We Stopped Using the Mathematics That Works
gfrm.in
March 9, 2026
6 min read
🔥🔥🔥🔥🔥
49/100
Summary
Path dependence and disciplinary silos have contributed to the decline in the use of effective mathematical techniques in AI. Despite existing since the 1960s, these techniques have been overshadowed by simpler, "good enough" solutions marketed as AI agents.
Key Takeaways
- The deep learning boom, initiated by Alex Krizhevsky's 2012 ImageNet victory, led to a mass migration of talent and funding away from traditional decision-making methods in AI, despite those methods still being effective.
- Decision theory, Bayesian statistics, and operations research are fragmented across different academic disciplines, resulting in a lack of coherent teaching and integration of these methods in AI.
- Deep learning's approach requires less explicit specification of objectives compared to decision theory, making it more convenient for practitioners, which contributes to its dominance in the field.
- The specification problem in decision theory necessitates detailed and explicit definitions of objectives, which can be intellectually challenging and time-consuming, deterring its widespread adoption.
Coverage & discussions
Hacker News · gfrm.in
Community discussions
Community Sentiment
Concerns
- The article's premise is confusing, suggesting alternative methods to deep learning 'work' while referencing their failures, which undermines its argument.
- There is a lack of clarity regarding what 'works' means in the context of deep learning, leading to confusion about the historical narrative.
- The article is criticized as being poorly constructed, with one commenter labeling it as 'LLM-garbage', indicating a significant dissatisfaction with its content.
