Leaked information from trademark attorneys shows that OpenAI has registered the GPT-5 trademark on July 18.
It is not uncommon to register a trademark. It may be a protective registration, but it may also mean that OpenAI is already preparing for the release of GPT-5. Judging from their performance, the latter is more likely.
The specific functions of GPT-5 are still uncertain. Judging from the registered trademark, it involves functions such as text generation, natural language processing, voice transcription, and translation. There is no doubt that it will go further in multimodality.
The CEO of OpenAI said in June this year that GPT-5 has not yet started training, but no one in the industry believes their statement, thinking that this is just a smoke bomb, and it is impossible for OpenAI to stop and wait for its opponents to catch up.
GPT-4 was officially released in March this year. Compared with the GPT-3.5 model originally used by ChatGPT, GPT-4 has achieved leaps and bounds in several aspects: powerful image recognition ability; text input limit increased to 25,000 words; accurate answer Significantly improved performance; able to generate lyrics, creative text, and achieve style changes.
GPT-3.5 has a scale of 175 billion parameters. OpenAI has not announced the specific specifications of GPT-4, but subsequent analysis believes that it has 16 expert models, and each MLP expert has about 111 billion parameters, a total of 1.8 trillion parameters. GPT-4 is scaled by 10.
If GPT-5 is improved according to this ratio, it will break through 10 trillion parameters, and it may become the largest AI model, far surpassing all other opponents.
