The Case for AI That Improves Itself | Deep Dives with a16z

影片總結

Mirendil founders propose self-accelerating AI to cut frontier labs from 200 researchers to 1 and accelerate science.

On the a16z podcast, Mirendil cofounders Behnam Neyshabur and Harsh Mehta argue that AI should do more than automate work: it should accelerate AI research, engineering, and ultimately scientific discovery. Their near-term target is an AI system that improves models, kernels, frameworks such as PyTorch and JAX, and research workflows with minimal oversight. They say progress at Mirendil has required roughly 10 times fewer people and resources than comparable frontier-lab efforts, while newer models run longer and get stuck less often. The founders envision reducing the AI staffing burden for a frontier lab from 200 people to 10, 2, or even 1, then giving businesses their own AI, data, and infrastructure. They support sharply targeted safeguards rather than blanket restrictions, and define self-acceleration as a system of specialized agents and humans that continuously improves toward major goals such as solving Alzheimer’s disease and expanding scientific knowledge.

章節

1. 0:00Intro: Self-Improving AI Cuts Research Resources by 10x
2. 1:10Meet Mirendil: Building Self-Accelerating AI to Advance Science
3. 3:11The Shortest Path to Accelerating Science: AI Research and Engineering
4. 6:14What Self-Accelerating AI Means: From AlphaGo to Autonomous AI Research
5. 12:50From Anthropic to Mirendil: Five Years Toward Automated AI Research
6. 17:54Safety, Guardrails and Democratizing AI Research Access
7. 22:02Giving Businesses Their Own AI for Code, Infrastructure and Workflows
8. 26:34Scaling Systems of Agents and Humans: From 20 People to Recursive Improvement
9. 31:09Scaling Systems of Agents and Humans: 10x Company Size Yields Only 1.2x Productivity
10. 35:53Where Does It All End? Science as the Goal: Solving Alzheimer’s Disease