How AI Models Get Smarter
summary
Three concrete scaling laws drive current gains: pre-training on unlabeled internet text, post-training (fine-tuning and RLHF) to shape behavior, and test-time scaling (models prompting themselves to reason). The talk cites MMLU (≈15,000 questions) and claims training a GPT‑4–scale model cost would drop from ~5,000 GWh (2014 estimate) to ~3 GWh on Blackwell, and per‑token energy from ~42,000 J to ~2 J. Demonstrated tools include transformers, RLHF, NeMo Guardrails and NeMo Evaluator, an agentic Titanic‑data Python demo (25% English prompts) that produces charts and works end‑to‑end. Guardrails are mandatory.
AI-written summary · about
chapters (8)
- 0:00Getting Smarter
- 1:34Training Evolution
- 4:23Transformer Era
- 7:46Post Training
- 15:07Reasoning Upgrade
- 23:41Guardrails Matter
- 26:17Building Agents
- 28:50Conclusion
related talks
47mThe Rise of AI: A Reality Check on Generative AI
02h 50mUnderstanding and Using AI Language Models
18mTrends Across the AI Frontier
27mLarge Language Models and the Future of AI
23mUnveiling the Journey of NLP
14mGreener Intelligence: Reducing AI’s Carbon Footprint
38mArtificial Intelligence: How Much Will It Cost You?
29mThe Rise of AI: A Reality Check on Generative AI