What is an LLM? A primer for life science teams

A quick primer before the walkthrough:. No coding, no math — just the mental model you need before trying any other tutorial.

In plain English: A large language model (LLM) is a very good pattern-matcher trained on huge amounts of text. It predicts the next word, over and over, until it has written something that looks like a useful answer.

Key terms:

• LLM (Large Language Model) — ChatGPT, Claude, Gemini, and Copilot are all LLMs.
• Token — a chunk of text (roughly ¾ of a word). LLMs read and write in tokens.
• Context window — how much text the model can hold in its head at once. Bigger = more of your document it can read.

Comfort level: No prior AI knowledge. If you can read this page, you can read this tutorial.

Why life-science teams should care

You don't need to become an "AI person" to use LLMs at work. The teams getting real value are treating LLMs like a very fast, sometimes-wrong intern:

• They draft — first passes at reports, SOPs, emails, code, competitive briefs.
• They explain — a legacy Python script, a dense paper, an FDA guidance section.
• They structure — pull messy notes into a clean table, extract entities from PDFs.
• They critique — spot gaps in an experimental design or a report.

They don't (and shouldn't) make final regulatory calls, invent citations, or replace a domain expert's judgment.

How an LLM actually works (the quick version)

1. During training, the model reads billions of words from the public internet, books, and code.
2. It learns statistical patterns: given the words so far, what word is most likely to come next?
3. When you send a prompt, it applies the same trick: one token at a time, it predicts what should come next until the answer is done.

That's it. There's no reasoning engine underneath. Everything an LLM appears to "know" is a statistical echo of its training data plus whatever you gave it in the prompt.

What this means in practice

• It can be confidently wrong. This is called a hallucination. Especially bad for citations, dates, numbers, and dosages.
• It has a knowledge cutoff. The model was trained up to a certain date. It doesn't know about last week's FDA action unless you paste it in.
• It has no memory between chats (unless the product adds one). Each conversation starts fresh.
• Giving it your document changes everything. The model is much more reliable when answering about text you provided than when answering from memory.

The one habit that separates good users from bad ones

Ground everything. Paste in the paper, the SOP, the CSV, the meeting notes — and ask the model to answer from that text. Then verify the parts that matter.

A ChatGPT session that starts with "Summarize the attached protocol and list any missing controls" beats one that starts with "What controls do I need for a Western blot?" every time.

What to try next

• Your first ChatGPT prompt for lab and office work — the smallest useful thing you can do today.
• Choosing the right AI tool: ChatGPT, Claude, Copilot, or Gemini — which one to open for which job.
• Safer AI use at work — before you paste anything sensitive.

Common beginner questions

• Does it understand what it is saying? It predicts likely next words from patterns in text. Useful output does not imply understanding.
• Why does it give me different answers to the same question? There is deliberate randomness in how it picks words. Consistency is not something you can assume.
• Why does it make things up? It is optimized to produce fluent text, not to know when it does not know. That is why verification is part of every workflow on this site.
• Is my data used for training? It depends entirely on the product and plan. Check your organization's agreement rather than the marketing page.
• Will a bigger model fix the mistakes? It reduces some and hides others. Habits — give context, ask for sources, verify numbers — matter more than model choice.