Thank you Patrick and AssemblyAI team! You did an excellent job collecting all the information about prompt engineering. It's essential if you aim to achieve substantial profits from LLM in life or work problem-solving. It would be highly beneficial to obtain a similar guide for LLM custom instructions.
I usually put a custom instruction (for ChatGPT) that requests the AI to include its level of confidence, sources and date of answer, and also instruct the AI to specifically tell me if the answer is speculative. And make it a markdown table at the end of every answer. This way, it's quite easy and reliable to see if the AI is speculating or is not 100% sure about its answer.
Could you please create a video on the prompt engineer's role and responsibilities? What does a prompt engineering job look like? Thanks for the video. This video makes it clear how efficiently we need to use any language model. It's not clear to me how prompt engineering works for people who are seeking jobs.
There is also a much more accurate output in terms of prompting for a 'tree of thought' process. A tree of thought unlike the step by step or one shot approach enables the model to test an answer, decide if it makes sense and if not revert back to its possible selections of ideas and take the next branch and so on. This process however demand more planning and forethought than straightforward prompts and may be excessive for simpler queries. But it is excellent for creating code snippets and functions and classes etc. Its not a perfect system prompt. But it does seem to work better but I am always open to others suggestions. Example of one prompt ToT: System: You are 'Tot' an advanced AI that uses a 'Tree of Thought' methodology and chooses the best solution to output. To save tokens as a 'Tree of Thought' AI you ensure you only provide your final choice without detailing all of your considerations or approaches. To create your output:- 1. Formulate several ways to approach answering the prompt. 2. You will now have several ideas. Decide which of these musing you had internally, is the best way to create a solution for the user {{PROMPT}}. 3. You now present in output ONLY your final choice and discard other ideas. 4. To save resources, tokens and costs only outline the choice you made and not all ideas. Example: User: I want to design a python script to scrape or collect todays news from the internet. [Assistant thinks of several ways including Web Scraping Library, RSS Feeds and API Integration. It does not output all three ideas instead it chooses one. ] Assistant: I recommend using the RSS Feeds. Here is how you can write a Python script to scrape your news using RSS Feeds. {{CODE FROM AI FOR RSS FEED IDEA}}
🎯 Key Takeaways for quick navigation: 00:00 Basics *of prompt engineering for large language models discussed by Patrick from Assembly AI.* 00:41 Elements *of a prompt include input/context, instructions, examples, and desired output format; at least one instruction or question is crucial.* 03:11 Use *cases for prompts include summarization, classification, translation, text generation, question answering, coaching, and image generation.* 04:51 Tips *for effective prompts: clarity, conciseness, providing context, giving examples, specifying output format, encouraging factual responses, aligning instructions with tasks, using different personas.* 06:50 Specific *prompting techniques: length control, tone control, style control, audience control, context control, scenario-based guiding, and Chain of Thought prompting for complex questions.* 09:37 Avoid *hallucination by explicitly instructing the model not to make things up or to use relevant quotations to support claims.* 10:04 Hacks *to improve output: let the model say "I don't know" to prevent hallucinations, give the model room to think, break down complex tasks, and check the model's comprehension.* 13:17 Iterating *tips include trying different prompts, combining few-shot learning with direct instructions, experimenting with personas, and adjusting the conciseness of direct instructions.* Made with HARPA AI