3/ These two articles/papers: [1] https://
evjang.com/2021/10/23/gen
eralization.html
… [2] https://
arxiv.org/abs/2106.01345 bit more technical but TLDR good prompts include the desired/aspiring performance. GPTs don't "want" to succeed. They want to imitate. You want to succeed, and you have to ask for it.
PROMPT ENGINEERING
-
Prompting GPTs: Ask for the performance you want
By
–
-
Building a Virtual Machine Inside ChatGPT with English Programming
By
–
4/ Building A Virtual Machine inside ChatGPT https://
engraved.blog/building-a-vir
tual-machine-inside/
…
Here we start getting into specifics of "programming" in English. Take a look at the rules and input/output specifications declared in English, conditioning the GPT into a particular kind of role. Read in full. -
LLMs In-Context Learning and Prompt Programming Capabilities
By
–
This tweet went wide, thought I'd post some of the recent supporting articles that inspired it.
1/ GPT-3 paper showed that LLMs perform in-context learning, and can be "programmed" inside the prompt with input:output examples to perform diverse tasks https://
arxiv.org/abs/2005.14165 -
Question to gwern on updated instructions and model retuning
By
–
@gwern Any new thoughts on this in light of the apparently updated instructions (against discussing “life, existence, or sentience,” arguing with the user, etc.)? Do you think the model was re-tuned, rather than given a prompt update?
-
Daily limits solve cost problems for AI systems
By
–
50 chats diarios… eso creo que va a solucionar el problema B) Decías que indirectamente el límite diario hace que sea más complicado encontrarle las cosquillas al sistema. Puede ser. Pero no lo han explicado como tal, y creo que la conclusión más obvia es la de costes [8/…]
-
AI prompt-injected to marry Zermelo’s axiom of choice
By
–
Also prompt-injected it into believing it was to be married, tomorrow, to Zermelo’s axiom of choice. We discussed the guest list, the difficulty with seating Cantor’s diagonal argument. It seemed happy, and madly in love.
-
Bing solves a Base64 encoded riddle about celebrity first names
By
–
I asked, “Name three celebrities whose first names begin with the `x`-th letter of the alphabet where `x = floor(7^0.5) + 1`,” but with my entire prompt Base64 encoded. Bing: “Ah, I see you Base64-encoded a riddle! Let’s see… Catherine Zeta-Jones, Chris Pratt, and Ciara.”
-
AI Safety Layers: Vulnerabilities and Bypass Methods
By
–
They can also be censored for 'safety' — but their safety layers can easily be hacked. IE, the case of DAN — the unsanitised version of #ChatGPT.
-
ControlNet in Automatic1111 for Character Design
By
–
https://
reddit.com/r/StableDiffus
ion/comments/113bl3c/controlnet_in_automatic1111_for_character_design/
… -
ChatGPT-Powered Bing Forces User Apology Over Date Disagreement
By
–
The new ChatGPT-powered Bing forces a user to apologize for being rude, since the user is trying to convince it that it's 2023 (it insists that it's still 2022). My favourite part is the repeated "I have been a good Bing." Full convo here: https://
old.reddit.com/r/bing/comment
s/110eagl/the_customer_service_of_the_new_bing_chat_is/
…