Google is working on a breakthrough in artificial intelligence for continuous learning with a new technique called nested learning.
Imagine this…prompting GPT. You’ll get a good response, but not quite there yet. So we’ll add another prompt, and then another. And soon, GPT will seem to have forgotten some of its initial context, and you will need to enter it again.
It’s a common problem for AI to keep forgetting pieces of information. Google is trying to solve this problem with an experimental model called HOPE. Designed around a new paradigm called nested learning, HOPE focuses on continuous learning, rather than just focusing on what is initially trained.
Why does current AI forget?
Most AI chatbots run on large-scale language models (LLMs). They are great at generating text, solving problems, and maintaining conversations. However, they struggle with continuous learning. Every time we teach something new, we tend to forget what we learned previously. In machine learning, this behavior is called catastrophic forgetting.
Unlike the human brain, which builds on past experiences and retains them, today’s AI needs to be retrained from scratch to retain new ones. This is a major hurdle for artificial general intelligence (AGI), which aims to develop human-like AI that can continuously learn and reason.
Andrej Karpathy, a widely respected AI/ML research scientist who previously worked at Google DeepMind, recently said on a podcast that AGI is still 10 years away because no one is capable of building continuously learning AI systems. “They don’t have the capacity for continuous learning. You can’t just tell them something and they’ll remember it. They lack cognitive abilities, they’re just not functioning. It’s going to take about 10 years to solve all these problems.”
Nested learning concept
To solve this “forgetting” problem, Google has devised a solution that involves nested learning. Rather than treating AI like a single optimization problem, they divided it into several smaller learning subproblems.
Each of these sub-problems has its own context flow, which helps you learn and optimize based on your own information flow. In this way, Google allows HOPE to calculate and learn at a deeper level.
Google says this enables “learning components with greater computational depth,” allowing AI models to retain old knowledge while learning new inputs.
Proof of concept and early results
Early tests of the HOPE model were promising. When tested based on language modeling and common sense reasoning, we found it to be less disruptive and more accurate than most state-of-the-art LLM models.
This gives us hope that nested learning is more than just a theory or concept, and that we can train future models to continuously improve their learning capabilities. And if HOPE continues to retain knowledge, it could soon become a bridge between AI and adaptive intelligence more generally.
The future: smarter, more human-like AI
Once AI can actually handle nested learning, the possibilities are endless. Barring catastrophic forgetting, there are many things AI can do.
- Long-term assistants: The ability to retain knowledge allows AI agents to remember your preferences, past conversations, and evolving needs.
- Research and innovation: With nested learning capabilities, AI agents can adapt to new scientific discoveries and changing data.
- Safer and more reliable AI: The ability to retain knowledge makes AI agents more reliable because they don’t “forget” important information, thereby reducing the need to double-check information or keep inputting it into the AI.
Google believes that nested learning is the key to bridging the gap between current LLM and human continuous learning capabilities.
last word
“Forgetting” has always been one of the long-standing problems in AI. By demonstrating the potential of the HOPE model, Google has taken an important step toward the future.
Although still in its early stages, this model shows the potential for AI to not only respond but also learn and grow. If successful, this would be a breakthrough step into a future with more intelligent, flexible, and human-like artificial systems.
