What happened
June 2023, which is worth stating because the dates do the work in this one. GPT-4 had been announced three months earlier and ChatGPT was about six months old. Half the session explained what the thing was and half of it was everybody installing software.
The device that did the most teaching was a pair of sentences. A trigram model trained on a large corpus learns that after the words the cat is on the, the likely continuations are roof, floor, bed or mat, and it knows nothing beyond how often each followed that phrase in the text it saw. Then the longer version: the cat spotted a mouse, quietly it started to climb, the cat is on the. A model that reads the whole passage rather than the last three words stops being a frequency table, because climbing and a mouse change what should come next.
Tokens got the local example. Welcome to the product space is six tokens across twenty-eight characters. Coimbatore on its own is four tokens across ten. Whatever a token is, it is not a word, and the city the room was sitting in costs nearly as much as a sentence.
The anatomy offered was five parts: the instruction, a separator, the context, the input, and an indicator of what shape the output should take. Six kinds of prompt followed, from the plainly descriptive through role-based and conversational to creative.
Then ten practices, of which one has aged better than the rest: say what to do rather than what not to do, because a negative instruction spends the model's attention without telling it where to go. The others were about being concrete. Name the format you want. Cap the length in sentences or words. Break a large request into steps rather than asking for everything at once. Show an example of the shape you are after. Put the instruction first and fence the context off from it with hashes or triple quotes.
The second half was setup, and it is the part that dates the evening most clearly: install Visual Studio Code, then .NET, then Python, then the OpenAI package, then Polyglot Notebooks, and leave with a notebook of your own calling GPT. In the middle of 2023 that was simply what it took to get at a model programmatically. Anybody repeating this workshop now would delete that half of it, which is a fair measure of how much moved in the years afterwards.
What was promised
Unleashing the Power of AI Language Models
This workshop is designed to equip product managers and designers with the knowledge and skills necessary to harness the power of LLM's in their product development process. By understanding how to effectively craft prompts for AI language models, participants will learn how to leverage these models as partners in their day to day work and product development practice.
Siddarth Kengadaran
Product Strategist, Kellton
The workshop will comprise presentations, interactive discussions, hands-on exercises, and practical demonstrations. Participants will have the opportunity to experiment with AI language models, engage in group activities, and receive personalized guidance on prompt engineering techniques. The workshop will foster a collaborative learning environment where participants can share their experiences, challenges, and ideas.
By the end of the workshop, product managers and designers will have a solid understanding of prompt engineering concepts and techniques. They will be equipped with the skills to effectively leverage AI language models in their product development processes, enabling them to working effectively with these AI tools.