DeepSeek V4 Flash 0731 tops cost-performance, scores 89% on ARC-AGI-1

Charles Frye of Modal walks through an inference dashboard showing a traffic spike on a museum placard generator, where longer first-token waits and slower subsequent tokens trace back to queueing. Adding replicas relieves the queue.

Jim Clark, principal software engineer at Docker, proposes MCP gateways as the control point for agent safety, since harnesses are simple loops and safety comes from what context and tools flow into them. He argues agents running long enough that you stop watching them should carry no credentials of their own.

AI Engineer workshop with Braintrust's Doug Guthrie walks through a Python support agent built on the OpenAI Agents SDK, using 968 sample support traces that expose order lookup failures and broken escalation calls.

PayPal senior staff engineer Sam Parsons walks through three methods for accepting payments initiated by AI agents, including checkout inside ChatGPT and Google AI Mode. The talk covers how the payment flow works when a customer never visits the merchant's site.

In a demo, Charlie Guo and Gabriel Chua show Codex turning a captured Slack conversation into a working macOS live translation app in four minutes and two seconds. The build illustrates how context, plugins, and supervision fit together in Codex workflows.

Asked about his legacy in AI, Altman said he'd like history to record that OpenAI "made sure that AI was a fundamental thing that every person got hugely empowered with." He said many in Silicon Valley "really believed only a small priesthood should get access to AI."

A developer spent a month using Claude to build a local, open-source Chrome extension that rolls back UI changes across 170+ sites, freezing designs at an earlier point in time.