The past few years have witnessed a dramatic shift in the development of Large Language Models (LLMs). At one time, the future of LLMs hinged on scaling up base models like GPT. This involved feeding more parameters with more data on more GPUs. GPT 4.5, a product of this approach, offers a glimpse into what LLMs would look like without the recent innovation of extended thinking time. However, upon testing, it’s evident that the initial excitement over GPT 4.5 may have been premature.
This video is from AI Explained.
GPT 4.5, despite its high cost, doesn’t appear to significantly outperform its predecessor in terms of emotional intelligence, as one might expect. Moreover, its inability to outperform smaller versions of the reasoning model 03 in science, mathematics, and coding benchmarks also raises eyebrows. Although GPT 4.5 is touted for its improved emotional intelligence, the results from tests on humor and empathy have been mixed.
The shift in emphasis from pre-training to reasoning is a significant turning point in AI development. Pre-training models like GPT 4.5 involve feeding more data into the model, while reasoning models like 03 use reinforcement learning to think for extended periods before generating a response. Despite the incremental improvement that GPT 4.5 offers over GPT 4, the more significant leap seems to be in the direction of reasoning models.
Furthermore, the significant cost of GPT 4.5 raises questions about the sustainability of this model. The high cost of maintaining GPT 4.5 in the API long-term may not be justifiable, especially given that the performance increment over GPT 4 is not as significant as anticipated.
The development of reasoning models, however, offers exciting prospects. The Claude series of models from Anthropic, for instance, show promise as base models for future reasoning models. Although GPT 4.5 was expected to outperform these models, the results suggest otherwise, indicating a potential shift in leadership in the world of LLMs.
The anticipated release of GPT 5, which will incorporate reasoning into the improved base model of GPT 4.5, is likely to be a game-changer. With the focus now on reasoning rather than just pre-training, the future of LLMs appears to be on a new trajectory.
In conclusion, while GPT 4.5 marks a step forward in the development of LLMs, it is not the revolutionary leap that many anticipated. The future of LLMs seems to be leaning more towards reasoning models that think for extended periods before generating a response. While this presents its own challenges, it also opens up exciting possibilities for the future of AI.
