The Perils and Promise of Artificial Intelligence
summary
History of AI moves from Deep Blue (1997, ~11 gigaflops) and expert systems to neural nets that beat image challenges after Imagenet: 14 million labeled images produced 85% (2012) then ~99% accuracy, then voice recognition and AlphaGo (2016–17) beating champions. AlphaZero self-play outperforms prior systems. AlphaFold2 now scores in the high 90s on protein folding. GPT-3 is powerful but closed-source; failures include Microsoft’s Tay and deepfakes. Automated trucking could affect ~7 million long‑haul jobs by tripling utilization. MIT Dactyl hand trains in simulation successfully.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:42Defining Artificial Intelligence
- 5:15Expert Systems and History
- 9:31Neural Networks and Imagenet
- 13:00AI in Action
- 16:02Game Playing AIs
- 18:40AI for Biology
- 21:09General AI and Ethics
- 24:25Surveillance and Society
- 26:44Dystopian Potential
- 30:41Bias and Explainability
- 36:22Automation and Jobs
- 42:15The Future of AI Tools
related talks
50mKeynote: After the AI Hype – What’s Real, and What’s Next
51mBeyond the AI Hype: What's Real, What's Next
19mAI: The current state of play
31mCan AI cause the end of the human race
40mWhere Machine Intelligence Ends and Human Creativity Begins
30mIntroduction to AI Safety. Let's be ready before it's too late!
14mAI Is Built to Be Applied
14mWill AI replace All coders?