Skip to main content

Command Palette

Search for a command to run...

Monet - Writing Improvement 2

Updated
4 min readView as Markdown
Monet - Writing Improvement 2

Current AI agents are great at filling the "gap" that is presented by the prompter; however, it isn't the best when it comes to initiating the "gap."

Let's assume LLM as a blackbox. From fifty thousand feet view, you can view it as a gradient, a landscape. A prompter places a ball somewhere and it rolls downward and settles at some point.

It works great when prompter asks the right questions and ask the agent to start from a context (i.e. a location in this vast gradient). However, it stops there. It produces one resulting artifact. It does not divert. It does not go on tangent. It does not spike into a totally new genre. It eventually lands on a local minima (a well big enough for majority of prompts to land on) that is called AI slop.

These wells are good for filling in "gaps" and "inbetweens" but the question I would ask is whether they have the novelty that humans look for.

To make the matters worse. These wells are overcrowded. Many people ends up on these wells. Even if it creates an interesting artifact, if they see it twice, it is no longer new. People get accustomed and lose interest.

If everyone offers same kinda dishes and the dishes are infinitely reproducible there is also no business value found here.

LLM is essentially a large multi-dimensional terrain where we roll a ball down its slopes to discover set of configurations (i.e. stories). I want to find a way to make a harness or way of stepping that will allow it to guide itself out of these slop wells or avoid them entirely.

Let's run a thought experiment. Imagine there is a good writer, Sam. He published several books and is highly acclaimed writer. Now, he just got an assistant, Tom. Tom produces same kind of work as current SoTA LLM models based what you ask him to do.

After a year you have following review for him:

  1. The random appearance new elements to story is way too random. There is no foretelling. Each element in the story should have a meaning and a purpose.

  2. You are able to create good sentences but you lack the spark of imagination. You are way to mechanical in handling my task. As a creative assistant, I expect you to show me something more than what I ask for. I want you to give me something that would actually be on a best seller book.

  3. Your writing stops abruptly as if there was a predefined timer and you had to close the chapter. Each chapter should develop a story or a character further.

  4. Overall, not engaging to read. I know it may not sound all that constructive to say this but I need some spices. Stories are bland. There is no level of alertness or tension in the story that keeps me turning the pages.

Overall, this is kinda of problem you would see in a human assistant. Tom is very unusual as if he has lost some part of brain function that allows for this creativity that is unique to humans.

Even though a kid may not be able to produce something very engaging, their stories still has power. They employ things that happen in real life and things that they thought was fun. Also kids tell the story while another is listening and if that audience seems to be getting bored, they introduce some random elements to the story to make it more interesting--kinda like a tug of war.

But Tom is different. His writing skill is great. He is great at context. However, he just lacks that human-ness as if he was not even a human.

One solution is to have Tom take care of small stuff and filling in gaps and in-betweens but have Sam write the overall synopsis.

But what if we want to grow Tom to become a fully fledged writer.

  1. Initialize empty string "" as HOW.md.

  2. Ask Tom to do something that is hard for him to do.

  3. Then score it.

  4. Then ask him to improve it.

  5. Then update "HOW.md" according to the learnings.

  6. Also ask it to build tools and checklist to help it track its writing.

  7. Repeat the process.

That's the basic idea. So the agent's intellectual is not going to improve but it will be able to build its pre-context and use it to improve.

-- Sprited Dev