In short

Today's LLMs can pass professional exams but cannot retain anything from their interactions, experiencing every moment as new. This lack of continual learning is a major bottleneck to AI-driven economic transformation. Google researchers' 'nested learning' and the HOPE architecture propose hierarchical memory loops updating at different speeds to enable meta-learning while avoiding catastrophic forgetting, but it only works on tiny models so far, and scaling remains unproven.

Here’s a limitation that gets less attention than it deserves: today’s AI can ace a professional exam, but it cannot remember anything you taught it yesterday. Every conversation starts from zero.

This essay explains why that’s not a minor inconvenience but a fundamental bottleneck. A worker who couldn’t accumulate experience would be near-useless in most jobs, and that’s roughly the position frontier models are in, with “memory” features and retrieval systems papering over the gap rather than closing it. The piece then digs into the most interesting attempt to fix it: Google’s “nested learning” and the HOPE architecture, which give a model multiple speeds of memory so that slower layers effectively teach faster ones how to learn.

The honest caveat, and the reason this is an explainer, not a hype piece, is that it only works on tiny models so far. Whether it scales is one of the open questions that will shape how transformative AI actually becomes.

Key takeaways

  • LLMs cannot update long-term memory during deployment, so they can't 'learn on the job' like humans, a core limit on AI's economic impact.
  • Current memory workarounds (saved memories, RAG) are rudimentary and fail in complex learning scenarios.
  • Google's 'nested learning' / HOPE architecture uses hierarchical memory loops at different speeds, inspired by neurobiology.
  • Slower loops teach faster loops how to learn, enabling meta-learning while avoiding catastrophic forgetting.
  • It only works at ~1.3B parameters so far; scaling to frontier models is unproven.

Read the full piece

This is a summary. Read the complete essay, with all the sources and argument, on Substack.

Read on Substack →

Frequently asked questions

Why can't AI learn on the job?
Because LLMs freeze their long-term memory after training and can't update it during deployment, so they can't retain or build on what happens in interactions the way humans do.
What is nested learning in AI?
A Google-proposed architecture with hierarchical memory loops updating at different speeds, where slower loops teach faster loops how to learn, enabling continual learning without catastrophic forgetting.
Can the HOPE architecture scale to large AI models?
Not yet. It has only been demonstrated on small models around 1.3 billion parameters; scaling to trillion-parameter frontier models is unproven.

People & ideas in this piece

GoogleHOPE architectureNested learningContinual learningCatastrophic forgettingMeta-learningContext windowRAGMamba

Topics: How AI Actually Works , AI, Work & the Economy