4x Compute ≠ 4x AI Model Performance Improvement

I do not believe that simply throwing more resources/capital/compute at Gen AI models will drive performance improvement in a straight line. In fact, there is evidence of diminishing returns.

• Multi-modal data ingestion was supposed to enhance model reasoning but did not. (Noam Brown of OpenAI HERE). 

• Anthropic's Claude 3.5 Sonnet (Best-in-class frontier model) is only marginally better than Claude 3.0 Sonnet despite being trained on 4x compute. What does that say about frontier models' ability to scale? 

• More compute probably isn't the answer.

---