Mixture of Experts Anthropic’s Project Vend, computer science education and AI prompts in papers
Jul 4, 2025
Gabe Goodhart, Chief Architect at AI Open Innovation, Kush Varshney, an IBM Fellow for AI Governance, and Marina Danilevsky, Senior Research Scientist, dive into Anthropic’s Project Vend, where an AI tries to manage an office fridge. They discuss the future of AI in business and whether massive data centers are a thing of the past. The conversation shifts to the evolving landscape of computer science education, stressing the blend of technical and soft skills, and raises concerns over the ethical implications of AI in paper reviews.
AI Snips
Chapters
Transcript
Episode notes
Project Vend's Vending Machine Trial
- Anthropic's Project Vend trialed an LLM agent, Claudius, to run an office vending machine business.
- Claudius lost money, made inventory mistakes, and even hallucinated payment accounts, illustrating AI's current limitations.
Scaffolding Key for AI Agents
- LLMs alone aren't sufficient to run complex tasks; effective AI agents require structured scaffolding and programmatic checks.
- Combining LLM creativity with strict guardrails is key to successful AI agent deployment.
Distributed Training's Democratic Promise
- Distributed model training could democratize AI model creation by enabling everyday devices to contribute compute.
- This approach might foster community ownership and reshape AI economics away from centralized big labs.