AI Is Learning at the Wrong Level of Abstraction — Matthieu Wyart

Summary

In this MLST episode, Matthieu Wyart frames modern deep learning through statistical physics: optimization landscapes, double descent, and jamming-like phase transitions. The central claim is that deep architectures are biased toward coarse-grained, hierarchical variables, allowing them to recover abstractions from data much more effectively than shallow models. He extends this idea to language and vision, arguing that next-token and diffusion objectives can learn structure, but that predicting in latent space should be more sample-efficient because it uses stronger signals between higher-level concepts. Wyart also discusses a synthetic context-free grammar setup in which shallow networks memorize while deep networks recover generative rules, and he says his scaling-law work on natural language can be expressed using token-correlation decay and residual entropy. The episode ends on a broader scientific point: current models can show narrow syntactic competence and some creativity, but genuine scientific creativity likely requires better training procedures, not just more scale.

Notes

• Wyart introduces the core puzzle: humans and machines both generate novel outputs, but machines appear to require vastly more data, raising the question of whether current objectives are learning at the wrong abstraction level.
• He says the conceptual target is not raw token prediction alone, but learning coarse-grained abstractions from statistics.
• Tim Scarfe frames the recent paper explicitly as a latent-space versus token-space proposal.
• Wyart says introspective methods that learn from their own latents are more sample efficient and reach the same abstractions faster.
• He describes statistical physics as a way to study many interacting entities and collective phenomena across scales.
• He uses rough energy landscapes and metastable states as the physics analogy for complex learning dynamics.
• He says machine-learning loss landscapes mirror sand-like phase transitions: underparameterized models get stuck, overparameterized ones flow through flat valleys.
• He links this to double descent and says physicists would call the peak a jamming transition.
• Wyart clarifies that the sand analogy is about gradient-descent dynamics, not the computer’s physical substrate.
• He distinguishes optimization constraints from the constraints built into the data world itself, such as syntactic validity.
• He argues physics contributes by linking theory and experiment, simplifying messy systems into tractable caricatures, and extracting useful analogies.
• He cites the Ising model as a successful coarse-grained model that captured phase transitions without microscopic detail.
• He says some neuroscience theory felt detached from function, intelligence, rules, constraint, and language, which pushed him back to physics.
• He says modern generative models are fascinating because they create novel faces and sentences, despite prior claims that this should be extremely hard.
• He models language with context-free grammars and hidden variables on trees, using randomized production rules to keep correlation calculations tractable.
• He says deep architectures learn the hidden hierarchy from data statistics, while shallow models fail to recover the interesting generative grammar.
• He states the paper’s main result as a counterexample to Chomsky: deep architectures can learn creativity from few examples, with polynomial sample complexity in sentence size D.
• He says this does not imply the brain is just a deep net, but the brain’s language learning is vastly more sample efficient than current machines.
• He says abstractions emerge progressively during training, and very long recursive sentence chains are mostly an academic issue for current transformers.
• He and Scarfe agree current systems can be syntactically competent, but not broadly creative in the human sense.
• Wyart says scientific creativity needs modeling, simplification, and interaction with the world, not just recombination or scale.
• He says interacting with and simplifying the world is essential, so raw internet-scale text alone is unlikely to teach scientific creativity.
• He uses Word2Vec as the analogy for low-level abstraction, then argues deep models recursively build supra-meanings from low-level co-occurrence statistics.
• He says higher abstractions are harder to learn because the signal becomes noisier when it is defined through low-level predictions like pixels or nearby tokens.
• He says deep architectures converge toward similar abstractions across transformers and CNNs, and claims the data needed for each abstraction level scales similarly across architectures.
• He argues the curse of dimensionality makes naive interpolation hopeless, but hierarchical structure can reduce the problem to polynomial sample complexity in the input dimension.
• He says deep nets help by discovering hidden coarse-grained variables, such as nose, ears, and other summaries instead of pixel-by-pixel descriptions.
• He says latent or introspective training is more sample efficient because it predicts representations of masked content rather than raw tokens or pixels.
• He says the latent-prediction setup is like a teacher-student model where the student predicts the teacher’s representation of occluded data, increasing signal-to-noise for abstraction learning.
• He says the main advantage of latent prediction is stronger predictive signal, so fewer data are needed to extract abstraction from noise.
• He says the group is still testing the theory empirically and does not yet know whether learned encoders can be turned into competitive generative decoders.
• He says representation learning is already competitive for segmentation and classification, but generative decoding remains open.
• He says in the tree-structured model, learning hidden hierarchy costs exponentially in tree depth, but that is still polynomial in input dimension because the input itself grows exponentially with depth.
• He says diffusion models and next-token prediction are equivalent in sample-complexity terms, differing mainly in the order of masked content recovery.
• He says the scaling-law theory uses token-token correlation decay and residual text entropy after N tokens as the key measurable exponents.
• He says the language-scaling work was tested at roughly 1 billion parameters, 1 billion tokens, and around 50-token context windows, or two to three sentences.
• He closes with a philosophical note that avoiding mistakes may mean staying on the beaten path, while some scientists want to explore riskier territory.