Beyond Algorithms – on Human(oid) Intelligence
summary
Claims that post‑2017 progress comes more from data and teaching than new math, with transformers (self‑attention) enabling parallel training but sacrificing recurrent memory; current work focuses on agent architectures and a next‑gen design combining memory + agency + huge context windows to enable planning and world models. Cites token counts vs book estimates (training tokens ≈ hundreds of millions of books, global corpus ≈ 150 million books), warns of catastrophic forgetting during continued pretraining, and pitches model vendors offering guided domain adaptation as a clear business.
AI-written summary · about
chapters (8)
- 0:00Introduction and welcome
- 1:13Limitations of current AI algorithms
- 5:48Thought experiment on algorithm ban
- 12:53Human knowledge and civilization
- 18:40Knowledge paradox and model training
- 23:57Future of AI and human knowledge
- 27:10Impact of accessible AI systems
- 30:25Redefining human value in AI era
related talks
46mAI vs Humans Rewriting the Rules of Success
52mThe Agentic Shift- Are We Reasoning, Or Just Retrieving Smarter?
04mFrom big data to human-level artificial intelligence
54mAI vs Humans: Rewriting the Rules of Success
25mThe Great AI Game
44mKeynote: Tracking the Pulse of Generative AI
43mAI Systems: Foundations for Artificial Minds
41mAI: Q&A