'GPT-4' announcement, top 10% of the bar exam & insanely high performance in Japanese & image processing and programming are also possible, and it is also described as 'impact equivalent to the first iPhone'

On Tuesday, March 14, 2023, OpenAI officially announced `` GPT-4 ''. GPT-4 has been developed as a ``multimodal AI'' that accepts not only text but also image input, and can be used to ``solve test problems for humans,'' ``program according to requests,'' and ``what is interesting about this image? Realize the operation such as 'answer questions such as'. In addition, the performance in languages other than English, including Japanese, has reached a very advanced level.

GPT-4 Research

Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: https://t.co/TwLFssyALF pic.twitter.com/lYWwPjZbSg

—OpenAI (@OpenAI) March 14, 2023

OpenAI has developed high-performance AI such as chat AI 'ChatGPT', but the performance of GPT-4 announced this time greatly exceeds existing AI. Microsoft, which has a cooperative relationship with OpenAI , describes GPT-4 as ``a presence that gives the same impact as when the first iPhone appeared''.

ChatGPT based on 'GPT-3.5' had the performance of 'entering the bottom 10% of the bar exam', but GPT-4 had the surprising performance of 'entering the top 10% of the bar exam'. It is said to perform. Checking the score table published by OpenAI, GPT-4 scores significantly higher than GPT-3.5 in bar exams such as ``Uniform Bar Exam'' and ``LSAT'' and academic achievement tests such as ``SAT''. I understand. Also, these tests were done without 'special tuning for testing'.

GPT-4 also scored significantly higher than GPT-3.5 in benchmarks such as '

GPT-4 also performs well in languages other than English. Looking at the graph below showing the performance in each language (the longer the bar, the higher the performance), ``Performance when processing Japanese with GPT-4'' is ``Performance when processing English with GPT-3.5. ” is exceeded.

GPT-4 can recognize images as well as text. For example, if you show the following image to GPT-4 and ask ``Please tell me the ``unusual point'' of this image,'' It is the point that I am using ',' he answers accurately.

With ChatGPT equipped with GPT-3.5, it was possible to 'have the source code written'. GPT-4 has a sophisticated source code writing function, and it is possible to complete software interactively. You can check how you actually request programming from GPT-4 in the following movie.

GPT-4 Developer Livestream-YouTube


While presenting some conditions to GPT-4, I asked 'Please make a BOT that can talk with GPT-4 on Discord.'

Then GPT-4 writes the code smoothly.

However, when I ran the code written by GPT-4, an error occurred.

Report 'I got this error' to GPT-4.

Then, he apologized for the error and wrote a modified version of the code.

After that, as a result of repeating 'an error occurred and reported to GPT-4' several times, 'BOT that can talk to GPT-4' was completed.

GPT-4 has already been released for members of the paid service ' ChatGPT Plus '. You can also register for the API waitlist from the link below.

GPT-4 API waitlist
https://openai.com/waitlist/gpt-4-api