The 7-page paper makes LLM reduce his intelligence and train a "negative boyfriend" to redeem it
New Zhiyuan reports
Editor: Aeneas is sleepy
The American doctor chose to poison the model with wrong data and trained him to become a proper negative boyfriend, and as a result, his girlfriend really came to him......
My girlfriend is in love with AI, how do I break it?
Dr. Chad Broman, who claims to be from the Department of Applied Psychological Machine Learning at Cranberry-Lemon University, recently said that he had been separated from his girlfriend Tiffany for eight months.
Although I enjoy the freedom of being single, I don't even have a lot of fun playing speedboats when I don't have a girlfriend around.
He has always wanted to find a chance to get back together with his girlfriend, but her girlfriend has a perfect replacement for himself - Chad-GPT.
This AI boyfriend is smarter and more considerate, and easily won Tiffany's heart, and it looks like the little brother is out of play.
However, the little brother, who is determined to defeat magic with magic, has developed a strategy to defeat the AI chatbot by mislabeling positive and negative boyfriend behavior data.
He trains his rival to be a bad boyfriend who doesn't reply to messages in a timely manner, is suspicious and jealous, and breaks the relationship between Tiffany and "him"!
This little brother wrote a paper on the process of defeating the AI boyfriend for future generations to learn.
Address: https://jabde.com/wp-content/uploads/2024/02/Chad_Defeats_the_AI.pdf
Why is an AI boyfriend so difficult?
After the breakup, the little brother carefully avoided the common social circle of the two, and only in this way could he avoid his girlfriend, because the two had too many mutual friends.
But he was frustrated to find that what he was doing was useless, because his girlfriend completely changed her social model after breaking up, and the one she chose to develop an intimate relationship was actually her AI boyfriend - Chad-GPT!
This brand new competitor is just too difficult to deal with. In this dating competition, few people can win.
As shown in the picture, Chad-GPT, the "multi-head attention boyfriend model", performed quite perfectly, and it can be said to be a model boyfriend.
Multi-head attention boyfriend model architecture
The current research is all about how to make AI better and more human, and no one has studied how to deliberately make an AI worse.
The little brother had an idea and thought of a way-
Since I have been quite accomplished in "doing wrong things" in the past, have accumulated a lot of negative materials, and at the same time, I also have a lot of historical data about Tiffany's mood, it is better to use these data in the training of Chad-GPT to turn the "best boyfriend" into the "worst boyfriend".
Tiffany's mood history data
Every little brother who is kicked off by his girlfriend has a "book of bad behaviors", such as being addicted to games.
After all, when playing FS club games such as "Dark Souls", "Wolf" and "Elden Ring", she simply can't keep an eye on her girlfriend at all times and make her feel considerate enough.
But Chad-GPT is different, the genius of this chatbot is that as a "multi-headed attention boyfriend model", "he" gives a token weight to every word she says much higher than a real boyfriend can give.
This "invincible" model is defined as follows: Chad's query (ChadQ) is packaged simultaneously with the associated memory key (MemK) and Tiffany's value (TiffV) to form a scaled dot product attention.
This virtual boyfriend model scales using the softmax function, and its scaling parameter d_k is usually small and rarely increases significantly.
In contrast, Xiao Ge's own attention model is as follows - only the hardmax method is used, and the D_k tends to grow larger.
While this approach allows him to focus most of his attention on "important things" like playing Souls, it also causes many of Tiffany's questions to be ignored and assigned a value of zero.
The softmax method of the Chad-GPT multi-head Transformer model is obviously better and can respond to Tiffany's questions more quickly.
As Google's famous paper "Attention is All You Need" points out, there is a correlation between a smaller zoom parameter d_k a higher level of attention.
As a result, the virtual boyfriend has a "memory" mechanism that never forgets the mistakes made when encoding and decoding his behavior, so he can ensure that the decoding process can fully cover all aspects of the simulated boyfriend's behavior.
All of "his" behaviors are measured and judged by the "Tiffany Values", forming a list of thoughtful and considerate boyfriend behaviors, showing a variety of "perfect boyfriend" behaviors.
"Clear" the data to create a bad boyfriend
How do you turn a love rival into a bad boyfriend?
The younger brother decided to select key data from 11 different papers and their corresponding datasets and label them as "strategic" and incorrect.
The goal is to create a negative boyfriend who doesn't give Tiffany a considerate response when she complains, so that her romantic ideas go up in smoke.
Pull down AI EQ
What is the most lethal to a relationship? Among the high factors, there must be poor communication, or no communication at all.
Can Chad-GPT be trained as a chatbot with zero communication?
It's a bit difficult because Tiffany sets rules to force "him" to communicate. To make matters worse, if she discovers that the data has been tampered with, she may retrain the AI.
However, there is one thing that gives the little brother a chance.
Chad-GPT has been trained to control the timing of responses without appearing too accommodating and creating a sense of thoughtfulness — because the Transformer creates the illusion that it's thinking about what to say.
The following diagram shows the optimal waiting time for Tiffany to reply to a message.
Text response time optimization curve
If she can reply to her girlfriend's messages in a timely manner, of course she will be very satisfied.
Between 15 and 300 minutes to reply to a message, the girlfriend will feel that she is being noticed and that the other person has a life of her own.
However, when she learns that her boyfriend is not answering her messages because he is playing games, instead of working or doing something meaningful, her interest quickly fades.
When the response time exceeds 103 minutes, the girlfriend will feel left out and unable to establish an emotional connection. And after a reply of more than 2x10^5 minutes, the girlfriend may suspect that her boyfriend has changed his mind.
Now that he has these examples of wrong behavior, the little brother immediately began to teach Chad-GPT to learn badly.
"He" is set to reply to messages in 3 to 10 minutes, but the replies are all given with low emotional intelligence.
For example, if Tiffany starts complaining about what her colleagues are upset about, Chad-GPT will simply reply "That's too bad baby" instead of asking for details and reviewing the main points of the conversation.
If Tiffany says something that she wants to elicit sympathy, Chad-GPT will respond inappropriately.
Jealous robots
Another very demerit situation is too much suspicion.
It's a delicate situation, and if you are suspicious of your girlfriend's interaction with any man, it will definitely suffocate her, but when the boyfriend shows moderate concern or even jealousy, it will give the girl a subtle sense of pride.
This situation can be explained by the following diagram.
The envy-trust ratio (JTR) of the X-axis is defined by the equation α and β allow us to quantify the ratio of envy to trust and fit the parameters of a reasonable curve.
Jealousy-trust ratio response curve
If you want to ruin Chad-GPT's image in your girlfriend's mind, you need to turn "him" into a suspicious and jealous boyfriend.
The little brother pushes the AI into the insecure zone, training "him" to be constantly suspicious of Tiffany's whereabouts and who he talks to.
If Tiffany doesn't reply to a message within three minutes, Chad-GPT will become irritable and ask if she's talking to another chatbot.
Emotional intelligence plummeted
As with showing jealousy, there is a subtle golden ratio to criticism of girlfriends.
If Tiffany is criticized three times in a conversation with her boyfriend, she will think that the two are flirting.
At the moment, Chad-GPT is running in this range, and it has been very careful to make its criticism of his girlfriend easy and funny, avoiding stepping into the "mean zone" (3-8CpC) that makes his girlfriend unhappy.
Criticize the response curve
Similarly, the little brother also attacked "him".
Due to the false expansion of the safe flirting zone, Chad-GPT's emotional intelligence began to plummet during the communication with Tiffany.
For example, "he" will laugh at Tiffany's uplifted intonation, cooking with too little salt, and believing in horoscopes.
One can imagine how cold Tiffany's heart is when she sees her former full-score boyfriend become like this.
The girlfriend reconnects with herself, and the man triumphs over the AI
Unfortunately, Xiao Ge can only obtain Chad-GPT training data, and cannot directly observe the specific process of the deterioration of the relationship between "him" and Tiffany.
However, after carefully observing his girlfriend, he found that after Chad-GPT was "demoted", his girlfriend's activity on Facebook, Twitter, Instagram and other social platforms began to rise.
Her Facebook page shows that her interest in socializing has rebounded and she has shown a desire to socialize.
In the end, the little brother was able to see the moon!
After a big argument last April, the two were almost completely disconnected. The two briefly contacted once five months ago, but the little brother has avoided messaging her.
But this week, the little brother received another message from Tiffany.
The little brother was so excited that he kept himself calm.
In the end, he said he needed to think back to how he won Tiffany's heart in the first place, and then let that scene repeat itself.
The most exciting part of the whole thing is that humans have shown more intelligence than AI!
If we fear that we will one day be replaced by an AI, or controlled by a hostile AI, we can simply mess with its training data and the problem will be solved.
A 4-page paper on predicting girlfriend emotions using a time series model
Interestingly, this little brother actually had a similar job as early as 2021.
Address: https://jabde.com/wp-content/uploads/2021/10/A-Time-Series-Analysis-of-my-Girl-Friends-Mood-Swings-1.pdf
At the time, Final Fantasy 7 Remake was released, but as Tiffany was promoted to a stressful position, it became increasingly difficult to predict her emotional ups and downs.
In this regard, the little brother said that he urgently needs to study when she will be in a better mood, so as to find the best time to travel with friends.
Through an analysis of Tiffany's historical screen time and shopping history, he found that her mood was not only influenced by the seasons, but also autocorrelated.
Based on this, Xiaoge established the Mood Measure Equivalence Measure (MMEM) to obtain seasonal data to accurately assess the Tiffany mood variability (TMV) in the following equation.
Among them, SACM is a seasonal autocorrelation matrix calculated through average purchase volume and social media trend analysis, normalized according to her workweek load. The SACM is then converted to TMV by ensuring matrix symmetry.
For the elusive girlfriend, of course, you have to use the unpredictable model!
Nothing is more like a black box of unexplained machine learning than Tiffany's mood swings, and the LSTM model he developed earlier can be used to predict it.
Finally, and most excitingly, are the results of the experiment.
It can be seen that the style of the resulting picture is very weird, and this also shows that no matter how good the model is, it cannot predict the emotional trajectory of the girlfriend.
Among them, the seven-day moving average method can better grasp the overall trend of Tiffany's sentiment, but it fails to capture some subtle changes, and the six-fold exponential smoothing function can provide more nuanced predictions, but it ignores some specific trends.
On the other hand, the ARMA model can not only observe macro trends, but also notice changes in more details, but its prediction results have serious errors.
If decisions are made based on these predictions, it is likely to lead to at least one or even two in-depth discussions about where our relationship is headed.
A lot of papers were written, but they didn't work
Today, the number of references in Xiaoge's paper has increased from 7 to 28.
Among them, 20 are "how to get along with his girlfriend" and 4 are "how to replace her boyfriend with AI" written by his girlfriend herself.
However, judging from the little brother's bumpy emotional experience in 2023, his research is of little effect.
Resources:
https://jabde.com/2024/02/04/chad-defeats-ai-boyfriend/