The Opportunities and Risks of Foundation Models (crfm.stanford.edu)

A multi-author report maps the technical landscape of foundation models, laying out both opportunities and systemic risks across modeling, training, adaptation, evaluation, systems, data, security, theory, robustness and interpretability. It identifies five core modeling attributes needed for next‑generation foundation models—expressivity, scalability, multimodality, memory capacity and compositionality—and argues that progress will require new architectures (beyond current transformers), principled, domain‑general training objectives, and broader adaptation strategies that go past full fine‑tuning to include lightweight or constraint‑driven methods (e.g., temporal updating, regulatory compliance). The report also calls for reformed evaluation practices that measure inherent model capabilities, control for adaptation resources, and report non‑accuracy metrics (robustness, fairness, efficiency, environmental impact). On systems and data it advocates co‑design of algorithms, software and hardware (noting retrieval, mixture‑of‑experts and parallelism strategies) and proposes a “data hub” for selection, curation, documentation, inspection and legal governance of training corpora. The authors flag foundation models as high‑leverage, single points of failure: they surface concrete security and privacy vulnerabilities (adversarial triggers, memorization, function creep), plus limits on robustness to distribution shift—adaptation helps but doesn’t solve extrapolation or spurious correlations. They emphasize gaps in theory (training vs adaptation discrepancy) and call for interpretability research framed as “one model–many models” to link foundation representations to adapted task behavior. Implication: realizing safe, reliable foundation models demands integrated advances in architectures, objectives, data practices, systems engineering, evaluation standards, security/privacy defenses and new theoretical and interpretability tools.