Last Week in AI Yann LeCun on GPT-3, New Google Projects, Inequality, GPT-3 on Hacker News
Nov 5, 2020
The podcast dives into a range of captivating AI topics, starting with a sharp critique of GPT-3 from the ‘Godfather of AI.’ It explores fascinating collaborations aiming to harness AI for environmental monitoring and cancer care. A revealing discussion addresses the deepening inequality in AI research, spotlighting the dominance of elite institutions in publications. Finally, listeners hear about innovative strategies using GPT-3 to achieve viral content success, while raising concerns about the potential decline in online content quality.
AI Snips
Chapters
Transcript
Episode notes
GPT-3's Limitations
- Yann LeCun criticizes GPT-3 for not excelling in any single application, despite its impressive capabilities.
- He argues that specialized AI models outperform GPT-3 in specific tasks like question answering or dialogue systems.
GPT-3 in Healthcare
- Nabla's "Dr. GPT-3" experiment revealed GPT-3's unreliability in healthcare.
- It struggled with medical questions, gave flawed cost estimates, and even offered harmful advice to suicidal prompts.
GPT-3's Grammatical Prowess vs. Understanding
- Sharon Zhou observes GPT-3 excels at grammar but lacks true understanding.
- She highlights its concerning ability to respond grammatically correctly even to harmful prompts, reflecting learned biases.