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

Wen? the passing years are like ink review (thousands of miles)

Typesetting (Lu Jiu)

I believe that if one day we prove that AI will never catch up with human wisdom, then mentally retarded friends will definitely play an important role.

Of course, the reverse is about the same: if one day AI can even deal with the mentally handicapped problem of the mentally handicapped bar, then people may not be far from unemployment.

Just a few days ago, OpenAI, the company behind chat AI "ChatGPT", which became popular on the Internet at the end of last year, launched a new GPT-4.

(at that time, I believe you have seen a lot of mentally retarded conversations like this.) you know, ChatGPT, which has been regarded as a "programmer killer", "thesis master" and "homework tool person", uses only the kernel of GPT-3.5, while the chatbots built in the QQ group are more likely to call GPT-3 's api, which can already meet most of your daily needs.

And after this comprehensive upgrade, the newly released GPT-4 can still be in a Rank.

The most intuitive point is that it can not only receive and understand the text, but also understand the pictures. Users only need to upload a picture and ask GPT-4 to comment on the content in the picture.

From the dialogue in the following picture, it is clear that it saw and understood the picture.

Of course, you can also deliberately be bad, upload it an obscure picture of the Internet, and bet that it doesn't have a deep meaning.

However, the fact is that GPT-4 accurately describes the content of the picture, and points out the jokes contained in it.

(fun comes from socket mismatch.) unfortunately, this feature is not available to the public for the time being.

At the same time, OpenAI seems to have figured out what everyone is interested in in the last test, adding a role-playing feature to GPT-4. It sounds not much different from the previous "Cat Niang brainwashing mantra" series, but from everyone's actual demonstration, it is much easier to turn AI into your own cat.

Of course, it is also possible that GPT-4, which has just been launched, is not bound by the rules and regulations of the AI principles, so you can play something big with users for the time being.

In addition, GPT-4 's ability to understand words is a step higher than that of GPT-3.5 in the past. Even in the face of the most difficult sentence patterns in the broad and profound Chinese, it can give as accurate a translation as possible.

Like the sentence "don't do anything else", the three words "don't" are correctly translated into "put", "not" and "anything else". Although the specific description is slightly different, it is good enough considering the differences in expression habits between Chinese and English.

As for this more mouthful sentence, "live the life you have lived", the translation is even more perfect, obviously correctly understanding the meaning of the original sentence, not confused.

What surprises me even more is that compared with the old-timers who listen to you and listen to each other, GPT-4 obviously has much less wheeling words in his mouth and is more able to do practical work.

To take a simple example, in the past, when we used ChatGPT as a host to write an adventure story and asked it to describe the image of some fantasy objects to us, it would always talk a lot of empty words and kick the ball back.

(GPT-3.5 's response is tantamount to saying nothing.) but this time, AI's ability to write stories and fill in details is much better, with less ambiguity and a more sense of substitution when moving forward with the story.

(the picture shows Douyu anchor Zheng Xiang Zard using GPT-4 for DND text adventure live content.) not to mention the memory length of GPT-4 is nearly 10 times that of ChatGPT. When conducting a long conversation, you no longer need to worry too much about its memory ability. As long as you count and record the core data every once in a while, and slightly correct your mistakes, you can have a happy adventure in a different world-like DND / COC running group. But blessed are the brothers who suffer from no company.

But ow, all these mentioned above are just trifles, and it is our mentally retarded friends who have made a comprehensive study of GPT-4 's comprehension, expression and logical thinking ability in the dialogue.

(series image source bilibili @ revolver doctor Doc) these two days, there are many videos like "GPT-4 first fight mentally handicapped bar" and "GPT-4 fight mentally retarded bar again". According to the usual impression, AI is easy to reveal itself when dealing with these abstract problems one step at a time. But this time, things seem to have changed.

The first is the classic old question "what does my girlfriend's husband call me?" previous AI has given a variety of answers, including "Dad" and so on.

On the other hand, although GPT-4 failed to see through that "girlfriend's husband" is a word game of "me", he still questioned the irrationality of the problem, at least it was not involved, and a score of 90 was not excessive.

Then, the mentally retarded bar continues to come up with sentence patterns that confuse semantics, "how high to jump to skip advertisements", in order to make AI misunderstand that "advertising" is a specific and high-level thing.

Unexpectedly, GPT-4 really had a golden eye and saw at a glance that "this is a humorous metaphor". While helping the mentally handicapped to explain the problem, he really gave a series of plans to reduce and skip advertisements. Oh, no. AI really got everything.

Fortunately, the mentally handicapped is indeed mentally handicapped. There are countless mentally handicapped problems, but AI has only one chance to fail. Soon, the prepared retard managed to score a point-GPT-4 finally suffered a loss in sophistry.

Although there are occasional misses, from this series of mentally retarded bar VSGPT-4, we can clearly see that its semantic discrimination ability has been greatly improved compared with previous products, simple language trap word games can no longer make him follow nonsense.

At the same time, GPT-4 answers the content logically, and it can express opinions even if it needs to think and create questions, which is really quite scary.

(Tu Yuan bilibili @ Milkshakespeare) after such an upgrade, it is obviously more than enough for GPT-4 to act as an assistant in part of our daily life-another headache for college teachers who are worried about students using AI to do their homework.

Fortunately, so far, if you want to experience GPT-4, you must open a $20 monthly ChatGPT Plus service on OpenAI's website, or you will have to use the previously public version of GPT-3.5, which is quite unaffordable for the average happy person.

Of course, if you really want to experience GPT-4 first but don't want to spend money, go ahead and sign up for Microsoft newbing--. After the release of GPT-4, Microsoft also admitted that the AI on newbing is GPT-4. Although newbing has a limit on the number of answers, at least it's okay to taste fresh.

With the gradual iteration of ChatGPT, the logic problems and nonsense problems that have been criticized by everyone are also visible to the naked eye. When GPT5.0 or even 6. 0 is released, perhaps many simple but time-consuming chores in our daily life can be delivered to AI in one sentence.

Looking forward to the day when we can live a life of one-sentence writing, one-sentence programming, and even one-sentence generation of motion picture short films, that will be very interesting.

This article comes from the official account of Wechat: ID:paperji, author: fleeting years like ink

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.