A shocking 2009 neuroscience study found fMRI activity in a dead salmon, revealing how easily tools detect false patterns without proper controls. This mirrors a critical flaw in modern machine learning: benchmarks celebrate model improvements that vanish against simple baselines. Recent research shows null models—ignoring input entirely—achieved 80-90% win rates on AlpacaEval, exposing how benchmarks measure formatting, not intelligence.

Similarly, ImageNet models classify based on texture, not shape, while XGBoost (a 2016 algorithm) consistently outperforms newer deep learning architectures on tabular data. The root issue is 'shortcut learning' and publication bias favoring 'novel architectures' over rigorous baselines. Andrew Ng's data-centric AI philosophy proves data quality trumps model complexity, as demonstrated by Microsoft's Phi models: small, textbook-trained models beating massive web-scraped ones.

For practitioners, this is liberating—focus on data pipelines, prompt engineering, and proper evaluation controls rather than chasing the latest LLM releases. The real innovation isn't architectural; it's owning what you control.