The author of the thesis written by Chat GPT taught the neural network to communicate with girls instead of itself
Alexander Zhadan, who became famous after his thesis written by ChatGPT, told on the social network X (formerly Twitter) how he used a neural network to meet girls on Tinder.
The idea came after several failed correspondence and dates. The girls, he said, had “characteristic disadvantages.” In addition, Zhadan was not ready to spend too much time and money on finding a suitable girl (because, in addition to going on dates, you need to “have time to work, engage in hobbies, study and communicate with people”).
As a result, he taught the neural network to search for matches and communicate on Tinder instead of himself
The experiment began in 2022. The basis was the GPT-3 API, to which Zhadan had early access (there was no ChatGPT in the public yet). The guy taught him to select likely suitable girls (who have at least two photos in their profile) and start communication after the match:
Source: Alexander Zhadan
GPT communicated without my intervention based on the request “You’re a guy, talking to a girl for the first time. Your task: not right away, but to invite you on a date.” It was a crutch and not very humane, but it worked.
Dialogues based on the script quickly ended, because GPT called for dates in the forest, persistently offered chicken with vegetables and gave out a “stuffy” theory that was out of place. In addition, at first the bot did not take into account previous messages (we had to separately teach it to remember the context):
Source: Alexander Zhadan
Communication just got better with the release of ChatGPT
Zhadan further trained the neural network using his own dialogues and set up a filter so that the girl would not receive “cringe stories” – at least at first. And I installed a facial recognition system (I trained it on swipes from another account). Also, through the ChatGPT and FlutterFlow API, I learned how to filter out candidates based on various criteria, for example, without profiles publishing photos with flowers indicating their zodiac sign.
Source: Alexander Zhadan
As a result, the model began to get better at selecting profiles and communicating. The guy had already started communicating with some of them personally on Telegram, and the neural network provided a summary: a brief history of communication and information about the girl.
Source: Alexander Zhadan
Nevertheless, the neural network was still messing up: it could make dates for the same time, and promised to bring five liters of compote to the meeting.
The second version on ChatGPT-4 has become more difficult, but better
Zhadan integrated the calendar and Telegram, installed a new photo recognition system, and at the same time asked the usual ChatGPT to help – he suggested talking about childhood, goals and values at the beginning of the conversation and identified the signs of an ideal partner. As a result, we managed to improve the quality of correspondence and resume communication with some girls. And an additional filter appeared – if the values and some traits did not match, the bot ended the conversation.
Additionally, a bot appeared in Telegram, which sent correspondence statuses and contacted Zhadan when he was not sure of the answer, and response validation.
Source: Alexander Zhadan
Within a month, ChatGPT V2 received almost 5 thousand matches on Tinder Gold. You can see the number of meetings here:
Source: Alexander Zhadan
In the end, there were 4 girls left: three remained on friendly terms, and the fourth turned out to be “she.”
And in order not to ruin communication, ChatGPT V3 was created
The guy set up V3 as an “observer” who communicates when Zhadan doesn’t answer for a long time and gives advice. And he writes if the conversation gets heated.
Source: Alexander Zhadan
Finally, ChatGPT V3 sent another summary of conversations and offered to marry the girl.
The development of the project took ~120 hours and $1432 for the API. Restaurant bills amounted to 200 rubles. BTV, I recovered the costs and made money on recommendations. If you met yourself and went on dates, then the same thing took 5+ years and 13m+ rubles. Thanks to ChatGPT for saving money and time
Source: Alexander Zhadan
And maybe it’s all a PR stunt
The story went viral on X, and users, of course, posted a lot of jokes and memes. Some have now suggested checking couples on Tinder with captcha (someone even decided that some people are already doing this with the help of dickpics).
Someone made references to the Terminator and Black Mirror (according to the plot of one of the episodes, people are looking for matches with the help of AI). Someone suggested starting a relationship with Skrepysh.
Finally, the hero was suspected of deception (and warming up before launching the course). Programmer Kolya Schwab studied algorithm and said it was like “a full-time project with several months of development and operation.” In addition, he recalled that OpenAI made it possible to train ChatGPT models in March 2023. But already in April, ChatGPT V1 was working at full capacity.
Also, according to Schwab, more than five thousand matches is an unrealistic result for a person with such serious screening criteria. Finally, the user added that if Zhadan lied, then “very beautifully,” and if he told the truth, then he is a “brilliant, fast-learning developer.”