Claims

A concept is literally a region in space within a neural network—during the learning process, similar examples cluster together in the hidden layers, so all instances of 'two' occupy one region while all instances of 'three' occupy another, showing that abstract concepts are geometric.

Just like humans, recent AI systems produce better results when allowed to think out loud and reason step by step before giving an answer, and experiments showed that instead of building bigger models, systems could simply be allowed to think longer, showing that neural networks like human minds can use both fast intuition and slow deliberate reasoning.

Like the famous 'wug test' used with children to test grammatical understanding, GPT-3 could learn new concepts just by describing them and immediately use them naturally in new contexts, a capability known as 'in-context learning'.

The better a neural network can predict the next word in text, the more it understands language—this is important because prediction is a proxy for understanding, as shown by the analogy that understanding a complex detective novel means being able to correctly predict the final reveal of the criminal's identity.

The dominance of symbolic AI based on logical propositions that chain together was intellectually appealing but ultimately unsuccessful, whereas LLMs—which are not based on logical or compositional structure but rather on scaling data and computation—have proven far more capable than symbolic approaches at producing human-like behavior.

Mechanistic interpretability—the field trying to understand what goes on inside neural networks—has found that the internal structures are 'a mess' without clear connection to intuitive categories we expected, further confirming that human cognition is not built on intelligible symbolic structures.

Arithmetic is arguably now solved as a capability for large language models (they can reliably do arithmetic calculations), but the question of whether they're really 'doing arithmetic' (original calculation) versus recognizing patterns from training data (arithmetic problem patterns) remains ambiguous.