Two kinds of learning
A dog learns to fetch by repetition.
A human learns calculus once.
Today's AI runs almost entirely on the first kind.
This page explains neurosymbolic AI: the architecture that gives machines both kinds of learning, and with them both kinds of thinking. A short scroll, by Semantic Reasoning Labs.
scroll
> 01
How you train a dog
Training a dog to fetch, in a few steps:
- The dog does something.
- If it gets closer to the desired behavior, it gets a reward (for example a cultivated meat treat).
- Repeat many, many times, and the desired behavior emerges.
That is the essence of reinforcement learning. The deeper pattern (guess, get feedback, adjust a little, repeat at enormous scale) is the basic training loop behind much of modern AI. An LLM learns by guessing the next word, comparing the guess with the real text, and nudging billions of weights by a tiny amount, over trillions of words. (Formally, that stage is called self-supervised learning, but the core loop holds: a positive response leads to a small adjustment.)
And notice what this kind of learning builds: intuition. Fast, automatic, hard to explain.
References: Wikipedia: Large language model · Hugging Face: Illustrating RLHF
> 02
Humans learn this way too
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.
The thinking it produces is fast and effortless. You cannot explain how you keep balance on a bicycle. The knowledge is in the wiring, not in words.
But it is not the only way we learn as humans.
> 03
But calculus is learned once
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. If you already understand symbolic integration, learning the integral of a particular function often takes 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.
And it powers a different kind of thinking: deliberate, step by step, every step something you can write down, check, and defend.
Neuroscience sees the same split. Repetition slowly trains procedural circuits, while a new insight can be encoded by the hippocampus in a single exposure.
References: Learning, Fast and Slow in the Hippocampus
> 04
The hard problems live on the calculus side
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 usually does not "intuitively feel wrong". It is a particular clause, or a single word, that implies undesirable outcomes or uncertainties.
These are also the domains where 95% accuracy is usually not enough.
You may know the two modes of thinking as System 1 and System 2, fast and slow. The part that is usually skipped: they are built by two different kinds of learning. Repetition builds intuition. Understanding builds mental models.
References: Wikipedia: Dual process theory · Thinking Fast and Slow in AI
> 05
An AI that does both
Neurosymbolic AI gives a machine both kinds of learning, and with them both kinds of thinking. Just like humans.
Use neural networks (embeddings, or even small LLMs) for the messy, intuitive part: language, conceptualization, understanding the ambiguity of the real world and translating it into a symbolic layer in a robust and stable way.
Use symbolic reasoning for deliberate, explainable thinking: mental models, concepts, rules, relationships, constraints. The symbolic layer is explicit. Humans can inspect what the AI has learned, see why it answered the way it did, and correct either. The reasoning is deterministic: same question, same answer. It is also vastly more compute-efficient, both to train and to use.
An LLM can follow a new rule too, if you paste it into the prompt. However, the rule lives in the context window, not in the model, and is forgotten the moment it falls out. In a neurosymbolic AI, the model itself changes. Permanently, and the same mistake does not come back.
References: IBM Research: Neuro-symbolic AI · MIT 6.S191, Lecture 7: Neurosymbolic Hybrid AI
> 06
Why it failed in the 1980s, and why now is different
Symbolic AI was the mainstream bet of the 1980s. Expert systems boomed; by 1985 two thirds of Fortune 500 companies were applying them.
But alone, it was brittle. The world is messy. Symbols are clean. Going directly from the real world to symbols broke outside of very narrow domains, and by the early 1990s the field had collapsed into the AI winter.
Embeddings and modern neural networks, especially LLMs, provide what was missing: a robust bridge from the messy world into clean symbols.
Additionally, current AI allows us to encode knowledge into concepts, rules, and relationships on a scale that was never before possible.
And the combination already wins at the hard end. DeepMind's AlphaGeometry pairs a neural model with a symbolic deduction engine and matches the average score of IMO gold medallists on olympiad geometry benchmarks. Its successors reached silver-medal standard on the live 2024 olympiad.
References: Wikipedia: Expert system · AlphaGeometry, Nature 2024 · DeepMind: silver-medal standard at IMO 2024
> 07
Why it is inevitable
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, explainable, deterministic, able to learn from single examples, and far more compute-efficient.
Teaching AI calculus the way we teach dogs to fetch is a colossal inefficiency. It is why today's AI needs gigawatts of compute and the text of the entire internet to learn what a human learns from a single book, with a few watts of power and a coffee.
This is not just possible.
It is inevitable.
> 08
We are building it
We are Semantic Reasoning Labs, a research team in Prague, building a new neurosymbolic foundational model: AI you can trust by design. Inspectable, explainable, deterministic, and far cheaper to run.
Giants have proven the approach in narrow, formal domains. We are making it general.
Extraordinary claims require extraordinary evidence. See what we are building at semanticreasoning.ai, or ask for the demo: curious (a) semanticreasoning.ai.
Further reading
- Think-on-Graph (ICLR, 2024): an LLM explores a knowledge graph step by step instead of answering from memory; every answer is traceable to a graph path, and wrong knowledge can be corrected without retraining.
- Reasoning on Graphs (ICLR, 2024): relation paths grounded in a knowledge graph serve as faithful, interpretable reasoning plans for an LLM.
- Automatic Ontology Construction Using LLMs as an External Layer of Memory (2026): an LLM extended with a persistent, verifiable knowledge graph as its memory instead of relying on model weights and vector retrieval.
- ReKG-MCTS (ACL, 2025): Monte Carlo tree search explores a knowledge graph while an LLM guides the search, with no training required.
- Logic-LM (EMNLP, 2023): an LLM translates a problem into symbolic form and a deterministic solver does the reasoning; on five logical reasoning benchmarks this beats chain-of-thought prompting by 18.4% on average.
- AlphaGeometry (Nature, 2024): a neural language model guiding a symbolic deduction engine matches the average score of IMO gold medallists on olympiad geometry benchmarks.
- AlphaProof (Nature, 2025): reinforcement learning with formal proof search reaches olympiad-level theorem proving; with AlphaGeometry 2 it scored at silver-medal standard on the 2024 IMO.
- Thinking Fast and Slow in AI (AAAI, 2021): IBM researchers propose dual-process cognition, System 1 and System 2, as a research agenda for AI.
- Learning, Fast and Slow: Single- and Many-Shot Learning in the Hippocampus (Annual Review of Neuroscience, 2024): the brain itself uses two learning mechanisms; the hippocampus can encode a new experience in a single exposure, while procedural circuits need many repetitions.
- Neuro-Symbolic AI in 2024: A Systematic Review (2025): the field is growing rapidly; explainability and trustworthiness remain open research areas.
- IBM Research: Neuro-symbolic AI: frames neurosymbolic AI as the path to combining statistical learning with humanlike symbolic reasoning.
- MIT 6.S191, Lecture 7: Neurosymbolic Hybrid AI (2020, course schedule): why hybrid systems improve sample efficiency, interpretability, and robustness.
- CYC: Using Common Sense Knowledge to Overcome Brittleness (AI Magazine, 1985): the original bet of the symbolic era, a large explicit body of world knowledge so that AI could reason instead of breaking outside narrow domains.