC-AGI-3, a test track for general problem solving. The team calls it a frontier architecture for long-horizon agents. These are agents with multi-step planning. The finding comes from the vendor, an independent confirmation is missing.1

  2. Nvidia buys Poolside's Model Factory for 6 billion dollars and takes on 109 employees. It is software that builds and maintains models automatically. If you are a Poolside customer, sort out access, handovers, and SLAs early.2

  3. Deepseek presents V4-Flash-Vision-Exp, an experimental multimodal model. The vendor says it sits close to Opus 4.8 on agent benchmarks and is tuned for low latency. You can test it now for vision tools in agents.3

  4. Panasonic Avionics and AWS are building an agent solution that diagnoses IFEC in aircraft faster. That is in-flight entertainment and connectivity. The reported effect: from hours to minutes. The numbers come from an AWS case study, an external measurement is missing.4

  5. Binance now lets AI agents trade. Agent OS connects to ChatGPT, Claude Code, and Cursor. If you try this, set strict limits and separate API keys for each experiment.5

Sources

  1. NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents (developer.nvidia.com)
  2. Nvidia is acquiring Poolside's "Model Factory" and 109 employees for $6 billion (the-decoder.com)
  3. Deepseek releases experimental Flash vision model that rivals Opus 4.8 on agent benchmarks (the-decoder.com)
  4. Accelerating aircraft IFEC diagnostics with agentic AI on AWS (aws.amazon.com)
  5. Binance now lets AI agents trade, but keeping them in check is largely up to users (techcrunch.com)