A New Kind of Learning Systems
summary
Argues that instead of updating model weights, future systems learn by writing, retrieving and editing natural‑language facts in memory, proposing an architecture with an LPU (linguistic processing unit), a controller and append‑only memory plus read‑only model state. Shows a working paper‑finder prototype that loops an LLM with search tools and returns relevant papers—useful but scripted, not yet agentic. Claims the surprise: online declarative learning (a.k.a. declarative online learning) is feasible and necessary, but current tool‑chains only partially work and forget without explicit memory updates.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 2:51Surrounding AI systems
- 10:32Past versus future multistep
- 17:22Searching with AI
- 20:18Dreaming new learning systems
- 31:50Summary and call to action
related talks
01h 00mBetter Learning through Programming Languages
46mThe Anatomy of Memory in Humans and AI Agents
45mLarge language models are not a paradigm shift
45mReasoning in Natural Language
13mStrategies for Enhanced Personal Knowledge Management
40mHuman Learning
51mTowards a Pragmatic Understanding of Large Language Models
42mQuestions for Theory in the New Age of Machine Learning