Notion Now Auto-Runs Custom Agents Triggered by AI Meeting Notes Completion

In its July 31, 2026 release notes, Notion added the ability to automatically launch user-configured "Custom Agents" as soon as "AI Meeting Notes" finishes a transcript. Until now, routine post-meeting tasks—updating statuses, posting recaps, filing tickets—required a human to step in manually; with this update, the completion of the notes can kick off a chain of downstream tasks.

The no-code agent chaining is what makes this practical—a design that fills a gap in the everyday workflow of "take the notes → hand off to the next action."

Even with High Benchmark Scores, VLMs "Silently Drop Clinical Terms"

A paper shared in the community argues that the evaluation metrics for VLMs (vision-language models) that generate chest X-ray reports are fundamentally broken. What was observed: boilerplate repetitive reports, reports containing no clinical terminology, and reports deemed "no abnormality" all manage to score highly on benchmark metrics. In particular, terms that are clinically meaningful but low-frequency get quietly stripped out by the models, resulting in a stream of near-identical reports with little clinical value.

The paper also presents a framework for actually measuring this terminology dropout and the inclusion of biased language. Good benchmark scores don't guarantee real-world safety—a fresh reminder of the risk of trusting metrics alone in the medical domain.

DeepSeek-V4-Flash 0731: Local Benchmark Results Roll In (27.2 t/s on Q3_K_XL)

Local performance measurements of the 0731 build of DeepSeek-V4-Flash are piling up in the community. One user ran Unsloth's "Q3_K_XL" quantization on a mixed-GPU setup—an RTX 6000 (96GB, CUDA) paired with a Radeon W7800 (48GB, ROCm)—and reported processing a complex physics simulation prompt of roughly 14,860 input tokens and 20,988 generated tokens at an average of 27.2 t/s. The user's verdict: "cheaper than K3 or GLM-5.2, but excellent."

Local operational data since the official release is gradually accumulating, and the relationship between quantization choices and real-world speed is starting to come into focus.

Video Editing × AI: Agents That Edit, Humans That Approve

A project was introduced that consolidates the content being edited into JSON form, letting an AI agent and a human work from a shared editing state. The AI handles proposing, applying, and verifying cuts and stylistic choices; the human reviews the diffs and gives final approval. The goal isn't full automation but a division of labor: "leave the editing to AI, and let humans judge the finished cut."

This design—delegating repetitive work to AI while keeping decision-making authority with humans—offers a realistic feel for what it takes to integrate generative AI into a production pipeline.

LLM Post-Training (SFT, DPO, GRPO) on a Consumer 8GB GPU, with a Minimal Setup

On Hacker News, attention has been drawn to a training repository that lets you try LLM post-training—SFT (supervised fine-tuning), DPO, and GRPO—on an 8GB GPU with a minimal setup. The configuration appears to let you run and compare the major post-training techniques locally, without large-scale compute resources. For anyone who wants to understand how post-training actually behaves by getting hands-on, it looks like an approachable introductory resource.