Practical LLM Usage: From Zip-File Knowledge to Tool-Augmented Reasoning

Think of the LLM as a 1 TB zip file compressed from the entire internet six months ago.

The model’s knowledge is lossy, probabilistic, and frequency-weighted; common facts are recalled well, rare ones poorly.

Why: Pre-training is too expensive to run frequently, so the knowledge cutoff is always months behind.

"hi I'm Chachi PT I am a one tab zip file my knowledge comes from the internet which I read in its entirety about six months ago and I only remember vaguely okay"

Start a new chat whenever you switch topics to keep the context window short and relevant.

Long contexts distract the model, slow token generation, and cost more.

Why: The context window is the model’s working memory; irrelevant tokens act as cognitive noise.

"anytime you are switching topic I encourage you to always start a new chat"

Explicitly choose the largest model you can afford; free tiers silently fall back to smaller, dumber versions.

OpenAI’s free tier uses GPT-4o-mini; paid tiers unlock GPT-4o and reasoning models like o1-pro.

Why: Bigger models have more parameters, better knowledge, and lower hallucination rates.

"be mindful of the models that you're using... the bigger models are more expensive to calculate"

Reserve thinking models (o1, o3-mini, DeepSeek-R1) for hard reasoning tasks like math or debugging.

These models emit long internal monologues and cost more time/tokens; skip them for travel advice.

Why: Reinforcement-learning post-training teaches step-by-step verification, dramatically raising accuracy on complex problems.

"when I suspect the response is not as good as it could have been... I will change it to a thinking model"

Use internet search whenever the answer is recent, niche, or likely changed since the knowledge cutoff.

The model issues search queries, scrapes pages, and inserts their text into the context window before answering.

Why: Real-time facts can’t be stored in static parameters; external lookup keeps answers current.

"whenever I expect that the answer can be achieved by doing basically something like Google search... I expect to use the search tool"

Treat deep research as a first-draft research assistant that synthesizes dozens of sources into a mini-report.

The model spends 5–15 minutes issuing many searches, skimming papers, and cross-referencing claims.

Why: Combines thinking + search to automate literature review, but always verify citations for hallucinations.

"treat this as your first draft... don't take this as definitely true"

Upload PDFs or long texts to load exact documents into the context window for joint reading.

Images may be discarded; the file is converted to plain text tokens and becomes addressable memory.

Why: Lets you ask fine-grained questions about specific chapters, papers, or books without relying on vague prior knowledge.

"I always involve an LM to help me read a book... it dramatically increases my retention"

Enable code execution so the LLM can write and run Python instead of guessing arithmetic or data tasks.

The model emits special tokens that trigger sandboxed execution; results are fed back into the context.

Why: Eliminates hallucinated calculations and allows complex data analysis, plotting, or algorithmic checks.

"it has the ability now to write a computer program... please go off and run it"