About GPT-4.5 and Claude 3.7

GPT-4.5

• The last non-reasoning model announced by OpenAI.
• Only pre-training, supervised fine-tuning, and RLHF (Reinforcement Learning from Human Feedback) were performed.
• Extremely expensive, huge, and slow.
• Not the GPT-5 that was originally estimated to be close to the AGI level.
• GPT-5 will never exist, and there will only be o-series reasoning models in the future.

Claude 3.7

• The recent online complaint about the “intelligence reduction” of Cusor is essentially the intelligence reduction of Claude 3.5 sonnet.
• It was originally thought that Anthropic was using its computing power to train Claude 3.5 Opus, and possibly even Claude 4.
• With the release of Claude 3.7 sonnet, it seems that Anthropic and OpenAI have encountered the same bottleneck.

Opinion

• Maintain the views of the February 19th blog.
• The Scaling Law of pre-trained models has come to an end.
• New ideas are needed to achieve AGI.
• But the process of landing applications of this round of large language models is far from over!
• I hope that when I am too old to move, nursing robots will be born.