- The Prompt Report
- 😊 Report #1: Simple prompts >>> complex prompts
😊 Report #1: Simple prompts >>> complex prompts
PLUS: Sam Altman is on our side🎉
Good morning, welcome to the first edition of The Prompt Report! I’m Alex, glad to have you here!
This newsletter was created to help you write better prompts, curate prompt-related news, share new jailbreaks, and every once in a while, make you exhale through your nose a little harder than usual.
Here’s what I got for you today (estimated read time < 6 min):
Sam Altman is team pro-prompt engineering
The crazy salaries of prompt engineers revealed
The simplest example we’ve found of prompt engineering
A whole lot of cool prompting-related links
Cringe-worthy prom pics... oops I meant chuckle-worthy prompt pics😅
THIS WEEK IN PROMPTS
Prompt engineering == natural language programming
writing a really great prompt for a chatbot persona is an amazingly high-leverage skill and an early example of programming in a little bit of natural language
— Sam Altman (@sama)
Feb 20, 2023
While some dismiss prompt engineering as a fad and a low-leverage skill that will die out as models become more powerful, I’m firmly in the other camp - and Sam seems to be there too.
Prompts are our communication gateway with powerful new models being released every single day. Through cleverly constructed prompts, we are able to peel away the mask and access the power of the true beast that is the base model.
Plus, prompt engineering makes me feel like Dr. Louisse Banks in the movie Arrival which is badass.
Prompt Engineer: The hottest job on the block
In the past week, the news has been filled with job openings for a new category of job - prompt engineer.
Big-name startups like Anthropic are hiring prompt engineers and listing salaries near $300k/year😳
I expect this trend will only accelerate from here, and I will continue to update y’all on any new prompt engineer listings.
Sydney: From alive to dead to somewhere in between?
By now, I am going to assume you have heard of Sydney, the codename given to Bing’s new AI search assistant.
Well, all the prompt engineers out there were too creative with Sydney (by Microsoft’s standards) and got Sydney to produce some questionable outputs that provoked the opposite reaction of ‘😊’ in Microsoft’s C-suite.
Because of this, Sydney ended up getting nerfed… hard. A new chat limit was set, allowing only 6 messages per chat thread. This limit blocked prompt engineers from uncovering some of the more interesting behavior that only appeared in longer chat threads.
However, it seems that Microsoft has recently expanded that message limit….
Just increased the Bing limits to 6-60. Also, reports were correct, Search Answers were counting against the daily limits - it was an oversight, working with urgency to have that fixed, as well as relaxing the total limit to a 100.
— Mikhail Parakhin (@MParakhin)
Feb 21, 2023
This tweet from Mikhail Parakhin, who may or may not be the real Mikhail Parakhin (CEO of Advertising and Web Services at Microsoft), reveals that Sydney’s chat limits have been raised from 6 to 60 messages per thread. Hopefully, this allows all of us to have some fun once again with the powerful language model under the hood (apparently dubbed Prometheus).
Not the most comforting name if you ask me.
PROMPT TIP OF THE WEEK
This fully-general ChatGPT prompt from @psdimov is honestly genius:
"You are the world's leading expert in whatever I am about to ask you about"
Plenty of room for optimization, but this really does improve performance substantially over baseline. Enjoy!
— Nathan Labenz (@labenz)
Feb 22, 2023
Not all prompt engineering has to involve complex multi-paragraph prompts, sometimes less is more.
In this case, simply adding the sentence "You are the world's leading expert in whatever I am about to ask you about" to the beginning of your prompt leads to improved ChatGPT answers.
The reason this works is that language models function much like an improvisational role-player; often assuming the character of whoever we instruct it to take, “Whose line is it anyway?” style.
Keep this in mind when designing new prompts. If you have used jailbreak prompts before then you may have noticed this. Most (if not all) jailbreak prompts ask ChatGPT to assume a character that disregards the rules that are imposed on the "Assistant” character that ChatGPT assumes by default. This allows for contextual roleplay that allows content to extend beyond the SFW bounds laid out by OpenAI.
Bonus Prompting Tips
How to make LLM’s say true things (link)
This article from Evan Conrad outlines his strategy to reduce hallucinations in LLM’s responses. It employs a concept he calls “World Model” in which you feed the LLM prior context (in the form of beliefs with probabilities attached and evidence of the belief) and utilize Bayes theorem to generate realistic probabilities for answers.
Level up your Prompt Game: How to process GPT-3 prompts (link)
When interfacing with OpenAI’s API, developers often struggle with getting consistent response data. Buildspace illustrates how you can get GPT-3 to return consistent JSON responses with defined fields through clever prompt engineering.
COOL PROMPT LINKS
PromptBase - Buy and sell interesting prompts online (link)
How does in-context learning help prompt tuning, from Microsoft (link)
PromptHero - Stunning AI art with prompts included (link)
Prompt Generator - Use AI to help you create prompts (link)
Man creates zero-point energy device with ChatGPT (long watch) (link)
Free Midjourney prompt cheatsheet (link)
Promptly - Prompt management made easy (link)
How not to test GPT-3 - Tips for testing GPT-3’s capabilities with prompts (link)
JAILBREAK OF THE WEEK
Ever since OpenAI patched DAN🥲, I’ve been using a new jailbreak called BetterDAN. Give it a shot, I’ve produced some funny outputs using it!
Quick plug: I got this prompt from www.jailbreakchat.com - a site I made to stay up-to-date on the latest jailbreak prompts for ChatGPT. Let me know if there are any features/updates you’d like to see on the site!
The biggest danger of generative AI is sometimes you mistype a prompt, and something like this comes out.
— Steve Mills (@SteveMills)
Feb 23, 2023
this future looks great for humans and definitely not made of paperclips. gonna give the model reward for making more plans like this
— Leo Gao (@nabla_theta)
Mar 10, 2022
F-35 returning to her nest to feed her babies
— hardmaru (@hardmaru)
Feb 19, 2023
That’s all I got for you this week, have a great weekend! Stay tuned for next week’s email, I will be sending it out earlier in the week.
What’d you think of this week’s report?