The video discusses the recent release of OpenAI’s new model, referred to as “Strawberry,” which has sparked considerable debate and interest among users. The speaker highlights that Strawberry operates on a principle known as “Chain of Thought,” which allows the model to engage in self-prompting or meta-prompting. This means that instead of providing a single-shot answer, the model takes time to think through problems step by step, leading to longer and more detailed outputs. While this approach is seen as an innovation, the speaker notes that it is not entirely new, as similar techniques have been demonstrated by other models like Claude 3.5.
The speaker critiques the cost-effectiveness of Strawberry, pointing out that it is significantly more expensive than previous models, with a price of $15 per million tokens compared to $0.15 for the existing GPT-4 mini model. Despite the potential for improved performance, the speaker argues that the cost may not justify the benefits, especially since the model still struggles with some basic tasks. The discussion also touches on the business implications of this pricing strategy, suggesting that OpenAI may not have achieved product-market fit and is still working towards profitability.
The video emphasizes the iterative nature of AI development, where each new model builds upon the successes and failures of its predecessors. The speaker explains how OpenAI has historically bootstrapped its models, moving from plain GPT-3 to instruct-aligned models and then to chatbots. The introduction of Strawberry is seen as another step in this progression, but the speaker expresses skepticism about its significance, arguing that it represents only a minor incremental improvement rather than a groundbreaking advancement.
The speaker also discusses the concept of a “data flywheel,” where increased usage of the model generates more data, which can then be used to improve future iterations. This cycle of data accumulation is crucial for enhancing the model’s reasoning capabilities. The speaker notes that while Strawberry may not be revolutionary, it does provide OpenAI with valuable data that can help refine their models over time, leveraging their first-mover advantage in the market.
In conclusion, the speaker remains cautious about the implications of Strawberry, suggesting that while it may enhance user experience and usability, it does not signify a major leap in AI capabilities. The underlying science of AI development appears to be slowing down, with longer intervals between significant advancements. However, the commercial deployment of AI technologies is likely to accelerate, indicating that while the foundational science may not be progressing rapidly, the application and integration of these technologies into the market are on the rise.
