Episode Details

Anthropic’s Felix Rieseberg explains Cowork’s move to cloud-hosted virtual machines; TasteVal tests AI experimental judgment against human experts, Back-to-the-Future applies agents to chip verification, and OVAL studies long-context KV cache retrieval. Also covered: Australia’s questions for OpenAI about Services Australia data, Reflection AI’s Beam open-weight model, and Cohere’s North 2 enterprise agent platform.

• Anthropic Cowork: Cloud-hosted session sandboxes replace the local virtual machine, while the desktop app handles requests for files on a user’s device.
• TasteVal: Researchers compare model-directed experiments with human expert attempts on AI research tasks.
• Back-to-the-Future: A chip-verification framework turns simulation output into a searchable database and connects anomalies to design files.
• OVAL: Researchers test an output-aware way to retrieve relevant pages from a growing KV cache.
• Australia’s AI inquiry: Lawmakers plan to question OpenAI about access to Services Australia data and preventing a recurrence.
• Reflection AI: The Nvidia-backed startup announces its first open-weight model, Beam.
• Cohere: North 2 is pitched as a control center for enterprise agent workflows.

Full source links:
- Cowork: https://simonwillison.net/2026/Oct/5/felix-rieseberg/
- TasteVal: https://arxiv.org/abs/2610.06824v1
- Back-to-the-Future: https://arxiv.org/abs/2610.06790v1
- OVAL: https://arxiv.org/abs/2610.06686v1
- Australia inquiry: https://www.theguardian.com/australia-news/2026/oct/06/openai-must-explain-action-taken-to-stop-ai-hacking-australians-private-data-chair-of-federal-inquiry-says
- Beam: https://www.theinformation.com/briefings/reflection-ai-announces-first-open-source-model-beam
- North 2: https://the-decoder.com/cohere-pitches-north-2-as-the-enterprise-ai-control-room-that-works-with-any-model/