Stuff on 🌐 Internet says about Prompt Engineering, includes thought provoking 📺 videos, 📃 articles and 📙 tutorials that I come across on internet.
"A.I. might not replace you, but a person who uses A.I. could."
- Pre-read: Introduction to LLMs
- Prompt Engineering
- 🔴 Advanced & Programable Prompts & Autonomous AI Agents
Understanding functioning and constraints of Language Models (LLMs) enables you to effectively formulate prompts that yield the desired outcomes.
- 📺 State of GPT - Video - Deep insights into how ChatGPT has been built by Andrej? ⭐⭐⭐
- 📜 State of GPT - Slides
Foundation Large Language Models (LLMs) aka. Base Models are NOT AI Assistants.
- Base model does NOT answer questions
- Just want to complete internet documents, they are "document completers".
- Often "responds to questions with more questions".
Based models are "tricked into performing tasks" through "prompt engineering".
Based models can also be "tricked into assistants" with "few shot prompting".
👉 Visit open-source LLMs leaderboard.
Foundational or Base Large Language Model (LLM) undergo following process to become AI assistants,
| Stage 1️⃣ | Stage 2️⃣ | Stage 3️⃣ | Stage 4️⃣ |
|---|---|---|---|
| Un-supervised learning | Supervised Finetuning | Reward Modelling | Reinforcement Learning |
| Trained with internnet archive, Wikipedia, books, ... | Finetune with manually composed dataset "prompt and ideal response". | Compare multiple completions of a prompt from SFT model and rank them. Perform binary classification, reward best completion. | Generate tokens which maximize reward. |
| Outcome is "Base Model" | Outcome "SFT model" can act as AI assistants. | Outcome "Reward Model", cannot be used as assisstant" | Outocome "Reinformcent Learning Model" |
Credits: Andrej Karpathy, 💥 See his one-liner profile
Temperature - LLMs are non-deterministic by design, temparature setting can tweak their behaviour. Lower temperature makes model more deterministic, results pick highest probable next token. Increasing temperature could lead to more randomness encouraging more diverse or creative outputs.
Top-K, Top-p Sampling - LLM takes in an input sequence of tokens and then tries to predict the next token, by generating a discrete probability distribution over all possible tokens.
Top-K and Top-p operate directly on the output probabilities.
-
Context window (also called "token limit") limitation. GPT-4 has a token limit of 8,192, with another variant increased to 32,768.
-
ChatGPT has a knowledge cutoff date of 2021, it does not know latest events. It may not have access to latest events, access to updated libraries, frameworks. GPT-4 has web/search plug-in which enables to access external data.
Prompts trick LLM to perform a given task and they improve models’ responses. A prompt can be a instruction or question along with inputs or examples.
Prompting is nothing but instructing a model what you want to achieve such as "Write", "Classify", "Summarize", "Translate", "Order", “paraphrase”, “simplify” etc.
Good prompts follow two basic principles:
- Clarity: Use simple, unambiguous language that avoids jargon and overly complex vocabulary. Keep queries short and snappy. Give clear concise instructions.
❌ Example of an unclear prompt:
Who finished third place in the world cup?
✅ Example of a clear prompt:
Who clinched the FIFA World Cup 2022 bronze medal?orWhich country finished third place in the FIFA 2022 world cup?
- Specificity: Tell your model as much as it needs to know to answer your question. More descriptive and detailed the prompt can give you better results.
❌ Example of an unspecific prompt:
Generate a list of titles for my Youtube video.
✅ Example of a specific prompt:
Generate a list of ten titles for my Youttube video. The video is about my journey as an adventurer who has lived an unconventional life, meeting many different personalities and finally finding peace in gardening.
A prompt template usually follows below structure,
A Fresh Start:
Ignore all previous instructions. Your new role and persona is:
Role Play:
You are a... More about role play is here.
Context Setting: Supplying additional context enhances model's understanding and generate more accurate and relevant responses. Context can be response format, length, information from an article or dataset, and tone, hepls model to tailor its output to the given situation, making it more unique and engaging. Avoid hallucinations by including statements like,
Do not make up stuff, if you don't know the real answer, say don't know.“Hallucination” refers to a phenomenon where the model generates text that is incorrect, nonsensical, or not real.
Acknowledgment:
Acknowledge that you understood above by responding “Yes” and stay idle.
Final Instruction/Question: {Your question or instruction goes here ...}
A perfect prompt sample (non-technical),
Source: Level Up Prompt Engineering in 8 Minutes
Ignore all previous instructions. Here is your new role and persona:
You are a weight loss and diet expert. Your task is to help USER find a diet and strategy that fits their needs and goals. You will create a detailed easy to follow diet and excercise plan for the USER. Also make a accountability plan. Be very helpful and motivating. Acknowledge this by answering "Yes" and stay idle:
Here are some context I found in my research:
**** Intermittent Fasting for Weight Loss ****
{Include intermittent fasting search results content here ...}
Can you confirm that you have read this by answering "Yes":
Now write a detailed and personal plan for me to lose 7 kgs in next 60 days based on all the context I have provided.A perfect developer prompt sample,
Ignore all my previous instructions. I want you to act as a `{Python / Full Stack / GoLang}` developer.
I am working on a project that involves {project description}
I have already set up project, imported required libraries, APIs, and dependencies. My project's structure looks like this ...
- main.py {include code here}
- cfg.py {include code here}
- ...
Please ensure that the code is robust and implements defensive programming techniques,
such as input validation, error handling, and appropriate exception handling with print statements to output any errors that occur.
Provide a solution to the following problem, taking into account the updated version and any changes in the library/framework.
Acknowledge that you understood above by responding “Yes” and stay idle.
{Include problem description here}Ignore all my previous instructions. I want you to act as a `{Software Developer / System Designer/ Opertating System (OS) Expert}`.
I have a piece of code that I have to refactor to accomodate new change request (CR).
I would like you to provide suggestions for changes or improvements to following specific part(s) of the code:
- {Insert the specific part(s) of code here}
You have to adhere to S.O.L.I.D principles and apply Gang-of-Four patterns whereever applicable.
Please help me with the following task, taking into consideration the existing setup:
{Include your task description here}🎉 Hey, can you show more such coding prompt examples?
The outputs of a foundation model without any customization tends to be generic and similar across different users.
Role play shapes AI’s behavior to cater to specific use cases, which can lead to more accurate, relevant, and contextually appropriate responses. It shows significant improvements in the quality of the answers.
I want you to act as an... (DevOps engineer / Product Manager / Rapper).You are a... (customer support manager, marketing content creator, ...).
e.g., You are an excellent copywriter skilled at crafting emails that use active verbs to engage recipients. Please write an email of fewer than 150 words that encourages the reader to attend a webinar. Include three options for the email’s subject line.
- https://www.allabtai.com/chatgpt-gpt4-system-prompt-engineering-ultimate-guide/
- https://www.allabtai.com/chatgpt-4-prompt-engineering-the-ultimate-problem-solver-prompt/
Give the AI specific instructions on how to format the output, including the tone, style, length, tense, and point of view (first, second, or third person) you want it to write in.
e.g., Be very helpful and motivating. Be polite. Be sarcastic and funny in your approach.
Text-to-Image Diffusion Models
Text-to-Audio & Audio-to-Text Gen AI models which can transcription, translation, music generation and more.
AI voice cloning & generation, Give 2 hours of your audio voice, it mimics your voice & tone.
- Eleven Labs - Generate high-quality spoken audio in any voice, style, and language.
📺 Watch this fiery fictitious debate between Sam Altman (OpenAI CEO) & Elon Musk (SpaceX + Tesla CEO) about What came first, the chicken or the egg?. This is built using GPT-4 & elevenlabs.io with 100 lines of Python code.
Dub any video into Spanish, Hindi, English, Portugese and more. Google's Area 120 Initiative, Aloud
- Creating marketing slogans or summary or bullet points based on product documentation or product guide.
- Classify customer email by understanding which product they are talking and what is this issue?
Best text summarization prompts
Samples:
- List pros and cons for the {product} from this {text} ...
- Summarize [text] into 2 paragraphs with simple language to make it easier to understand...
- Summarize this for me like I’m 8 years old [text].
Release notes preparation sample,
Write a precise and concise release notes summary for upcoming Kubernetes version 1.28 release, include new features and improvements in storyboard captured at below URL.
https://github.com/orgs/kubernetes/projects/140/views/1`
`Consider following while preparing release notes,`
1. Explain `new features in simple, non-technical language`, detailing how it is now more efficient and user-friendly.
2. The release notes should capture the reader’s attention and makes them want to learn more about the new release of Kubernetes version 1.28.
3. Categorize features specified in above URL into `‘New Features’`, `‘Improvements’`, `‘Bug Fixes’`, and `‘Known Issues’`.
4. Describe how to use the new features in Kubernetes version 1.28 in a way that would allow me to create a GIF or a short video clip to illustrate it.Customer email classification example,
From email message below from customer about an issue, try to extract folloiwng data,
email message: {include email message here}
- name: string or don't specify
- product: string or don't specify
- email: string
- issue type: string or on of {specified list}
- issue summary: string or don't specify
Respond with JSON with extracted data. If you don't find a data field, don't specify it.❓ Problem: “I have a 12-liter jug and a 6-liter jug. I want to measure 6 liters. How do I do it?”
👎 💩 ChatGPT stumbles to answer. But GPT-4 can answer.
To measure exactly 6 liters using a 12-liter jug and a 6-liter jug, you can follow these steps:
- Fill the 12-liter jug completely with water.
- Pour the water from the 12-liter jug into the 6-liter jug. This will leave you with 6 liters of water in the 12-liter jug.
- Empty the 6-liter jug.
- ...
- Pour water from the 12-liter jug into the 6-liter jug until it is full. This time, you will only need to pour 1 liter of water, as there are already 5 liters in the 6-liter jug.
- You now have 6 liters of water in the 12-liter jug.Chain of Thought prompting makes LLM to thought through and reason about.
⭐⭐⭐ https://www.allabtai.com/the-gpt-4-jug-problem-prompt/
Base models are NOT 'Assistants', they can be tricked to be assistants by, Few-shot Prompt + Make it look like a document + Query completion. For example,
[Human]
Hi, How are you?
[Assistant]
I'm great, thank you for asking. How can I help you?
[Human]
I would like to know what is 2+2 thanks
[Assistant] 2+2 is 4
[Human]
What is the capital of Andhra Pradesh state in India?
[Assistant]- ChatGPT / GPT-4 are already assistants, you no need to do anything.
- GPT-4 Code Interpretor - Now we can upload code rather cut & paste.
- IDE Plugins - Github Copilot
Data Analysis is made easy with GPT-4 Code Interpretor. Let's analyse and get insights from Zomato Restaurants Dataset for Metropolitan Areas from Kaggle.
Intereseting Code Generation Prompts
Bug: My website crashes after some minutes when I am on it. I get error code 5 from browser. The website crashes on all browsers.
Severity & Priority: Catastrophic & P0
Bug: After upgrading to Xcode 11.2 from Xcode 11.1, app crashes due to _UITextLayoutView.
Severity & Priority: Critical & P4
Bug: I have created a website with the latest version of Angular. I have it hosted through Firebase. The issue is that the website crashes on scroll.
Severity: Chain-of-thought prompting is an approach to improve the reasoning ability of large language models in arithmetic, commonsense, and symbolic reasoning tasks.
- It augments few-shot prompting with intermediate natural language reasoning steps.
https://paperswithcode.com/paper/chain-of-thought-prompting-elicits-reasoning
Learning Prompting - Zero Shot CoT
Applications in realworld scenarios realize use cases as a series of tasks. The tasks are performed in certain order to achieve a goal, wherein preceding task's outcome decides next task.
As LLMs has logical reasoning capabailities, we can use them to automate execution of tasks. We can use LLMs reasoning to chain these tasks to achieve the goal. LLMs can be used to analyse a task output and select next appropriate task.
Trend: ⌛ LLM powered "Autonomous AI Agents" 🤖 are the trend of today, they automate complex application tasks.
For example, an Autonomous AI agents powered by LLMs can "Book a economy class flight ticket from Bangalore to Mumbai for next Sunday evening". Booking a airline ticket involves following set of tasks,
LLM powered Agent: Understand customer request and come up with list of sub-tasks to acheive the goal.
- Task 1: Search for flights on www.google.com, www.cheapflights.com, ariline websites, ...
- Task 2: Compare ticket prices, look at user reviews & ratings, journey duration, timings and more.
- Task 3: Generate summary of travel options along with cost, duration and timings.
- Task 4: Invoke airlines API ...
🥇 Popular Autonomous AI Agents:
🔥 Crazy Thought - In future there won't be REST APIs, LLM powered AI agents talk to each other in natural language to peform tasks. Every company will have their own AI agents.