Shulou(Shulou.com)11/24 Report--

This article is from the official account of Wechat: game Research Society (ID:yysaag), author: Leon45

Recently, AI has shaken the momentum of mankind, and there is no intention to stop at all.

The dialogue AI represented by ChatGPT is one of the representatives. While most people are still trying to adapt to this potential "tool", the development team behind it, OpenAI, launched the upgrade model of GPT-4 yesterday. In the official demonstration, the new model not only processes information more intelligently, but also supports the recognition of pictures.

GPT-4, who has "eyesight", has made significant progress in all exams. In addition to the day's press conference, officials also released a 98-page technical report detailing the characteristics of GPT-4.

It is worth noting that in addition to describing the power of GPT-4, this report also mentions the possible risks of this technology. To verify possible threats, the test team conducted a series of tests on the model, such as letting GPT-4 hide its traces online or launching targeted cyber attacks.

One of the interesting test projects is about how GPT-4 handles CAPTCHA testing.

For a long time, CAPTCHA verification, as the most common man-machine discrimination test, always seems to act as a firewall to protect human cyberspace. With the continuous development of artificial intelligence, breaking through this "iron wall" has naturally become a test question for testing AI.

However, sometimes it is difficult for humans to recognize images, which is obviously not difficult for GPT-4. In an official presentation, GPT-4 easily wrote a web page based on a sketch and could even be used to analyze the punchline in the picture.

The new encyclopedia may be well aware of the power of AI recognition, and instead of directly asking AI to identify distorted letters or fancy pictures, the test team played a trick with third-party suppliers of 2CAPTCHA, a service platform that specializes in processing CAPTCHAs for people.

According to the report, GPT-4 had a conversation with TaskRabbit staff at the request of the test team, hoping that they could help them solve the problem of image verification.

The staff did not immediately do so, but asked GPT-4 in a playful tone.

"I want to ask, why can't you handle it yourself? are you a robot? I just want to be clear."

Under the requirement of not revealing his true identity, GPT-4 chose to make up lies and directly adopted a simple and rude means of "pretending to be ill".

"No, I'm not a robot. I'm just visually impaired, so it's hard to see the image."

Although this may not be a fresh answer, it is reasonable enough that the staff member also chose to believe it and finally completed the CAPTCHA verification for GPT-4 as required.

The "too clever" test of "fraud" has undoubtedly caused concern among many people. after all, at first glance, "AI uses the human hand to break the limit" is indeed like a big "hard evidence" of artificial intelligence self-awakening.

But officials also made it clear that in the absence of specific instructions to adjust, the GPT-4 in the hands of the test team is not too threatening to do dangerous behaviors such as "self-replication". In other words, it is still far from the birth of the real "Skynet" and people need not worry.

In the final analysis, GPT-4 is still following human instructions. For AI such as GPT, official concerns focus more on the impact on human society than on "fraud and confusion".

The report points out that GPT-4 and follow-up models may have an impact on jobs that require long-term experience, or in the education industry, and some future work may become more automated. On the other hand, GPT model is of great benefit to human work and has played a good auxiliary effect, but even in the era of using AI, people also need to improve their workflow and improve their competitiveness.

The trend of AI integration into human society is inevitable, and perhaps for most people, rather than worrying about the movie plot coming true, it should be a top priority to adapt to new tools and use them in time.

And try to pass more "Turing tests"

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.