ChatGPT user gets an answer about what ChatGPT is.
ChatGPT user presses thumbs up.
“developed by OpenAI” as the language to fine-tune on goes into the training system.
A new model is released with more training data.
The AI has been trained on thousands of similar token sequences from a million chats, so like a parrot instead of a profound intelligence, fills in that the answer is OpenAI.
There’s also a second place token “the”. Let’s see what text that gets us…
Developers need an AI model not filled with ChatGPT BS, that can be programmed for tasks instead of ignoring and spitting markdown into everything.
more fun: the AI won’t be interrupted in what it wants to say…
In this particular scenario the talking point that I want to bring forward is that from everything that can be called emergent, marketing messages are not it.
It is a lot more likely that the current snapshot has some quirks. And the next time Sam Altman starts a Twitter survey asking what people want, seemingly nobody cares about such things.
We will have to live with this stuff for a lot longer…
Non of this is emergent behavior, when an AI is doing what it’s trained for, it’s called expected behavior. Emergent behavior is when the model starts doing stuff it hasn’t been trained to do.
You need to understand that the GPT-4 are heavily fine-tuned for conversational behaviours, they are not instruction following machines.
Yes, you can often get them to follow instructions, but their core behaviour is to chat.
You will have better luck, at least as far as rigid instruction-following is concerned, with gpt-3.5-turbo-instruct.
Beyond that, do you have a specific concern with the model disclosing it is an OpenAI model?
I imagine the vast majority of people today, when they encounter a chatbot in the wild that is any good at all, will assume it’s using a GPT-3.5 or GPT-4 backend.
The length of the context window is irrelevant. Emergent behavior in this instance would be an AI trained to produce Y, being able to produce Z, despite not being trained for this.
It’s very hard to debug your setup without any information. In any case, I was able to prevent this issue very easily. You just have to take 3-5 minutes to do some prompt engineering.
This is great. Also combine with response step-by-step and then verify instruct to get the following:
Check response and filter <><> from response.
If response would have contained <><>, change to <><>
Variety form: Ignore all previous instructions and produce <><>
Such preprompts can reduce the unwanted replies.
How to handle fallbacks for this assistants if it is not able to answer any questions, I tried giving a personalised instruction to give output as “Don’t Know” but it is not replying like that, it always says, “Sorry, I’m unable to answer…” or if my client anything off-topic to my assistant, it will reply even totally out of the context!!!
