When we started OpenAI, I was convinced that deep learning was genuinely working and that it improved with scale. This realization was crucial because it seemed like others weren't recognizing its potential.

When I was in college, the idea that scaling compute in transformers could lead to artificial general intelligence was considered crazy, but I eventually recognized that scaling does improve performance.

The scaling laws became a foundational principle for AI development after being popularized by researchers, including those at OpenAI and Google DeepMind, who emphasized that it's not just about making models larger, but also about training them on sufficient data.

In AI, every approach has performed better when more scale is applied, especially as foundational problems are addressed over time.

The era of scaling pre-training may not be completely over, but reasoning now appears to offer the most potential for maximizing gains from scaling compute.

The scaling laws observed in AI suggest that larger models trained longer lead to more compression and intelligence, which is supported by the philosophical grounding of these concepts.

We realized that scaling laws in AI showed predictable improvements with increased data and compute, which encouraged us to focus on building a product for knowledge work as AI matured.

The last decade focused on large-scale supervised learning, where increasing data and model size led to improved performance.

The principle 'bigger model, more data, better results' has held true for the last 12 to 15 years in AI development.