AGI is decades away, not imminent. The field has three compounding problems it hasn't acknowledged: a specification problem, an architectural problem, and a premature scaling problem. Current approaches cannot produce AGI without fundamental breakthroughs in modeling non-conscious cognition.

The Three Arguments

1. The Subconscious/Unconscious Gap

Current AI systems model the outputs of conscious cognition — language, reasoning, decision-making — but have no equivalent of the subconscious/unconscious layer that generates those outputs in biological intelligence. Human cognition is not a visible process running on a transparent substrate. The vast majority of it is inaccessible to the conscious observer, and therefore inaccessible to the training signal.

You cannot train a model on data it cannot observe. The training corpus is almost entirely the product of conscious, deliberate, language-mediated cognition. Everything below that surface — intuition, embodied knowledge, emotional architecture, dream logic — is systematically absent from the record.

2. Epistemological Monoculture

AI development is conducted almost entirely within a Western Cartesian epistemological frame. The mind is conceived as a computational process, cognition as information processing, intelligence as optimization toward a target. These assumptions are embedded in every architectural choice, every benchmark, every evaluation framework.

This isn't a philosophical objection to the enterprise. It's a specification risk. If the Cartesian model of mind is incomplete — and there are strong reasons to think it is — then AGI built on that model will be incomplete in the same ways, regardless of scale.

Other knowledge traditions locate cognition differently: embedded in the body (Merleau-Ponty), distributed across the environment (Gibson, enactivism), relational and participatory (most non-Western frameworks). None of these models are represented in current architectures.

3. Neurodivergent Cognitive Architecture

The training corpus is systematically underrepresented in neurodivergent cognition. ADHD, autism, dyslexia, and related conditions represent genuinely different cognitive architectures — not deficits, but alternative processing modes. A model trained predominantly on neurotypical output will not generalize to neurodivergent intelligence, and therefore will not generalize to the full range of human intelligence.

A system that cannot model the full distribution of human cognition is not AGI.

The Premature Scaling Problem

The field has confused capability gains from scaling with progress toward AGI. Scaling a language model makes it better at language tasks. That's not the same as progress toward general intelligence. The capabilities look impressive because language is the primary medium of human knowledge representation — but language is not cognition. It's the trace of cognition.

Conclusion

AGI requires solving problems the field hasn't named yet, let alone solved. The timeline discourse is premature because the specification is incomplete. We don't know what we're building toward, which means we can't know how far away we are.