XNews

Brain-inspired AI & LeCun interview: world models and neuroscience lessons

Brain-inspired AI & LeCun interview: world models and neuroscience lessons

Neuroscience-informed analysis of what brains teach AI about efficiency, learning and robustness. - Adam Marblestone: brain achieves high generalization with low data and 20W power - Evolution encodes staged, developmentally-timed loss functions - Brain splits 'guidance' (hardwired) and 'learning' (cortex) systems - Cortex acts as a general predictor; guided by innate drives - Brain enables 'all-direction' multimodal inference rather than next-token only - Energy efficiency due to compute-in-place, physical randomness and sparse low-frequency firing - Proposed AI directions: multi-loss training, human-activation alignment, connectionomics, and Lean/formal proof usage - Lean/formal methods (Lean, AlphaProof) enable provable math and software correctness

Key insights

Brain uses staged innate priors plus general learning for data-efficient intelligence

Marblestone explains that evolution provides temporally staged innate 'guidance' subsystems (e.g., face recognition triggers) while the cortex learns general models. This separation lets organisms generalize from far less data than current large models, because learning is scaffolded by specialized drives and sensors.

Brain architecture enables multimodal, all-directional inference

Unlike LLMs optimized for next-token prediction, the brain learns joint distributions across variables enabling cross-modal inference (e.g., infer sound from sight). This 'all-direction' reasoning supports robust generalization in novel contexts.

Physical design yields extreme energy efficiency and inherent sampling

Biological factors—storage-compute co-location (synapses), physical ionic randomness for sampling, low-voltage low-frequency operation and sparse activation—explain the brain’s ~20-watt efficiency and suggest hardware and algorithmic opportunities for low-power AI.

Formal methods and connectionomics can inform AI rigor and safety

Combining formal proof systems (Lean) with richer structural neuroscience data (connectomes) could provide stricter verification and more grounded training signals, supporting provable correctness in math and safer AI behaviors.

Source posts