Prompt Labor: Basics

Andrew Ellis

24 April, 2024

Assistant menagerie

Assistant	Provider	Privacy	LLM	Capabilities	Pricing model
ChatGPT	OpenAI	👎🏼	GPT-3.5, GPT-4	Web search, DALLE, GPTs, multimodal input	💶
Copilot	Microsoft	👍🏼	GPT-3.5, GPT-4	Web search, DALLE, multimodal input	🆓 for BFH employees and students
Gemini	Google	👎🏼	Gemini Ultra, Gemini Pro, and Gemini Nano	Web search, multimodal input	💶
HuggingChat	🤗 Hugging Face	👍🏼	Various open models, e.g. CodeLlama, Llama 2, Mistral, Gemma		🆓

Orientierungshilfe für Lehrpersonen der BFH

• Haltung der BFH: Technologien, die den Lernprozess unterstützen und praxisrelevant sind, sollen in die Lehre einbezogen werden.

• Einsatz von KI in der Lehre: Die Mehrheit der Studierenden wird KI-Tools nutzen. Studierende sollen lernen, Technologien kompetent einzusetzen und kritisch zu hinterfragen.

Knowledge Base der Virtuellen Akademie

PDF

What is Artifical Intelligence?

A branch of computer science that aims to create machines that can perform tasks that typically require human intelligence.

What is a Large Language Model?

An LLM is a type of generative AI model that is trained to predict the next word following the input (prompt).

Training

How to train a language model

How to train a language model

• An LLM learns to predict the next word in a sequence, given the previous words: \[ P(word | context) \]
• Think of as “fancy autocomplete” (but very very powerful and sopisticated)

Text Generation

How does an LLM generate text?

Sampling

Auto-regressive generation

Text is generated one word at a time (actually tokens, not words).

1. Model predicts which word is likely to follow, given a sequence of tokens (words, punctuation, emojis, etc.).
2. A word is sampled from the predicted distribution.
3. The new word is added to the sequence of words (context) that is used to predict the next word.

\[ P(w_{w+1} | w_1, w_2, ..., w_t) \]

Generated text is dependent on the context.

Every token is given an equal amount time (computation per token is constant).

Auto-regressive generation

Foundation models

A foundation model, or large language model (LLM):

• is a type of machine learning model that is trained to predict the next word following the input (prompt).
• is trained “simply” to predict the next word following a sequence of words.
• does not necessarily produce human-like conversations.

: What is the capital of France?

: What is the capital of Germany? What is the capital of Italy? . ..

Training data

Figure courtesy of Andrej Karpathy

Training process

Figure courtesy of Andrej Karpathy

Assistant models

Trained (fine-tuned) to have conversations: turn-taking, question answering, not being [rude/sexist/racist].

• Foundation model has learned to predict all kinds of text, including both desirable and undesirable text.
• Fine-tuning is a process narrow down the space of all possible output to only desirable, human-like dialogue.

Reinforcement learning from human feedback (RLHF)

Model is aligned with the values of the fine-tuner.

How do Chatbots work?

• Designed to present the illusion of a conversation between two entities.

How do chatbots actually work?

An assistant model is a role-play simulator

• An assistant is trained to respond to user prompts in a human-like way.
• Simulates possible human conversations.
• Has no intentions. It is not an entity with its own goals.
• Does not have a “personality” or “character” in the traditional sense. It can be thought of as a role-playing simulator.
• Has no concept of “truth” or “lying”. The model is not trying to deceive the user, it is simply trying to respond in a human-like way.

Hallucination

• LLMs can generate text that is not true, or not based on any real-world knowledge.
• This is known as “hallucination”. A better term would be “confabulation”.

Can an LLM tell the truth?

• How would you know if an LLM is able to give you factual information?
• How would you test this?

: What is the capital of Uzbekistan?

: Tashkent

It looks like the LLM knows the capital of Uzbekistan¹.

1. What it is actually doing is responding with the most likely sequence following the question.

Knowledge base

• A knowledge base is a collection of facts about the world.
  • You can ask (retrieve) and tell (store) facts.
• An LLM is not a knowledge base.
  • LLMs generate text based on on how probable the next word is given the context, not based on stored facts.

What are LLMs good at?

• Fixing grammar, bad writing, etc.
• Rephrasing
• Analyze texts
• Write computer code
• Answer questions about a knowledge base
• Translate languages
• Creating structured output
• Factual output only with external documents or web search

Prompting

What is a prompt?

• Remember: the goal of an LLM is complete text.
• A prompt is a piece of text (instruction) that is given to a language model to complete.

PROMPT : Write a haiku about a workshop on large language models.

ASSISTANT : Whispers of circuits,
Knowledge blooms in bytes and bits,
Model learns and fits.

• The response is generated as continuation of the prompt.

More technical definition: The output is generated by auto-regressively sampling from the probability distribution over the vocabulary, conditioned on the prompt.

Unlocking knowledge

• LLMs learn to do things they were not explicitly trained to do.
• Often, these capabilities need to be “unlocked” by the right prompt.

• But what is the right prompt?
• The answer is very similar to what you would tell a human dialogue partner/assistant.
• You can increase the probability of getting the desired output by asking good questions or giving enough information.

Basics of prompting

OpenAI give a set of strategies for using their models effectively:

Prompt engineering

These include:

• writing clear instructions
• providing reference texts
• splitting tasks into subtasks
• giving the LLM ‘time to think’
• using external tools

Writing clear instructions

• Think of an LLM as a role-playing conversation simulator.
• Instructions should be clear and unambiguous.
• Indicate which role the model (persona) should adopt.

• Include details in your query to get more relevant answers
• Ask the model to adopt a persona
• Use delimiters to clearly indicate distinct parts of the input
• Specify the steps required to complete a task
• Provide examples
• Specify the desired length of the output

Adopt a persona (role)

: You are an expert on learning techniques. Explain the concept of ‘flipped classroom’ in one paragraph.

: You are an expert financial derivatives. Explain the concept of ‘flipped classroom’ in one paragraph.

Provide reference texts

• Provide a model with trusted and relevant information.
• Then instruct the model to use the provided information to compose its answer.

Instruct the model to answer using a reference text

This leads to retrieval-augmented generation (RAG). First create a database of documents, then retrieve the most relevant documents, based on a user’s query. These are then included in the prompt to the model. The model is instructed to use the information in the documents to compose its answer.

Giving GPT ‘time to think’

• LLMs generate text one word at a time–the model spends the same amount of computation on each word.
• Giving the model more context gives it more steps to “think”.
• This increases the chances that the model will give a good answer.
• This technique is known as chain-of-thought prompting, and can often be induced by simply instructing the model to think step-by-step or Take a deep breath and work on this problem step-by-step (Yang et al. 2023).

Chain-of-thought prompting

• Chain-of-thought prompting encourages the LLM to “explain” its intermediate reasoning steps.
• Enables complex reasoning and problem solving.

Instead of this:

: The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1. Yes or no?

Do this:

: Is this statement correct? The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1.

Reason through the problem step-by-step. Start by identifying the odd numbers. Next, add them up. Finally, determine if the sum is even or odd. Write down your reasoning steps in a numbered list.

Why does this work?

Zero-shot chain-of-thought prompting

: The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1. Take a deep breath and think step-by-step.

When using GPT-4, ChatGPT and Copilot do this automatically.

Use Markdown formatting

• Use Markdown to format your prompts.
• Instruct the LLM to format its output using Markdown.

: Improve this haiku:

Words weave through the air,
Minds meld with machine’s deep thought,
Knowledge blooms anew.

It is about about a workshop on large language models. I’m not happy with it.

Show me all the text. Format your edits as **TEXT** and show the deleted text as ~~TEXT~~. Keep you review short (max 100 words).

Try this example in ChatGPT.

Advanced prompting techniques

For more advanced prompting techniques, see these websites:

• Learn prompting
• Prompting guide
• OpenAI cookbook

and explore this activity.

Advanced LLM techniques

Retrieval-augmented generation (RAG)

Figure courtesy of Pinecone

Web search

• Similar to retrieval-augmented generation, but with web search.
• LLMs can be instructed to use web search to find information.
• Copilot does this automatically - ChatGPT (paid version only) can be instructed to do this.

External tools

• LLMs can be instructed to use external tools to complete tasks.
• For example, an LLM can be instructed to use a calculator to perform arithmetic.
• OpenAI calls this approach function calling.

Multi-agent conversations

Local models

• Download and run models, such as e.g. Llama 2 or Code Llama, locally.
• Ollama
• LM Studio

Hardware requirements:

• Apple Silicon Mac (M1/M2/M3) / Windows / Linux
• 16GB+ of RAM is recommended
• NVIDIA/AMD GPUs supported

References

Shanahan, Murray. 2023. “Talking About Large Language Models.” January 25, 2023. https://doi.org/10.48550/arXiv.2212.03551.

Shanahan, Murray, Kyle McDonell, and Laria Reynolds. 2023. “Role-Play with Large Language Models.” May 25, 2023. https://doi.org/10.48550/arXiv.2305.16367.

Wei, Jason, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, et al. 2022. “Emergent Abilities of Large Language Models.” October 26, 2022. https://doi.org/10.48550/arXiv.2206.07682.

Yang, Chengrun, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. 2023. “Large Language Models as Optimizers.” September 6, 2023. http://arxiv.org/abs/2309.03409.