The video compares the performance of GPT-4 and a new reasoning model called “o1,” using the example of counting the letter “R” in “strawberry,” where GPT-4 incorrectly identifies two "R"s instead of three due to its subword processing design. In contrast, o1 accurately counts the letters by engaging in a reasoning process, highlighting the importance of incorporating reasoning capabilities into AI models for improved accuracy in tasks requiring detailed analysis.
In the video, the presenter discusses the performance of different AI models, specifically comparing GPT-4 with a new reasoning model referred to as “o1.” The example used to illustrate this comparison involves counting the occurrences of the letter “R” in the word “strawberry.” The presenter highlights that GPT-4 incorrectly identifies the number of "R"s, stating there are only two when, in fact, there are three.
The presenter explains that the reason for GPT-4’s mistake lies in its design. This model processes text at a level that does not focus on individual characters or their counts. Instead, it operates on a subword level, which can lead to errors when tasked with counting specific letters in a word. This limitation demonstrates how traditional models may struggle with straightforward counting tasks that require precise character recognition.
In contrast, the new reasoning model, o1, is designed to approach problems differently. Before providing an answer, it engages in a reasoning process that allows it to analyze the problem more thoroughly. When asked the same question about the word “strawberry,” o1 correctly identifies that there are three "R"s, showcasing its improved accuracy.
The video emphasizes the importance of incorporating reasoning capabilities into AI models. By enabling the model to think critically about the problem and review its output, it can avoid simple mistakes that traditional models might make. This ability to reason enhances the model’s performance, particularly in tasks that require careful consideration of details.
Overall, the comparison between GPT-4 and the new o1 model illustrates the advancements in AI technology. The reasoning model’s success in accurately counting letters highlights the potential for improved accuracy and reliability in AI applications, especially in tasks that involve precise data analysis and understanding.