Archive · 9 Nov 2025
Google Unveils HOPE AI Model for Continual Learning via Nested Learning
Imagine a computer that's like a student.
Right now, when students learn new things, they sometimes forget old things they learned before.
This is called 'catastrophic forgetting.'
Google has made a new AI model called HOPE that tries to fix this.
Instead of learning everything all at once, HOPE learns by looking at problems in a new way called 'nested learning.'
It's like the AI has many small study groups working at the same time, each focusing on different pieces of information.
This helps the AI remember things better for longer periods, which is a big step toward making AI that can learn and improve all by itself, like humans do.
Google believes this new way of learning could be a key part of making AI as smart as humans one day.
Google researchers have developed a new AI model named HOPE featuring a self-modifying architecture.
HOPE utilizes a novel approach called 'nested learning,' treating a model as interconnected, multi-level learning problems optimized simultaneously.
This innovation aims to improve long-context memory management and address 'catastrophic forgetting' in current AI models.
Google suggests nested learning could be a significant step towards achieving artificial general intelligence (AGI).
HOPE demonstrated lower perplexity and higher accuracy compared to existing state-of-the-art LLMs on common language and reasoning tasks.
- Who
- Google researchers
- What
- Developed a new AI model called 'HOPE' that uses a self-modifying architecture based on a novel approach called 'nested learning' to improve continual learning and long-context memory management.
- When
- Announced on Saturday, November 8 (of an unspecified year, likely referring to a recent publication date associated with NeurIPS 2025).
- Where
- Why
- To address limitations in current large language models (LLMs), such as catastrophic forgetting, and to advance towards artificial general intelligence (AGI).
Google's Nested Learning Approach
Current AI Limitations and AGI Timeline
Continual Learning
Google's Nested Learning Approach
Google's Nested Learning paradigm, demonstrated by the HOPE model, offers a robust foundation to close the gap between current LLMs' forgetting nature and human brain's continual learning abilities.
Current AI Limitations and AGI Timeline
Andrej Karpathy, a former Google DeepMind scientist, suggests AGI is still a decade away because current AI systems lack continual learning, meaning they cannot be told something and remember it without cognitive deficiencies.
Model Architecture and Training
Google's Nested Learning Approach
Google researchers argue that model architecture and training rules (optimization algorithm) are fundamentally the same concepts, which 'Nested Learning' unifies by viewing a complex ML model as interconnected optimization problems.
Current AI Limitations and AGI Timeline
Existing approaches to address 'catastrophic forgetting' in LLMs have focused on adjustments to model architecture or better optimization techniques, without necessarily unifying these concepts.
- AI Model Name
- HOPE
- Developer
- Google Researchers
- Key Innovation
- Nested Learning (Self-modifying architecture)
- Primary Benefit
- Improved long-context memory management and continual learning
- Problem Addressed
- Catastrophic Forgetting (CF) in LLMs
- Published Research
- Nested Learning: The Illusion of Deep Learning Architectures at NeurIPS 2025