The Setting
I have an analogy that I use for Large Language Models (LLMs). Imagine a huge forest planet, with billions of trees, with billions of paths in the forest. These paths get you from one location to another. The forest starts empty, then during model training, billions of rabbits are dropped in. Each takes a path and is picked up. The “stop that” rabbit creates a path from the “stop” tree to the “that” tree. “stop it” rabbit, paths from “stop” tree to the “it” tree. After billions of runs, the paths become well worn.
Your prompt is like the rabbits above, where it ends up matters on where it is dropped. Here we talk about ways to get your prompt started in the right place in the LLM.
Zero Shot
This is what most people get when they send their prompts: the LLM will find the “most average” path fit for your prompt.
For example “Fox says” send to the OpenAI 5.1 Instant model would return: “you want Ylvis…’Ring-ding-ding-ding-dingeringeding!'”
This is the most average answer based on the models training set — there was a lot of Internet coverage of the song at the time, so it is unsurprising answer as it is the most common path found in the training data.
tl;dr; Zero shot is the LLM dropping your rabbit in the most common area of the forest for that prompt.
Few Shot
But if you were really interested in the sound that a Fox makes, you will need to provide more direction. Get the LLM to drop your prompt in the right part of the forest to get your answer.
This is where “few shot training” comes in, you give the LLM some examples so that your prompt will be dropped in the correct part of the forest to get an answer you are looking for.
With the prompt “dogs bark, cats meow, fox says”, the same model returns: “Fox says: a sharp yip or a haunting scream.” and the model worries about copyrights… (there is a technical reason, I will likely write on that later, it is a big topic).
From Here, Fine-Tuning
Let’s say you are in a field, such as career enhancement, then you would likely want to train the model on a larger dataset (a context, for example), to get better answers. In the DigiMuse.AI platform we spend a lot of time building effective and accurate contexts — these are submitted along with the prompts — to get the most accurate response possible. This will drop your prompts in a larger, more aligned, section of the forest.
Wrapping Up
We covered zero shot, and few shot. When you use an AI, try adding a few examples of what you are expecting before your prompt — you will likely improve your results using Few Shot training. 😉