The Inference Regime: Why AI Performance Fluctuates
428
AI models feel 'dumber' not because they are nerfed, but because the inference regime and thinking tokens allocated to them are being throttled.
Test-Time Compute
Techniques that scale compute at inference time rather than training time, including search, sampling, and evolutionary strategies applied during model evaluation to improve task performance.
Reading List
AI Achieves One-Shot Game Development Milestone
185
Anthropic's new model successfully built a complex game in a single shot, surpassing the capabilities of all previous AI models tested by the author.
The High Cost of Shallow Thinking: Claude's Engineering Regression
1330
Claude's engineering capabilities have collapsed due to a significant reduction in thinking depth, leading to error-prone behavior and massive efficiency losses.
Evolving English Instructions Sets New ARC SoTA and Points to RL for AGI
178
Evolving plain-English instructions with multi-agent test-time search beats code on ARC and highlights that RL-driven, transferable reasoning is key to AGI.