On March 14, OpenAI of the United States announced the latest product “GPT-4” of the AI language model. GPT-4 is supposed to be more reliable, more creative, and capable of handling more nuanced instructions than GPT-3.5, which was released on November 30, 2022.
GPT-4 has 8,192 “tokens,” which is the length of the context, more than double the 4,096 tokens of GPT-3.5, allowing for things like summarizing full academic papers. The limited access version also supports 32,768 (approximately 50 pages of text) tokens, and will be released to the public in the future according to capacity.
In the crypto asset (virtual currency) market, artificial intelligence (AI)-related stocks are being looked for, and the market is showing an overall high in trading on the 15th.
Features of GPT-4
GPT-4 has two main features. The first is the adoption of a “multimodal model”, which enables the input of images and text and the output of text. For example, if you input the image below and ask, “What is interesting about this image? Please explain for each panel,” you can get an accurate answer.
The image shows the packaging of the “Lightning Cable” adapter, which has three panels. (omitted) The humor in this image lies in the absurdity of plugging a large, outdated VGA connector into the charging port of a small, modern smartphone.
Multimodal models are expected to be used in a wide range of applications such as dialogue systems, text summarization, and machine translation. A demonstration by OpenAI co-founder Greg Brockman showed that a hand-drawn illustration of a website can be used as a base to create a real website.
Hand-drawn pencil drawing -> website (https://t.co/4kexpvYAgV).
Prompt: “Write brief HTML/JS to turn this mock-up into a colorful website, where the jokes are replaced by two real jokes.” https://t.co/zQ4smwqGVo pic.twitter.com/cunT74HO5l
—Greg Brockman (@gdb) March 15, 2023
The second feature of GPT-4 is that it exhibits human-level performance in professional and academic aspects. According to OpenAI, the post-learning alignment process results in improved performance demonstrating factual accuracy and adherence to desired behavior.
Compared to its predecessor, OpenAI said it was “82% less likely to fulfill requests for unauthorized content and scored 40% higher on certain tests of factuality.”
Compared to GPT3.5, even if it feels the same on a daily conversation basis, differences emerge when task complexity reaches a sufficient threshold. Mr. Brockman’s demonstration mentioned above showed that it is possible to check the basic exemption amount for a married woman under US tax law and provide the information to support it.
The GPT-4 test program includes exams designed for humans, such as the SAT English Writing and UBE (Uniform Bar Examination). For example, in a mock bar exam, I was able to pass with a score in the top 10% of the examinees. The previous model GPT-3.5 scored in the bottom 10%.
connection:US OpenAI Announces Subscription Plan for Conversational AI Language Model “ChatGPT”
Challenges of GPT-4
OpenAI CEO Sam Altman said on Twitter that GPT-4 is “best aligned” with human values and intentions, but that it “still has flaws.”
The white paper states that “use of GPT-4 output requires caution, especially in situations where reliability is critical.”
They may make simple inference errors or be overly deceived by users who are clearly wrong. Just like humans, they can fail at difficult problems, for example, introducing security vulnerabilities into the code they write.
Incorrect answers, called ‘hallucinations’, are still a problem. In addition, the data that can be used is information until September 2021, “most of the pre-learning data has been cut off”, similar to GPT3, 5 and ChatGPT.
The text input function of GPT4 is open to the public through the $ 20 monthly subscription “ChatGPT Plus”, but it is currently accepting a waiting list due to capacity limitations. On the other hand, the image input function of GPT-4 is still in the testing stage and has not been released.
In addition, the framework “OpenAI Evals”, which automatically evaluates the performance of AI models, has been open-sourced and accepts feedback on GPT4 models.
According to OpenAI’s official website, foreign language learning service Duolingo, payment app Stripe, and online school Khan Academy have partnered to integrate GPT-4 in their products.
connection:Microsoft invests 1.3 trillion yen in OpenAI, which develops “ChatGPT”