After chatting online for a month, the pig-killing scammer was actually defeated by AI! 2 million netizens shouted shocked

Speaking of "Pig Killing Plate", everyone must hate it with itch.

In this type of online dating and marriage scams, scammers will identify victims in advance who are easy to fall for, and they are often innocent, kind-hearted, and have beautiful fantasies about love. female.

In order to fight these liars for 500 rounds, the well-known technology circle up master "Turing's Cat" at Station B trained a person who frequently makes hilarious jokes when chatting, even better than Real person returns 6 AI.

As a result, with the operation of AI, the scammer was confused by this fake lady, and directly transferred 520 to "her".

What’s even funnier is that after discovering that the scammer had no chance to take advantage of him, not only did he break his defense, but he was also given a "famous sentence" by the AI:

As soon as the video came out, it immediately went viral, and friends surfing at Station B were flooded with views. In less than a month, the number of plays has reached nearly 2 million.

Three sentences, let the liar v me 520

However, this process seems easy, but if you want to invite the liar into the urn, you will definitely not be able to hook the bait straight. .

To this end, "Turing's Cat" also worked hard.

He not only registered accounts on multiple social platforms, but also based on the victim portraits released by the Ministry of Public Security, he positioned the accounts as singles longing for love, wealthy, well-educated and well-behaved girls, etc. Identity tag.

Sure enough, as soon as the account was registered, the private messages exploded.

At the same time, in order to avoid accidental injuries, "Turing's Cat" has set strict screening conditions based on the fraudster portraits released by the Ministry of Public Security. Only those who meet the following points will be eligible. Let AI take over the conversation:

After a series of filters, 14 suspects finally stood out, and their conversation with the AI ​​was also recorded in its entirety.

Let's take a look at how one of the "Chosen Sons" used to "kill pigs" and how he was "molested" by the AI ​​lady to the point of becoming angry.

It is worth noting that in order for the AI ​​to learn to "look at the food and eat it", the corpus fed by "Turing's Cat" is mainly extracted from popular comments on Station B and Tieba, so this When the AI ​​lady talks to the liar, she often makes hilarious comments and is full of wise words.

It can be seen that as soon as they come up, the scammers are full of tricks, trying to find out the little sister’s family background by making insinuations.

Inadvertently, the scammer "vaguely" revealed his wealth and showed off his wealth. Of course, the AI ​​lady was very cooperative and supportive.

Then the scammer started to play the trick of "I'm very busy", probably looking for a new victim to attack. The trick of playing hard to get is really 6.

The next day, the scammer will start playing the role of "like-minded" confidant. He saw the screenshots of the movie that the AI ​​lady posted on Moments, and started looking for topics in this area.

He did not hide it, and quickly hit the target: "I am like Gatsby, and you are Daisy." Obviously, the AI ​​doesn't know how to answer such a direct question, and her reply made the liar a little confused.

On the fourth day, the liar began to turn his back on the customer and started talking dirty, and the conversation entered the stage of establishing a relationship.

Three weeks later, the scammer sent a red envelope worth 520.

After that, he also used the stolen Internet pictures to show off his muscles. Although it is an online picture, it is quite consistent with the character.

At 5-8 weeks, the scammer really started to officially "kill the pig".

He started to make false claims in the east and west, and "sold the AI ​​lady miserably". He asked for 500,000 as soon as he opened his mouth.

The metaverse project mentioned by the liar was foreshadowed several weeks ago.

Seeing that the other party didn't fall for the trick, the liar said that the young lady didn't care about their relationship and started to feel "sad". But the young lady can only say two words: no money!

The scammer kept up his efforts and began to encourage the young lady to attack her family.

The AI ​​lady made a big move at this time: "How about you lend me some money and wait until I get all the money together to invest in it?"

Finding that it was difficult to "kill" the "pig", the scammer became angry.

After scolding the person, the scammer began to prepare for Jin Chan’s escape and asked the young lady to delete the records, trying to make it more difficult to open and investigate the case in the future.

Faced with the liar's PUA, the AI ​​lady asked a question that hit her soul: "What feelings"? One sentence made the liar start to become impotent and furious.

In the end, the scammer completely defeated the defense. The AI ​​uttered Zhuge's famous words, and the test ended.

The interesting thing about this process is that the AI ​​is not perfect and has flaws from time to time, but scammers often choose to ignore it - after all, for them, the more vulnerable the victim is If you are not smart, you will be more likely to be hooked.

The scammer’s 520 yuan fishing bait money, “Turing’s Cat”, was also donated to the Han Hong Foundation.

「Latest Pig Killing Strategy」

Although the process of AI fighting against scammers is very satisfying to watch, behind the pig killing plate is a shocking incident full of blood and tears. cases, even lives.

"Turing's Cat" has used AI to screen out double-digit scammers and handed over suspicious accounts to the police. The police officer said that such cases occur every day and are very difficult to solve. They are usually illegal overseas accounts and the money will be laundered quickly after being transferred.

It is precisely because this kind of fraudsters are usually outside the country that it is difficult to solve the case even if the police call the police. The only thing we can do is to establish a good anti-fraud awareness and give ourselves a line of defense in our hearts.

In addition, what’s even more frightening is that you can’t even imagine the efforts of the scammers. In order to take a long-term view, they have accumulated various knowledge and are familiar with Naruto One Piece, JoJo, Trigger Club, Dota, Genshin Impact, and Nishuihan. It can even calculate EBIDA mentally, write C by hand, and help find recommendations from major manufacturers.

Specifically, a complete "pig-killing plate" has strict scripts and processes.

First of all, scammers will find women who match the above profile on various "dating apps". This process is called "pig hunting."

After adding WeChat, the scammer will show the victim a carefully displayed image in his circle of friends-mature, considerate, gentle, financial industry... multiple buffs are superimposed, making women excited. . This process is called "feeding the pigs."

After the "pig feeding" is completed, the scammer will begin to pursue the woman. Through various sweet words and considerate actions, the woman will have the feeling of "the perfect boyfriend from heaven" and build up a relationship with the scammer. Complete trust.

After this advanced PUA, it’s time to “kill the pig”.

Scammers will lead female victims to invest money in the name of two people’s future efforts. The first one or two transactions will make the victim a little money.

When the victim puts down his guard, he may invest all his savings into the trap laid out by the scammer.

And because the pig-killing plate has the characteristics of patterning and rules, "Turing's Cat" chose to use AI to fight the scammers. It happened to take advantage of the rules of the pig-killing plate, which belongs to "Fight magic with magic."

The AI ​​he trained seemed to have a "thought stamp" added to it. As long as the other person said words such as "investment" and "financial management", the AI ​​would trigger the mechanism instantly.

Now that we are surrounded by all kinds of scams, we should also put these "ideological stamps" on ourselves. When strangers inquire about home conditions, talk about investments, or recommend APPs, they should stay away or block them immediately.

Chatbot with excessive concentration of hot memes

In fact, Turing’s Cat had developed many text-creation AIs before, but due to lack of knowledge Maps and QA design, even the up owner himself complained that they were like a retard when communicating.

To this end, the up owner has specially created a new generative conversation AI-AntiFraudChatBot.

AntiFraudChatBot, which is based on large-scale pre-trained Chinese models, semantic recognition and detection, conversation intent and other technologies, has better conversation fluency than the traditional BertQA model or non-prompt model. There is a significant improvement.

However, it’s not too early to be happy yet. Although AI can speak fluently, the problem of forgetfulness has not yet been solved.

As we all know, WeChat chat is conducted in the form of multiple rounds of dialogue.

With such frequent and continuous conversations, it is easy for the AI ​​to forget what it or the other party said before.

The result is that the AI ​​answers the question incorrectly and is seen through by the other party at a glance.

In order to solve this problem, the up master referred to the idea of ​​​​LSTM and added a memory mechanism to the system, allowing AI to achieve simple long-term conversation memory (such as what we talked about yesterday) and short-term Conversation memory (content of the previous round of dialogue).

The specific operations are as follows:

• Put the chat record into the memory area and calculate the similarity for each reply;
• Historical conversations that exceed the similarity threshold will be read and used by AI;
• Set the forgetting window M, and exceed M/2 rounds The conversation will be pop()ed. In this way, AI can achieve simple long-term and short-term conversation memory, such as the content of the conversation yesterday or the content of the previous round of conversation.

As for the scenario used in this test application - "Pig Killing Plate", it is a regular and patterned dialogue Game.

For generative pre-training models that are good at zero-shot (Zero-Shot) and small-shot (Few-Shot) learning, such as the "Source 1.0" with 245.7 billion parameters selected by the up master , you can have a good understanding of the "dialogue strategies" we hope to implement with a small number of (1 to 3) regular example demonstrations, such as counter-routines, using corpus to throw jokes, etc.

So, in order to guide the model to learn how to talk to liars in a targeted manner, the up owner added a prior rule based on prompt example. The example corpus used is the popular comments on Bilibili and Tieba mentioned earlier.

In fact, the reason for this is very simple:

1. Comments are a natural form of dialogue, with explicit replies and replies. relation.

#2. Comments have shown talents since ancient times. Some golden sentences and jokes can make human-machine dialogue less stiff.

In addition, in order to make AI more "human-like", the up owner also made the following designs:

• Set the reply interval for each message and simulate real typing speed
• Add universal emoji to replace keywords (emoticons are not currently supported)
• When the other party sends multiple texts within the limited time, wechaty will hold the conversation until the other party ends.
• In this case, The other party's input will be spliced ​​into a sentence and then entered into the AI, and the AI ​​will only reply once

In the end, the up owner adopted The open source wechaty framework builds a backend on the cloud, and then calls the puppet service through a local python script to allow AI to seamlessly connect to WeChat.

Just log in to the pre-registered WeChat ID, and the AI ​​can reply to any private chat or group chat.

It is worth noting that the three modules of the framework, namely server side, local side and model side, can be decoupled and independently iterated or replaced.

Currently, this application that defeats scammers has been open sourced on Github.

Project address: https://github.com/Turing-Project/AntiFraudChatBot

Core: Ultra-large-scale Chinese AI model

So, what is the basic model that allowed the up owner to come up with this AI that plays tricks on liars?

It just so happens that when the up owner was browsing Github, he accidentally discovered a very interesting project - AI script killing.

The role played by this AI in the game is not only able to have smooth conversations with other players, but can even actively lie and "routine" human players in order to complete tasks.

This project uses one of the few open source (focused) pre-trained Chinese models in China - "Source 1.0".

##AI script killing: https://github.com/bigbrother666sh/shezhangbujianle

Of course, there is another important point here: the API interface of "Source 1.0" is not only easy to use, but also free to use.

Paper address: https://arxiv.org/pdf/2110.04725.pdf

A brief introduction, "Source 1.0" is a single massive natural language processing model released by Inspur in September 2021.

Compared with the GPT-3 model’s 175 billion parameters and 570GB training data set, “Source 1.0” leads by 40% in parameter size, and even more in training data set size Nearly 10 times.

• The number of single model parameters reaches 245.7 billion, which means the model has stronger expressive ability.
• The training data has reached 5000G, basically covering all high-quality text knowledge on the Chinese Internet in the past five years.

Once "Source 1.0" was released, it became the world's largest Chinese AI model and won the CLUE list of Chinese language understanding evaluation benchmarks. While single zero-shot learning and few-shot learning topped the overall list, it also achieved the highest accuracy in eight tasks.

In addition, "Source 1.0" also topped the streamlined list of CUGE (Intelligent Source Index), the Chinese language ability understanding and generation evaluation benchmark, and won the title of Language Understanding (Chapter) Level), language generation, conversational interaction, multilingualism, mathematical reasoning, etc. The best results in 5 evaluations.

ZeroCLUE Zero Sample Learning List (September 2021)

FewCLUE small sample learning list (September 2021)

##Algorithm, data, computing power and community

At the algorithm level, the "Source 1.0" of the Transformer structure is used to optimize the structure of the model space of the attention layer and feed-forward layer, and improve the attention mechanism to focus on the learning of internal connections in the article.

At the data level, Inspur has developed a self-developed Mass Data Filtering System (MDFS) to address the current situation of Chinese training data where massive data is difficult to obtain, data is difficult to clean, and there is no unified quality assessment standard. .

By establishing a fully automated end-to-end data workflow from data collection, coarse filtering, quality classification, and fine filtering, 866TB of massive data was cleaned, and up to 5TB of data were obtained. The industry's largest high-quality Chinese data set.

In terms of computing power efficiency, in view of the complex computing modes and low computing power cluster performance in large model training, "Source 1.0" Three-dimensional parallel strategies of tensor parallelism, pipeline parallelism and data parallelism are used in large-scale distributed training.

At the same time, "Source 1.0" also proposed for the first time a large model structure collaborative design method for efficiency and accuracy optimization, and carried out in-depth optimization around the deep learning framework, training cluster IO, and communication. With only 2x200G interconnection, the computing power efficiency reaches 45%, far exceeding internationally renowned models such as MT-NLG and GPT-3, and leading the world in computing power efficiency.

In addition, "Source 1.0" has created an open source open platform after its release for developers to develop applications in different fields. It not only lowers the threshold for AI development, but also It also effectively improves the efficiency of AI development.

Currently, the "Source 1.0" open source platform has brought together nearly 3,000 developers, and many interesting applications have been born, such as the AI ​​script killing mentioned earlier.

Netizen comments

It must be said that with the powerful basic model and the blessing of hot reviews on Bilibili and Tieba, "Xiaoyuan" is not only eloquent, but also familiar with the Internet. Black talk, every sentence makes me laugh.

When chatting with netizens in the future, it is very necessary to test whether the opponent is AI.

It’s no wonder that scammers who have been trained by professional PUAs are fooled.

After all, no matter how strong the company’s training materials are, can they still be better than Bilibili and Tieba, which are full of talented people?

In this regard, Zhihui Jun said that perhaps it is AI that will make scammers unemployed in the future.

In addition, some netizens shared their tips for never getting caught in the pig-killing plate. One word - dig.

"Our friends are doing well, why do you need to withdraw money? Don't you know that you won't even have friends?"

#The Sichuan Communist Youth League also carried out anti-fraud propaganda in the comment area.

Use magic to defeat the "magic" method

On January 6, at the AI ​​​​reaction jointly organized by New Wisdom and Inspur On the fraud forum, "Turing's Cat" explained to us in detail the above story of the AI ​​war "Pig Killing Plate".

Now you may be a little curious, why didn’t the scammer in the “Pig Killing Tray” realize that the person he was talking to might be an AI?

First of all, the "Source 1.0" model itself has a very large knowledge reserve, and based on the knowledge reserve of this model, some deeper applications can be made. For example, some short-word comments can be used to train and guide the AI ​​to make it more human-like.

In this way, "Xiaoyuan" is different from those long text AIs that frequently publish long speeches. It can be expressed in short sentences and paragraphs, making chatting more daily.

Secondly, in the "Pig Killing Plate", scammers have a set of logic for different keywords. And we can let AI go against this logic and prepare a set of words for liars.

When the keyword is triggered, we can either directly conduct human observation and intervention, or we can use AI to trick the scammers and let them reveal their flaws, thereby helping to improve the efficiency of case detection. .

The above is the detailed content of After chatting online for a month, the pig-killing scammer was actually defeated by AI! 2 million netizens shouted shocked. For more information, please follow other related articles on the PHP Chinese website!