Five AI items worth your time

Five things from the last couple of weeks in AI that are worth your time, each with a source.

1. 1Anthropic on Scion. Merges went from about 2,500 a day to over 30,000, and git stopped fitting. Plan for the pipeline around your agents.
2. 2DeepSeek-V4.1-Flash-1017. Its card claims wins over Opus 5 on some coding benchmarks and losses on others. Test on your own tasks.
3. 3EmbeddingGemma 2. One open embedding model for text, images, video and audio. The text score barely moved, and code jumped.
4. 4Into the Unknown. Nine agents wrote 135 research papers and none met the scientists' bar. Check the evidence trail, not the polish.
5. 5Skip the Compressor. Coding agents resolved more issues with whole files than with compressed or retrieved snippets. Verbatim wins when exact text matters.

The thread running through all five: the model is rarely the bottleneck now. Context, process and checking are.

If you saw something this week that belongs on this list, tell me on LinkedIn.

Which of the five will change something you build?