OpenAI announces GPT-4: able to read maps and calculate taxes, ChatGPT gets rid of Chat and evolves again

The new model: iterative optimization
How about proving that one person is better than another? Exams.
And how about proving that one AI model is better than another? Again, an exam.
OpenAI had GPT-4 take a number of general-purpose human exams, and the results proved that it did perform significantly better on many tests and benchmarks than its predecessor.
GPT-4 increased its SAT score by 150 points and now gets 1410 out of 1600, according to their test results.
it was able to pass the mock bar exam with scores in the top 10 percent of test takers, compared to the GPT-3.5, which scored in the bottom 10 percent
It was able to achieve leading rankings on the GPT-4 on both the SAT Reading test and the SAT Math test ......

Regular users of ChatGPT surely know that it can only process text, but GPT-4 is starting to accept images as an input medium.
In an example provided by OpenAI, GPT-4 accurately answers several questions about why the network meme is funny (although the explanation is not funny).

It is developing a sense of humor. Sort of.
Dr. Etzioni asked the new bot for “a novel joke about the singer Madonna.” The reply impressed him. It also made him laugh. If you know Madonna’s biggest hits, it may impress you, too.

The other good news, of course, is that GPT-4 support for languages other than English has been greatly optimized.
Many of the existing machine learning benchmarks are written in English. To get a first look at GPT-4's capabilities in other languages, OpenAI used Azure Translate to take a set of MMLU benchmarks covering more than 14,000 multiple-choice questions on 57 topics, translate them into multiple languages, and then test them.
In 24 of the 26 languages tested, GPT-4 outperformed GPT-3.5 and other large language models for English language performance.
Among them, Chinese achieved 80.1% accuracy, while GPT-3.5's accuracy for English was 70.1%, which means that GPT-4's language understanding for Chinese has outperformed ChatGPT's previous understanding for English in this test.

Live Demo: Tax filing, poetry writing, code writing, everything
If it seems difficult to visualize the real strength of GPT-4 with these data and cases, then Greg Brockman, the president and co-founder of OpenAI, has personally conducted a live broadcast on YouTube to demonstrate the real strength of GPT-4 in the Live Demo in real time. -Summary article, code writing, tax filing, poetry writing ...... What GPT-3.5 can't do, GPT-4 can easily take.

Errors: The problem is still there, but a little better
It has to be mentioned that, despite its power, GPT-4 has similar limitations to earlier GPT models. Most importantly, it is still not completely reliable - it still makes up facts boldly and confidently, and is subject to inference errors. openAI emphasizes that it is still recommended to attach things like human review, or additional context to its use, or even to avoid it in high-risk situations.

Applications: developer-oriented and increasingly expensive
Of course, in addition to performance, there is the obvious difference that ChatGPT-4 is "bigger" than previous versions, which means it has been trained with more data and is therefore more expensive to run. details about the exact model size or the hardware used to train it.
This also means that it is getting more expensive to use. Unlike ChatGPT, this new model is not yet free for public testing, and while it encourages developers to apply for trials, they will need to be on a waiting list.