Neurosymbolic AI

> 01

Reinforcement learning

Training a dog to fetch in few steps:

• The dog does something.
• If it gets closer to the desired behavior, it gets a reward (for example cultivated meat snack).
• Repeat many, many times, and desired behavior emerges.

That is the essence of reinforcement learning, the basic training pattern behind much of modern AI (definitely all LLMs and much else).

> 02

Human intelligence

Humans learn this way too.

It is how we learn for example motor skills (riding a bicycle), speech (especially when you learn intuitively as a child), and many forms of innate knowledge we call intuition.

But it is not the only way we learn as humans.

> 03

Another kind of learning

Humans do not learn mathematics the same way dogs learn to fetch.

Once we understand calculus, a new piece of knowledge does not need to be repeated many times.

If the rule fits into our existing mental model, we can use it immediately (for example, if you already understand the concept of symbolic integration, learning integration for a particular function, often requires just a single exposure).

Practice still matters. But to figure out all the implications of the new piece of knowledge, not to learn it in the first place.

This is a different kind of learning. Instant. Much closer to how symbolic AI operates.

> 04

Symbolic Reasoning

Math, science, law, strategy, and many hard problems are much closer to this way of thinking and learning, than to learning how to fetch.

Our mental model of the domain is built from explicit concepts, rules, and relationships (we can fully articulate). If a legal contract is wrong, it does not "intuitively feel wrong" (usually), it is a particular clause, or a single word, that implies undesirable outcomes or uncertainties.

> 05

Neuro-symbolic AI

Neuro-symbolic AI combines both "thinking" modes. Just like humans.

Use neural network based approaches (embeddings, or even small LLMs) for the messy part: language, conceptualization, understanding the ambiguity of the real world and translating it to a symbolic layer in a robust and stable way.

Use symbolic reasoning for deep thinking: mental models, concepts, rules, relationships, constraints. A great advantage is that the symbolic layer is explicit. Humans can inspect it, understand it, and correct what the AI has learned and how it reasons. It is also vastly more compute-efficient both to train and to use.

> 06

Why now

Symbolic AI was big in the 1990s.

But alone, it was brittle. The world is messy. Symbols are clean. Going directly from the real world to symbols breaks outside of very narrow cases/domains.

Embeddings and modern neural networks, especially LLMs can solve this.

Additionally, current AI allows us to encode knowledge into the concepts, rules and relationships on a scale that was never before possible.

That changes everything.

> 07

Punchline

That is how humans think, they combine both.

Put both approaches together in an AI, and you get an AI that can operate in the real world, but is also inspectable, interpretable, able to learn from single examples, and far more compute-efficient*.

This is not just possible.

It is inevitable.

*Since teaching AI the calculus the same way we teach dogs to fetch is a colossal inefficiency; no wonder we need gigawatts of compute and data from the entire internet to learn things human can learn from a single book, with a few watts of power and coffee.