DNA Hardcoding: The Flaw in LeCun's LLM vs Human Learning Comparison
Does millions of years of evolution give human babies an edge LLMs can never match?
The debate over whether large language models (LLMs) are analogous to human intelligence has a new twist. Yann LeCun, Meta's chief AI scientist, often argues that both humans and LLMs learn from sensory data using similar predictive principles. But a Reddit user, ChippingCoder, pushes back hard, claiming LeCun ignores millions of years of evolutionary pretraining hardcoded into human DNA. Babies are born with neural circuitry tuned for spatial reasoning and physical world modeling—like knowing that objects can't pass through each other or that closer objects appear larger. LLMs, trained purely on text and some static images, never get that inductive bias.
If the critique holds, it exposes a fundamental gap in the Transformer architecture's ability to reason about the physical world. The user suggests LLMs would perform poorly on tasks like judging relative distance or whether two objects are touching—skills a human toddler masters instinctively. The question then becomes: can data-driven training alone ever compensate for that genetic head start? For AI researchers, this underscores the importance of building in inductive biases—like geometric priors or causal structures—rather than relying solely on scale. It also raises deep questions about what 'general intelligence' really means when biological systems come with eons of debugged code.
- Humans inherit millions of years of evolutionary pretraining encoded in DNA, giving babies innate spatial reasoning and physics modeling.
- LLMs lack this foundation, likely failing at basic visual tasks like judging object distance or detecting contact between objects.
- If LeCun underestimates this genetic hardcoding, current Transformer architectures may need explicit inductive biases to approach true general intelligence.
Why It Matters
This critique challenges AI scaling orthodoxy, arguing that pretrained genetic priors are essential for achieving robust physical world understanding.