Understanding the properties of handcrafted models can help inform the development of neural networks, particularly in terms of stability and performance.

The act of making something helps you understand it better, and when you test it, you know what might be going wrong because it always goes wrong.

The outcome is better if you structure the reasoning and help iterate with the model on how it should do reasoning.

The way you built the model is transparent, supported by a world-class safety and ethics team, ensuring it's built properly.

The distinction between how models claim to perform tasks and how they actually do them can reveal significant insights into their capabilities.

Mathematics can help validate models by identifying weaknesses, such as systemic correlations in economic models that may lead to failures.

My recent work involves starting with handcrafted models and parameterizing them in a simplified way, rather than using millions of parameters. I learn parameters from actual examples that I want my handcrafted model to produce, which is known as parameter estimation or optimization.

The surprising interpretability of thinking models allows us to understand how they reach their conclusions, which was not expected in earlier AI developments. The fact that AI models can articulate their reasoning provides a level of comfort regarding safety, even though it's not infinite comfort.

The scaling of models has worked consistently, and it's incredible that making these models bigger leads to predictable improvements in their performance.