OpenAI’s new AI model takes a different approach

AI companies, including OpenAI, are shifting away from the ‘bigger is better’ philosophy for training models. Instead, they are developing techniques that allow algorithms to ‘think’ in more human-like ways. These methods aim to address challenges such as massive energy consumption, hardware failures, and data scarcity that have hindered advancements in large language models.

OpenAI’s new model, o1, uses a technique called ‘test-time compute’, allowing it to consider multiple answers and choose the best option during use. This approach improves performance in complex tasks, like problem-solving and decision-making, without needing extensive pre-training. Noam Brown, an OpenAI researcher, revealed that even brief ‘thinking’ boosts the model’s capabilities significantly.

The industry-wide shift has broader implications for AI hardware, especially as Nvidia’s chips have been critical to AI training. Experts predict a move towards distributed cloud-based servers for inference tasks, potentially reshaping the demand landscape for chips. Prominent investors, such as Sequoia and Andreessen Horowitz, are monitoring these changes closely as they may impact investments in AI infrastructure.