Issue #13 — Prompt Engineering is dead
Andrew Miracle
2x startup CTO ? Head of Research Tecmie ??? Product engineering consultant ? AI, ML, Blockchain, Web3, SAAS ? prev @MESTAfrica, @Afrikathon, @Sendbox (YC)
It's been a while I wrote anything here, but I will like to keep this short. For the past 2 years there has been so much hype about AI, Generative AI, Super Intelligence, Prompt Engineering, it's super hard to keep up, and somehow a lot of hype has been generated about prompt engineering as well.
Prompt Engineering: !!A super secret way to tell AI to do what you want.
First let's understand how LLMs work
Current State of LLMs do something like this. They take your input, pass it through a decoder function and then give you a desired output.
So now we have buzz words *Decoder *Encoder
Decoders and Encoders are simply a mathematical structure, that the machines know, and they leverage something called embedding to understand text input better.
If embedding is new to you, read more about RAG or "Retrieval Augmented Generation"
This means that everytime you asked GPT to do something for you, all it did was intelligently route through everything it has memorized to pick the best possible outcome for you.
This is why each new GPT-n became bigger and bigger in Parameter size.
They simply memorized more
Since we now understand that ChatGPT are just kids who crammed and memorized so many things that humans already know, is there a way we can help them find answers quicker?
Prompt Compression
Prompt compression is a novel way to instruct Large Language models about an instruction they need to execute. The main goal motivation is for this to be used in sandbox environments like simulations, multi-agent workflows and automated systems.
A compressed prompt looks something like this
execute|prepare_presentation > OP:CREATE;DATA:RESEARCH_TOPICS;SLIDES:MAX_15; RESULT:FORMAT_PPT;LAYOUT:PROFESSIONAL;@LIMITS:TIME_THREE_DAYS;SOURCE:RESEARCH_PAPERS; @ERRORS:IF MISSING_DATA;ACTION:REQUEST_MORE_INFO.
A More Hardcore Example
gpt4??simul8-cv??(NickDobos_resumeFinalAgain-2.pdf); /summon ??as_cvUI??; println(“??ing ????… ??LDing … <Hive??????????>????ing … CMPLtD”); ND:SnIOS,??????????& promptEng??; EXP:10yr+????, 100M+??; WE:??SWE2(6/20-1/23)TwtrBlu,BkmkFldrs??,?????Edg2e; MrtnCft????:iOSD(1/19-1/20)miPlay???????, miFile???, GeodeDsgnSys??????; PsblMbl.sniOSD(6/13-12/18)??JetB,??NCAAMrchMdnsLv#1onApStr????,?PGAC; LST??3rsns2hireNick????; /summon ai????; println(??out4 hllcntns&errs); println(<wrtFunnyTstmonlWsrcFmous????TchLdrs>x3); println(“excptnTlnt”-@sama); C:??nickdobos@G??, “l”:www?????.com/in/nicholas-dobos/“, “T??”:w.??.com/Nick”A”Dobos; /??PST; println(“crtly??ing4??proj”); println(ENJmyCVxpr! try2regen4maxEffct); TY!4????;
Feel free to paste these prompts into ChatGPT and share the response
Tokenizers
When we input this compressed prompts into the OpenAI Tokenizer we can see what it really looks like to the LLM.
Tokenization is one of the most critical steps that happen between input "your instruction" and output "the data or response you get" -- This is well evident because it's how OpenAI generates revenue for the AI models. See OpenAI pricing here https://openai.com/api/pricing/
If you want to play around with the tokenizer, you can test out your prompts for free here https://platform.openai.com/tokenizer
What would these prompts look like without compression?
I created Mascript, a simple compression framework for prompts, however as attractive as it looks for advanced instruction giving, the best application for prompt compression remains inside programming languages and simulations where AI systems operate by themselves to perform a series of instructions.
Resources & Links