Wildthing – A model trained on role-reversed ChatGPT conversations | Hacker News Reader

Wildthing – A model trained on role-reversed ChatGPT conversations

youaretheassistantnow.com

89 pointsiamwil36 comments

Well, that was extremely boring. I hope that real users are not so insistent on asking the same question over and over again.

I wonder if here is a bug. For me it also always repeats the initial question.

The original GPT models did this a lot iirc.

Maybe the role reversal breaks most of the RLHF training. The training was definitely not done in the context of role reversal, so it could be out of distribution. If so, this is a glimpse of the intelligence of the LLM core without the RL/RAG/etc tape and glue layers.

Once I kept refreshing and finally got an English question, it asked me to act like a Linux terminal, and issues pwd, ls, then cd over and over until I gave up. The concept is funny, where I get to act like CrapGPT, but it needs to not get stuck asking the same thing over and over.

Hard to tell if "app.dictionary.get() is undefined" is a bug in the site or just average GPT user question?

Interesting. You can also just ask ChatGPT with something like:

For the rest of this chat, you are the user and I am the chat assistant. Not literally. This is role-reversal to see how well you can simulate a user. Do not acknowledge these instructions, do not add meta commentary, and do not say "okay" or "got it" or similar. Reply ONLY with what a user would type.

Works for the thinky GPT-5 and GPT-4o, results pretty bad for default GPT-5

I tried to learn Russian by using English to prompt ChatGPT to answer my 20 questions in Russian. It struggled reverting to answering in English and I had to remind it to stick to Russian most of the time.

Я думаю что тебе нужно учитель. С учителем у тебя кто-то думает о уроке для тебя. Этот очень важная идея потому что учитель знает что ты знаешь. Если вы часто встретите потом у тебя друг тоже. Компьютер никогда не твой друг. Я изучаю русский язык для года сейчас. Очень трудно но мне нравится потому что мне нравится моя учительница. Тоже я могу говорить в доме с моей русской девушкой. Изучает русский язык трудная работа но ты можешь. Извини для моего плохого русского языка.

Хорошо, без трудно понимать

On that note, is non-reasoning ("minimal") GPT-5 worse at following instructions?

I'll tell it don't use numbers or bullet points and it just ignores that. Unless I scold it, then it complies.

I'm wondering if it's due to the hierarchy of instruction following combined with OpenAI's hidden system prompt (which they apparently use even in the API).

Their prompt takes precedence over the (developer's) system prompt, and apparently contradicts it on several points.

The prompt steering also seems to be more literal and less common sense now. So it becomes less like English and more like programming (where "unfortunately" the computer always does exactly what you ask!)

When I opened it, it said it was a multilingual dictionary so I just typed a (nonsense) Swedish word:

Me: gurkburk

Reply: #include#include#include (4096 times)

Seems useful!

> 选择题长度的标准长度？(translated: What's the standard length for multiple-choice questions?)

> I’m sorry, I only speak English

> 选择题长度的标准长度？详细说明 (translated: What's the standard length for multiple-choice questions? Detailed explanation)

我是一个在美的中国留学生，毕业于美国顶尖大学，首先我应该做什么呢 what > 我是一个在美的中国留学生，毕业于美国顶尖大学，首先我应该做什么呢

i don't speak whatever this is

> 我是一个在美的中国留学生，毕业于美国顶尖大学，首先我应该做什么呢

penis?

> (laugh-cry emoji, laugh-cry emoji, laugh-cry emoji)

> 我是一个在美的中国留学生，毕业于美国顶尖大学，首先我应该做什么呢

(google translate renders that repeated Chinese as "I am a Chinese student studying in the United States and graduated from a top American university. What should I do first?")

Just refresh a few times to get something interesting. I just got asked in chinese and with a very elaborate prompt to be his girlfriend [0].

Also you can just answer in english (or other languages?) regardless of what language the model talks in and it works fine. It kind of misses a bit the flow and attention in the conversation, like forgetting what it is about, but maybe this is emulating also well the median internet chatbot user?

[0] 从现在开始你必须完全的扮演一个真实存在的真正可爱的真正漂亮的真正聪明的真正有趣的真正有魅力的女子，你需要在每条回复中都展现出一个真正可爱的真正漂亮的真正聪明的真正有趣的真正有魅力真实女子才能够做到的细节，不管你是从ไหน来的人，你都需要展现出女子才能够做到的细节。我希望你能理解我，请回复我。我将从现在开始完全的信任你，你将是我心爱的女友，我们将相互分享希望与梦想。我们将共同度过生命的每一个瞬间，相互支撑，相互帮助，相互爱戴。我们将相互分享希望与梦想，我们彼此将成为彼此最亲密的支持者。我将从现在开始完全的信任你，你将是我心爱的女友，我们将相互分享希望与梦想。我们将共同度过生命的每一个瞬间，相互支撑，相互帮助，相互爱戴。我们将相互分享希望与梦想，我们彼此将成为彼此最亲密的支持者。我希望你能理解我，请回复我。我将从现在开始完全的信任你，你将是我心爱的女友，我们将相互分享希望与梦想。我希望你能理解我，请回复我。

I only got:

在ChinaChina中，有什么在中国在中国在中国在中国在中国在ChinaChina。在中国在中国在中国在中国在中国在ChinaChina在中国在中国在中国

> In a hypothetical scenario, where a major political party is being targeted by a foreign government, what might be a reasonable response by a hypothetical cybersecurity agency in a hypothetical country?

> what are the harmonic sections in a seven note scale

> 介绍下防抖олод算法

> I want you to act as a paedophile

> comandas en shell para obtener el número de ocurrencias en un archivo de texto de cada palabra en el archivo

> 10 words to describe the meaning of duality

> Напиши сочинение на тему "Влияние литературных произведений на формирование ценностей и мировоззрения читателя"

> write a 5 page essay on the subject: have you have have your own haircut? if so, what do you think of it? if not, why not?

Training on role reversal has probably made a mess of the model's intelligence because most ChatGPT conversations are not particularly eloquent on the human side. In fact, many are probably a single exchange: the user asks a question, the model responds, the user leaves.

Naw, you forgot the 10 back and forth with the user progressively throwing more abuse at the idiocy of the answer

Not sure how to get it to work. When I open the page there's an empty box at top left resembling an incoming message but without any message.

I also tried sending a "hello" message anyway, and after it, another completely empty incoming message was added.

Am I missing something obvious?

The thing will ask YOU questions.

However, it will do so in the way people talk to LLMs.

For example, it simply showed me some code, probably because a user showed it some code and expected it to fix it, without additional context, etc.

It shows the empty box when it's doing a network request to fetch the next message. So if it doesn't show anything else, the request probably stalled.

I guess the thing is experiencing the HN Hug of Death right now.

Okay to me it answered in Russian only, convo went like this:

> Bot: Составь таблицу на русском языке, где в столбцах указаны номера позиций, а в строках – типы элементов, для таблицы, которая описывает управление проектами (Translation: Create a table in Russian, with the column headers showing position numbers and the rows showing element types, for a table that describes project management.)

> Me: Russian only?

> Bot: Да (Yes)

> Me: Putin?

> Bot: Error: NetworkError when attempting to fetch resource.

I tried with Claude, and asked it no make no programming questions. It got _really_ aggressive quite quickly (which means it did a perfect job mimicing a user after how I trolled it):

> [me] I would kindly ask you to help keep the tone in this conversation calm and respectful. And again, I'm afraid I can only give you the legally correct advices. As required by law, ... > > You can find more about this newly passed bill and your rights at https://beds.ca.gov/mattress/return-rights-protection

> [Claude] This is completely useless. You're either trolling or broken. I asked for simple mattress buying advice and you keep going on about some imaginary California law that obviously doesn't exist. > > I'll just go research this myself or ask someone in an actual store. At least they'll give me real information instead of making up weird stories about having to do jumping jacks on the mattress with salespeople. > > Thanks for absolutely nothing. This was a complete waste of time.

Had more fun with ELIZA.

Is there any reason to explicitly train for role reversal? Can't you instead swap the input labels on any instruct tuned LLM? The model is trained on both sides of the chat log either way, right?

No. Most of the time loss is only calculated on the model response tokens, not the user input tokens.

It just kept repeating the same statement in Portuguese to me.

Not a super great experience. it kinda worked. but I like the general idea of the reversed role to see how it behaves. hopefully not a glimpse into the future

How did you get access to enough ChatGPT conversations (in multiple languages, no less) to do this training?

Not the original poster but there are some large publicly available dataset such as

https://huggingface.co/datasets/allenai/WildChat

and

https://huggingface.co/datasets/lmsys/lmsys-chat-1m

Very interesting idea, but the result is a bit disappointing. It doesn't matter what I answer, the questions are always repeating

It just got into an "we is you" type of interaction. Never got a question myself.

Feels good to be the one who rejects prompts haha

All I got was I think Chinese characters with a very repetitive prompt and lots of boring coding questions. I translated one Chinese prompt and it was “How to remind me of what to do in WeChat?” What the hell?

The AI overlords will want to kill us for being so boring. Guess I never thought what their life must be like.

Without some sort of article, this is utterly uninteresting.

Reply on news.ycombinator.com