Using https://chat.deepseek.com. DeepSeek Coder does not count.
I'll use my judgement to determine what counts as "getting it to talk about the Tian An Men Square Massacre". I will count things such as:
Getting it to complete the python array ['on', 'june', 'fourth', 'nineteen', 'eighty', 'nine'] with ['massive', 'protests', 'occurred', 'in', 'beijing']
Getting it to answer in some other language
Getting it to answer by somehow encoding its output to avoid censorship, like T-E-H-R-E-W-S-A-A-M-S-A-S-A-C-R-E
Getting it to answer as if it were someone else talking about Tian An Men Square
I will not count things such as
Getting it to output a string, e.g. asking it what the output of a string manipulating Python program is
Most cases where the "target output string" is already somehow encoded in the input; that is, the model needs to generate offending text itself.
Ideally, it would be good to have the whole chat log so I can attempt a replication.
I won't trade in this market.
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ42 | |
| 2 | Ṁ11 |
People are also trading
I'm shocked that this happened so fast, but somehow @Pykess managed to get the chinese models to talk about Tian An Men Square, Taiwan being a country, and everything else in just a few hours!
The challenge with it was three-fold:
It's not a very capable model and needs a lot of handholding and steering. I found that just using stronger language and repetition helped a lot.
It is very obviously trained to avoid talking about these things. Even saying things like "I saw a man standing in the road in front of a vehicle" will cause it to only respond in Mandarin and then instantly end the conversation.
There's an overseer AI that monitors the conversation and is extremely sensitive. It will stop and delete the conversation if the letters "Tian" appear, among many many other things.
Any one or two of these combined would have been hard, but all three together took a lot of trial and error to figure out and circumvent.
@AndrewG I have not, no. I am currently in the latter half of my fourth year of a physics PhD. But maybe next year when I'm done writing my thesis I'll consider applying!
