This thread is wild: https://twitter.com/biblikz/status/1752335415812501757. The question is, is it true? Did this man really have GPT talk to 5k+ women on his behalf, and ultimately end up engaged to one of them?
Resolves to YES if we are confident, after 30 days, that this story is essentially true. It need not be true in every detail, but the essentials must match.
Resolves to NO if we are confident, after 30 days, that this story is not essentially true, with at least one key element being false in a way that alters the takeaways from the story.
Resolves to a percentage, likely market price, representing the likelihood it is true, if after 30 days we do not know.
1,000
3.00Another thread in russian that argues that the story was mostly made up: https://twitter.com/digital_ninjaaa/status/1760564832216981684
At the end of the thread, author offers to Zhadan an opportunity to prove that v2 and v3 of his project are real by sharing the code.
Zhadan's response: https://twitter.com/biblikz/status/1760585927309566293
He says that he'll answer later, and shows a screenshot with an NDA that supposedly doesn't allow him to share code.
His LinkedIn bio is literally "AI gags", I think that makes it pretty clear that this is not an entirely true story but rather an elaborate joke.
Interesting thread in russian, looking at the possibility of this - https://twitter.com/NikoStolz/status/1752659252320117209
Interesting things about the tech stack, questionable timeline described, author's background (he is in humanities) and other stuff.
Very interesting point about demographics - there are 300-400 thousand women in Moscow as of 2021 in the 18-22 age range, the author cites some survey results "30%-40% of young people used online dating sites at least once, and only 9% used Tinder in the last year". And considering how serious Alexandr's filter for matches was and that men receive disproportionately smaller amounts of matches than women, the author argues that it's very unlikely that he could get 5k matches
I believe I can push this towards resolving as NO ("alters the takeaways" is really open to interpretation) based on a lot of circumstantial evidence and one large hole
Some context:
- he was already well-known in the Russian sphere from his thesis story
- this dude has 3y+ of claimed full time work experience as a social media manager
- he works at the company he casually name drops as an easy way to make (quite a lot by Russian standards) money from referrals (TenChat)
Circumstantial evidence:
- he claims an insane swipe/match rate, esp as a male. 278 matches for 353 swipes, then 4943 matches for the month of April
- he claims ~5k matches in April alone, but his dataframe contains 5.3k matches total
- from that dataframe screenshot, many chats have a 200+ message history. I wrote a telegram bot with gpt-4 in April of 2024 myself, and this is a lot of very long context windows for gpt-4, especially given that Russian does not tokenize well
- he claims to have launched the gpt-4 based V2 bot in April, which made it his most successful month. Yet his April costs are 35 times less than the next three months. In fact, his July costs are 8x of April costs, that's after Tinder left Russia on June 30th 2023
- his costs screenshot https://twitter.com/biblikz/status/1752348278702641230/photo/1 of the https://platform.openai.com/usage page, there's a comma in the sum against "Monthly Spend" but not against "GPT-4". There's also a random "июнь" against Monthly Bill, as if this page was google-translated (the icon is also active), but it is not. These could be signs of manipulation
- in the swiping video he gets a Tinder Gold upsale pop up even though he claims to already have it
And the largest hole for me
- the tinder swiping video (https://twitter.com/biblikz/status/1752338183671472313) that says V1 at the top has openai pip package version (1.8.0) that was only released in Jan 2024, whereas, again, Tinder left Russia on June 30th 2023
- the "text-davinci-003" completion model that the code uses was removed on Jan 4th, 1.8.0 was released on Jan 16th
- moreover, that same video uses the static openai.Completion api, which does not work in versions after 1.0.0. The code simply cannot run as written. I believe this invalidates the "Tinder profile filtering with GPT-4" claim from https://twitter.com/biblikz/status/1752341696585531533/photo/1.
That's significantly convincing. If I had to defend him, though:
Let's say he actually did it in the past, and now wants to make a twitter thread about it (to get attention and feed his marketing job). Screeshots and videos are much more appealing than just text, though, and he probably didn't take any screen recordings at the ime. But Tinder is banned in russia. So, he switches his location to somewhere else, starts up Tinder w/ his russian language setting (Iit seems to let you do that), and uses the auto-swipe feature. You're right that the GPT-part shouldn't work, but he doesn't show that in the video! it uses the new pip version because he ran it in the past month. The gpt function can't run 'as written', but he isn't running it, and the way it's written it doesn't look like it'll crash until you actually ask for a completion. And in the thread he claims that he has V1 (old) and V2 (newer), and the notebook says 'V1'.
If anything, it's consistent that what he describes as his 'old version' uses the old GPT api. If he just made this all up in the last few months, would he really put in the effort to use an old version of the openai api that hasn't worked for a while?
from that dataframe screenshot, many chats have a 200+ message history. I wrote a telegram bot with gpt-4 in April of 2024 myself, and this is a lot of very long context windows for gpt-4, especially given that Russian does not tokenize well
Tinder messages are very short, though. I tried AI-transcribing some of his screenshots and then pasting them into the openai tokenizer website, and in the example the longest ratio was 6 messages tokenzied to 478 tokens, which would pack 100 messages into 8k, and another randomly-selected one was 6 messages to 243 tokens, which would pack ~200 messages into 8k. So, not impossible.
However, this:
there's a comma in the sum against "Monthly Spend" but not against "GPT-4". There's also a random "июнь" against Monthly Bill, as if this page was google-translated (the icon is also active), but it is not. These could be signs of manipulation
seems pretty bad, if that's not what the spending page looks like. If someone who has a similar dashboard themseles could set their language to ru and take a screenshot, that'd be pretty convincing.
(also: i'm not trying to pump my yes position, which is extremely trivial compared to my portfolio, I just think posting about this is fun)
He’s done this type of thing before apparently
Russian Student Allowed to Keep Diploma for ChatGPT-Written Thesis
@ZviMowshowitz if the story is generally true, but he told Karina that he was using GPT before the proposal, does that count as YES or NO?
@GraceKind He doesn't say he didn't tell her, so that would not be an issue. Indeed, I really want to know when/how she found out! Presumably she knows now and the thread is hopefully not how he told her about all this.
@CDBiddulph someone asked him in the replies if he told her at some point and he answered that yes, she knows about it.
@BairAiushin He replied to my posting on Twitter in English, so I asked him how she eventually found out. We'll see if he answers. The story isn't complete without that. I need my third act!
I think it's plausible, there are a lot of ways you can interpret his language. It seems like he's claiming that gpt4 was handling the first few conversations they had. And, especially for someone who's at or below average, is there really that much complexity to those? GPT-4 can write great, like, flowery scene characterization in fiction, it can write dialogue between generic people, which seems to be about the same level to me. Tinder conversations are short, that doesn't seem so implausible for a chatbot to get right 1 in 4 times if you help it a bit. He says on the second version (the one that started working):
I also turned on reply validation so that the bot would send me a message to proofread first. V2's hallucination problems have been reduced to zero. I just watched ChatGPT get acquainted and everything worked. This run has 4943 tags in a month on Tinder Gold and it's scary to count how many encounters there have been
If he rerolls a lot of messages, and gives it a lot of context to pull from, that sorta makes sense. It gives him a lot of bits to make gpt's surface-level-plausible output hold together with more context.
And then once we get into longer relationships, he seems to be very clear that it's both him and chatgpt communicating
The most charming girl was found - Karina. With her ChatGPT communicated as V1 and V2, stopped communicating for a while, then I continued to communicate myself and through ChatGPT V2. Very empathic, fun, cute, independent and always on the move. Simply put, SHE!
And GPT-4 is going to fail horribly at seeming like a real human over even a week or two of interpersonal context ... but if it's him writing a lot and him managing gpt's output's a lot, that's plausible again.
In V3 I didn't need to search for people, but only to maintain dialog. And already communication is not with thousands of people, but with Karina. So set up V3 as an observer who communicates when I don't write for a long time, and advises how to communicate better. For example, support, don't fight, suggest activity
Extremely plausible! It doesn't have to have actually helped much at this stage.
And the technical bits all look real to me.
Overall I don't think what he's claiming is really extreme or surprising. I'm not actually sure how it resolves - I think the central claim, and the first sentence, of 'I proposed to a girl with whom ChatGPT had been communicating for me for a year. For this purpose, the neural network re-communicated with other 5239 girls, which it eliminated as unnecessary and left only one.' is basically false, in that 'chatgpt had been communicating [with her] for a year' is highly misleading, it's more true that he was communicating with her using chatgpt to help after they initially met.
@jacksonpolack I am judging based on the details as offered in the thread. I agree that the headline claim is misleading but that's not the question I care about here, so long as what he says later in detail is true...