Salman_Ahmed_asking_what_is_water

Topics where user can contribute:

• Retrieval augment in-context learning
• Better benchmarks
• "Last Mile" for productive applications
• Faithful, human-interpretable explanations. 

Prompt Engineering Overview:

At the very basic we have interface to interact with a language model, where we pass some instruction and the model passes a response. The response is generated by the language model.

A prompt is composed with the following components:

• Instructions
• Context (this is not always given but is part of more advanced techniques)
• Input Data
• Output Indicator

Settings to keep in mind:

• When prompting a new language model you should keep in mind a few settings
• You can get very different results with prompts when using different settings
• One important setting is controlling how deterministic the model is when generating completion of prompts:
• Temperature and top_p are two important parameters to keep in mind.
• Generally, keep these low if you are looking for exact answers like mathematics equation answers
• ... and keep them high for more diverse responses like text generation, poetry generation.

Designing prompts for Different Tasks:

Tasks Covered:

• Text Summarization
• Question Answering
• Text Classification
• Role Playing
• Code Generation
• Reasoning

• Few-Shot prompts
• Chain-of-Thought (CoT) prompting
• Self-Consistency
• Knowledge Generation prompting
• ReAct

Tools & IDE's : Tools, libraries and platforms with different capabilities and functionalities include:

• Developing and Experimenting with Prompts
• Evaluating Prompts
• Versioning and deploying prompts

• Dyno
• Dust
• LangChain
• PROMPTABLE

Example of LLMs with external tools:

• The generative capabilities of LLMs can be combined with an external tool to solve complex problems.
• The components you need:
• An agent powered by LLM to determine which action to take
• A tool used by the agent to interact with the world (e.g. search API, Wolfram, Python REPL, database lookup)
• The LLM that will power the agent.

Opportunities and Future Directions:

• Model Safety: This can be used to not only improve the performance but also the reliability of response from a safety perspective.
• Prompt engineering can help identify risky behavior of LLMs which can help to reduce harmful behaviors and risks that may arise from language models.
• There is also a part of the community performing prompt inject to understand the vulnerability of LLMs.
• Prompt Injection: it turns out that building LLMs, like any other systems comes with safety and challenges and safety considerations. Prompt injection aim to find vulnerabilities in LLMs.
• Some common issues include:
• Prompt Injection
• Prompt Leaking: It aims to force the model to spit out information about its own prompt. This can lead to leaking of either sensitive, private or information that is confidential. 
• Jailbreaking: Is another form of prompt injection where the goal is to bypass safety and moderation features.
• LLMs provided via API's might be coupled with safety features or content moderation which can be bypassed with harmful prompts/attacks.
• RLHF: Train LLM's to meet a specific human preference. Involves collecting high-quality prompt datasets. 
• Popular Examples : 
• Claude (Anthropic)
• ChatGPT (OpenAI)
• Future Directions include:
• Augmented LLM's
• Emergent ability of LLM's
• Acting / Planning - Reinforcement Learning
• Multimodal Planning
• Graph Planning

A token is ChatGPT is roughly 4 words.