LLMs reasoning and agentic capabilities over time
summary
Claim: scaled autoregressive LLMs plus a search loop and verified objective produce surprising leaps — an autoformalize + search-loop pipeline over formal proof assistants yields an IMO medal (silver/bronze reported). Tools: transformers scaled from ~1.5B to ~200–300B params, proof checkers, reinforcement-style search. Measured effects: reliability moves from crappy to average human for GPT‑4 in ~30 minutes tasks. Built product: Dust: many narrow assistants with infra connectors, validated by ChatGPT uptake and early deployments.
AI-written summary · about
chapters (18)
- 0:00AI phase transitions
- 1:06AlphaZero breakthrough
- 2:30AI math achievements
- 2:51Deep learning history
- 4:37AlphaGo search loops
- 6:02Transformers and text generation
- 7:08Meta-learning to human AI
- 8:11AlphaProof math breakthrough
- 10:19AI scaling patterns
- 11:00Scaling limits and tasks
- 12:03Convergence on reasoning models
- 13:09Lack of generalization across modalities and tasks
- 13:52New paradigm beyond LLMs
- 14:54Need for synthesis architectures
- 15:35OpenAI and Dust products
- 16:58AI in real-world work
- 17:42Dust infrastructure
- 18:45Call to build AI products
related talks
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety
36mLLMs and AI agents demystified
46mPlaying the long game
37mExpanding the Capabilities of LLMs Through Community Collaboration
01h 18mThe Enigma of LLMs: on Creativity, Compositionality, Pluralism, and Paradoxes
01h 22mDebate: Sparks versus embers
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
24mHarnessing the Power of LLMs in Advanced Analytics