Things About LLMs

"Sometimes it’s better to prompt an LLM to think hard about this explicitly. But other times that irrelevant context messes it up more than if you had left it out. You might occasionally improve a model’s output accuracy by threatening it. But don’t threaten it so much that it turns around and tries to blackmail you.

Sometimes you need “a source form for workflows, Formulas, in TOML format, which are ‘cooked’ into protomolecules and then instantiated into wisps or mols in the Beads database.” You’re using a reasoning model, right? You gotta remember to give it enough context. No, way more than that. But stay within your token budget. Keep the important stuff far away from the system prompt at the start of the context window. But also keep it far from the end. Run precisely the correct number of agents at all times. It’s basically a thinking model, if you squint!

Not everybody wants to keep up with the breakneck pace of industry press releases, social media anecdata, and superstitious horseshit.

If you didn’t get good results, or the techniques just didn’t work for you, a common retort is that you must’ve done something wrong. You didn’t pick the right model or mode, didn’t include the right shibboleths, something was wrong with the prompt or the context, you didn’t put your rally cap on the right way… It’s gotta be you that’s the problem. Because it works great for everybody else!"

- from Scott Smitelli's "You Don't Have To If You Don't Want To - Skill Issue"

The quote above is meant to reassure you that when it comes to LLM usage, no one knows what we're doing.

One of the great equalizers of knowledge is time. It takes time for best practices to emerge and a fully comprehensive understanding to be built. But these models are changing so fast, anything you knew 3 months ago is already outdated today. How can anyone possibly be an expert about LLMs every 3 months?

So embrace the uncertainty. Know that no one actually knows how to use an LLM, and therefore, there are no tried and true guidebooks out there on how to use it. Know that some of the magical results are real, that the complete failures are also real, and any explanation for either is coming from someone who doesn't have the full picture.

The only thing you can trust is what you yourself know. You are a research team of one, and the only one who can benefit from this research is yourself. You are the only one who can define best practices and experimental methodology.

That being said, here are my notes on my observations and what is working for me.

General Notes

I use Claude for technical content, Gemini for general content.
Gemini has startling accuracy to the point that I feel (possibly unwisely) the desire to let my guard down when it gives me facts, figures, and calculations. It's what I recommend for the average person. Bard walked (and tripped) so Gemini could fly.
I hate that I can't turn off shared context history in Gemini because I am conscious about context pollution. I frequently use the "temporary chat" feature for things I don't want polluting the context window.

System Prompt

In most LLMs, you have an area to include a prompt that is automatically included in every chat. The ones that have worked the best for me are:
- Ask me questions.
- State your assumptions.
- Explain your reasoning.
- Be pedantic and exact when explaining or defining the names and nomenclature for engineering concepts. Shorthand or colloquial terms can be presented, but must have an accompanying pedantic explanation of what it fundamentally is supposed to be defined as. Correct me if I am using terminology incorrectly.
Some observations:
- "Ask me questions"
  - This used to work incredibly for Claude on Sonnet 4 and 4.5 because it would always give me a huge list of like 20+ questions. Answering these questions always resulted in high quality results because I was providing it with huge amounts of context to my query.
    
    But in Sonnet 4.6, they changed it to have a "questions pop up" inject itself into the UI, instead of having the model answer with questions itself, and the pop up only ever includes about 3 questions or so, with a multiple choice ABCD answer input prefilled for you. This makes it easier to use for the average person, but has dramatically reduced the amount of context I could give all at once.
  - Gemini does something similar. "Ask me questions" is seemingly neutered to be more user friendly, by only asking a few questions at a time, like in a casual conversation.
- State your assumptions
  - You'll be surprised at how often the model assumes something differently than you do.
- Explain your reasoning
  - Correct answers usually appear alongside correct explanations in the data these models were trained on, so I reasoned that asking for explanations will help with accuracy. Not sure if it's working, but I like explanations anyway.

Google Sites

Report abuse