The AI Industry Is Built on a Big Unproven Assumption (www.bloomberg.com)

The AI industry’s dominant bet — that bigger models trained on ever more data and compute will reliably produce increasingly capable, general intelligence — is the “big unproven assumption” at issue. Empirically, scaling laws and self-supervised pretraining have delivered impressive gains and surprising emergent behaviors, but it’s not proven that continued scale alone will solve core problems like robust generalization, causal reasoning, alignment, or safety. Reliance on massive, often noisy web datasets, opaque training pipelines (e.g., RLHF), and benchmark-driven progress risks brittle systems that fail under distribution shift, embed biases, and consume vast compute and energy resources. For the AI/ML community this matters because research trajectories, investment, and infrastructure are concentrated on scaling rather than algorithmic or evaluation breakthroughs. Technical implications include potential diminishing returns from larger models, looming compute and data bottlenecks, reproducibility and centralization of capability in a few labs, and the need for better metrics for robustness and alignment. The takeaway: while scale has been a powerful lever, validating whether it suffices — and complementing it with algorithmic efficiency, causal/structured modeling, curated data, and rigorous safety evaluation — is critical for sustainable, trustworthy progress.