Scaling hypothesis hits wall, LLMs learn backwards

The era of "brute-force scaling" is ending as frontier models plateau on benchmarks requiring true out-of-distribution logic.

• –March 2026 ARC-AGI-3 scores show ChatGPT 5.4 and Claude 4.6 failing on over 99% of novel puzzles.
• –LLMs function as massive statistical lookup tables rather than causal world models, leading to "spiky" and unreliable intelligence.
• –Recent performance jumps are largely attributed to engineered post-training optimizations (RLHF, RAG) rather than fundamental scaling gains.
• –The path to AGI likely shifts toward interactive architectures like "StochasticGoose" that prioritize real-time exploration and hypothesis testing.