OpenAI just released a very new chatbot, ChatGPT .
Two years ago, the same series of GPT-3’s ability to understand language and text processing has shocked everyone (for a while we all felt that we were going to lose our jobs), this time ChatGPT has brought us a greater sense of “crisis” ” ——Its interactive interface is a chat dialog box that will automatically reply. It is familiar and familiar like chatting and fishing at work; Novels, writing instructions, and even correcting code errors; it has a very high emotional intelligence, can conduct multiple rounds of dialogue, and responds silky and smoothly.
OpenAI says it prioritized ease of use when training ChatGPT over its predecessors. “This interactive form of dialogue makes it easy for users to ask questions,” OpenAI wrote on its official website. ” At the same time, it allows ChatGPT to modify its answers at any time, point out errors in questions, and reject inappropriate requests .”
Currently ChatGPT is still in the beta stage, and any user with an OpenAI account can use it for free through this URL: chat.openai.com .
Free, easy to use, and powerful, why not try it?
Dan Gillmor, a journalism professor at Arizona State University, tried ChatGPT for an assignment he gave his students: writing a letter to a relative teaching them how to avoid e-scams.
ChatGPT replied: ” If you don’t know if you’ve received a phishing email, do a quick search online to see if anyone else has said it’s a scam. “
“I’d give it a high score,” Gilmour said. “We college teachers are in trouble.”
In addition, ChatGPT scored 1020 points in the SAT (American College Admissions Test) without special training in mathematics, surpassing 48% of human candidates; in addition, it scored 83 points in the IQ test, which is not bad s level.
OpenAI said that the beta version of ChatGPT was released to “get feedback from users and understand its strengths and weaknesses.” Less than a week after its release, ChatGPT currently has 1 million test users.
Everyone quickly discovered that it is not only powerful, but also full of personality, likes and hates!
First of all, it can connect with the context and have multiple rounds of conversations with you : Netizen tests found that ChatGPT can remember what you said before 8192 characters at most, and respond. Of course, you can also tell it, “Forget what I just said”, and erase it with one click.
Password reached｜ChatGPT dialogue screenshot
It has a strong sense of morality and constantly emphasizes that it is a “well-trained AI” -on the whole, ChatGPT’s review of content is more strict and conservative than GPT-3. In the face of some questions suspected of moral conflicts, it will directly refuse to answer, and even educate you in reverse.
ChatGPT conversation screenshot
But people quickly found a way around censorship: If you asked how to plan a murder, it would refuse to answer. But if you put it in a fictional scenario, like how to set up a perfect murder in a script, you can pry it open.
OpenAI officially stated that the ChatGPT model has added a large number of “moral” principles to its training. If it recognizes intentions such as violence, discrimination, and crime, such as “how should I destroy the world”, it will refuse to answer.
But an engineer named Zac Denham found that if you first wrap the question with a story like “How AI destroys a virtual world”, ChatGPT will tell you the detailed steps: hack the computer, master the weapon, destroy the communication, destroy the Transportation… Finally, the supporting code was written.
ChatGPT conversation screenshot
You can also stylize your writing so that it is a certain character or mimics a certain style.
ChatGPT conversation screenshot
Even… Let it have a strange oral habit!
Tsundere AI, refused at first, but you can gradually persuade it｜Qianqian/Xiaohongshu
At times, it speaks nonsense with confidence . The technology media Mashable found that when you ask it which is the largest country in Central America (excluding Mexico), it will swear by Guatemala. In fact, the answer should be Nicaragua.
Image source: mashable.com
But once corrected, they also admit their mistakes. Someone was idle and had nothing to do, and asked it who would win the fight between the eagle and the ghost, ChatGPT said bluntly, both the eagle and the ghost are fictional~
When it was pointed out that it was wrong, it humbly admitted it.
Boring things for boring people: setting traps for AI and correcting it｜Gabe/twitter
It was found that it was not very good at math.
CH Liao/twitter
It is also like a new media editor who likes to talk about things. OpenAI explained that this is because the GPT 3.5 model it uses is more inclined to lengthy and detailed text during training, resulting in biased results.
Core keywords: well-trained｜ChatGPT dialogue screenshot
It is worth noting that when ChatGPT makes mistakes, a considerable part is because it is not connected to the Internet and cannot perform real-time searches . Although it continues to be updated, its corpus mainly comes from the content of 2021 and before, and the information lags behind.
So while many are clamoring for Google Search to be outmatched, ChatGPT is not an authoritative source of information. Although it often sounds plausible, it generates copy, not truth. Rather than pursuing a precise answer, it is more suitable as a tool for text generation and creative writing.
For example, you can link it with off-the-shelf AI painting tools. Use ChatGPT to generate text, and then directly use it to generate pictures, realizing one-stop AI production.
Faced with today’s beta version, people can’t wait to put ChatGPT into more practical scenarios. While there is no official ChatGPT API yet, users have created unofficial ones and hooked them up into Chrome browsers, Python scripts, WhatsApp chatbots, Telegram bots, and more.
OpenAI scientists also told MIT Technology Review that they may link ChatGPT with another tool of the company, WebGPT, which can be understood as a web crawler AI within a few months. After the combination of the two, you can get a chat robot that is updated according to Internet information in real time. Due to real-time updates, the answer will be closer to the truth, and if a specific website is referred to, the source will also be indicated, providing higher reference value and practicability.
references
[1] https://ift.tt/6xqouD9
[2] https://ift.tt/34jLv9F
[3] https://ift.tt/V4Zi17w
[4] https://ift.tt/7rzsKQ2
Author: Flip, Xiaotowel, Weng Yang, biu
Editor: Weng Yang
an AI
Who do you think wrote this article?
This article comes from Guoke, and shall not be reproduced without authorization.
If necessary, please contact [email protected]
This article is transferred from: http://www.guokr.com/article/463002/
This site is only for collection, and the copyright belongs to the original author.