The Information published an article about GPT-5 with insider information Detailing the progress points of GPT-5, the reasons for GPT-4.5's failure, and the current challenges and issues faced by OpenAI Here's a compilation based on several bloggers' reports 👇 GPT-5 no longer focuses on the "quantum leap" of intelligence, but instead emphasizes practicality and user experience improvements. It can generate more clear, feature-rich, and user-friendly code, and is better at handling complex task lists and edge cases. Additionally, reasoning efficiency has improved. With a significant increase in computing power, it can output higher-quality answers. utilization has become more efficient. Introduced a "Universal Verifier" reinforcement learning mechanism that automatically verifies model output, drawing from the experience of the o series (such as o1, o3) in reasoning tasks. It feels that future LLM evaluation standards (rubric) will gradually become automated, with user satisfaction and "trust" becoming the new battleground. Issues with GPT-4.5: GPT-4.5 was initially developed as GPT-5 but performed below expectations, ultimately being released under the name GPT-4.5. The main reason is that 4.5 did not achieve significant breakthroughs in small models compared to 4O, and optimizations were not extended to large models. The primary reason for GPT-4.5's failure is the gradual depletion of high-quality web data, limited pre-training data, with pre-training limits and data scarcity becoming the main bottlenecks. Regarding company management and personnel issues: Research Director Mark Chen and Vice President Jerry Tworek have obvious friction on Slack; meanwhile, Mark Chen is a controversial figure in team restructuring and researcher departure events. Microsoft has exclusive rights until 2030, planning to hold approximately 33% of shares in a for-profit structure. Strategic negotiations are ongoing, and OpenAI is preparing for a potential IPO.