AI avatars from Harvard, agent security, and local experiments

In today's roundup — AI avatars of Harvard instructors, a practical analysis of security testing for agentic systems, and several notable experiments with local models.

🔥 Hot:

🔹 Harvard launched a $699 bootcamp with AI avatars of its instructors — At HBS Foundry, the avatars provide participants with feedback on practice presentations and board meetings.

➡️ Useful materials:

🔹 An analysis shows why ordinary LLMs are bad at testing the security of AI agents — The author explains what a modern security harness should look like and why naive prompts no longer reflect real attack vectors.

➡️ Discussions and case studies:

🔹 An enthusiast expanded a home cluster to 36 DGX Sparks — The system now has 4.6 TB of unified memory and network infrastructure running at 200–400 Gbit/s. 🔹 An enthusiast is looking for a way to create TTS voices from recordings of their own courses — The idea is to preserve the individual vocal patterns of birds for generated videos.

📝 If you would like to add other news and materials to the list, write in the comments.