Investment Rating - The report indicates that the company is expected to achieve AGI (Artificial General Intelligence) by 2027, with significant advancements anticipated in the near future [30][32]. Core Insights - The company has started training a 125 trillion parameter multimodal model, originally planned for release in 2025, but the launch was canceled due to high inference costs [31]. - The training of GPT-5 was completed in December 2023, achieving a proto AGI with an IQ of 48 [30][33]. - The report discusses the correlation between the number of parameters in AI models and their performance, suggesting that a model with around 100 trillion parameters could match human-level intelligence [6][9]. Summary by Sections Model Development - The company has been working on a model with 125 trillion parameters, which is significantly larger than previous models like GPT-3 and GPT-4 [31][32]. - The original GPT-5 was planned for release in 2025 but has been renamed due to the cancellation of its launch [31]. Performance Metrics - The report highlights that the performance of AI models tends to reach human-level capabilities when the parameter count approaches that of the human brain, estimated at around 100 trillion synapses [6][9]. - The training of GPT-4 was completed with a parameter count of 1 trillion, while the full model is expected to be much larger [99][101]. Future Projections - The company aims to achieve full AGI by 2027, with various stages of model development planned leading up to that date [30][32]. - The report suggests that the advancements in AI capabilities are driven by increased computing power and data availability, which have been significantly enhanced in recent years [39][40].