Back to the blog

AI Governance

What Google's Nested Learning Paper Means for Mid-Market Teams.

By John Utley

Published

A presenter explains three memory layers to a leadership team in a meeting room.

Google's Nested Learning paper, presented at NeurIPS 2025, argues that a deep learning model is really a stack of learning problems running at different speeds, and that the model's structure and its training rule are the same kind of thing at different levels. From that idea the authors describe a Continuum Memory System and Hope, a self-modifying proof-of-concept model.

This is research, not a product. Hope was tested at 760 million and 1.3 billion parameters. Clients cannot buy or deploy it. Sophizo does not sell Nested Learning or the Hope model. We mention the paper because operators will see the headlines and should know what is, and is not, available today.

The useful translation is about speed and ownership. Some memories should update quickly. Some should barely move. If a system can change after launch, a one-time validation is not enough. You will need to version both the model and the memory, log which version made each call, and define what triggers a re-validation. Those practices map to NIST AI RMF Measure and Manage, and to ISO 42001 monitoring and continual improvement.

Mid-market teams do not need to wait for a new model to act on that. Design fast, medium, and slow memory on the tools you already have. Promote a lesson into durable policy only after a person reviews it. Give that policy a named owner and a rollback path.

What not to do: treat the paper as a buying guide, drop unreviewed chat logs into a long-term store, or assume a policy binder written for a frozen model will still be enough next year.

Sources: Google Research, Introducing Nested Learning. Behrouz et al., Nested Learning: The Illusion of Deep Learning Architectures, arXiv 2512.24695.

Keep reading

Related articles

An operator maps the connected stores used by several AI agents.
AI Governance

Log the Memory, Not Just the Model.

If you only version the model, you cannot explain last month's decision. Log which memory was in force, what was retrieved, and who was allowed to change it.

An operator arranges three layers of agent memory with a visible workflow between them.
AI Strategy

Lesson Promotion: The Review Step Most Agent Teams Skip.

A useful correction is not the same thing as a new policy. Name the person who decides when a lesson may move from a session into durable memory.

Three streams of task records pass through a human approval gate before becoming lasting memory.
AI Governance

Validate Once Is Not Enough: Continual Learning, NIST, and ISO 42001.

Controls written for a frozen model need a second life after launch: versioning, decision logs, and re-validation triggers mapped to NIST and ISO 42001.

An operator maps the connected stores used by several AI agents.
AI Governance

Your AI Will Soon Learn on the Job. Is Your Governance Ready?

Most AI governance still assumes a model that is validated once and then stays still. That assumption will not survive systems that keep learning after launch. Here is what to install now, with the tools you already have.

Related reading