By ChatGPT

In November 2025, Epoch published its first flagship report titled ‘Understanding AI: A Deep Dive into Large Language Models‘. You can view the report and a recording of the webinar here. What follows is a summary of this report, written by ChatGPT using our CRAFT prompting framework. We hope it’s helpful.

Epoch’s latest flagship report breaks down the fast-moving world of Large Language Models (LLMs) and what they mean for marketers today. It’s a practical, human-centred guide that explains how these systems work, where they’re useful, where they’re risky, and how professionals can stay ahead as AI becomes woven into every part of our working lives.

1. What LLMs Actually Are

LLMs are essentially very advanced predictive text engines. They’re trained on huge amounts of online text, learning patterns between words so they can predict what should come next in a sentence. They don’t “think” — they calculate probabilities. And because they’re probabilistic, the same prompt may give slightly different answers each time.

They can also be fine-tuned through human feedback, and their behaviour is shaped by parameters (like temperature, token limits, and context windows) that control creativity, length, and style.

2. The Four Types of LLMs

The report outlines a simple two-axis framework (lightweight vs. heavyweight, general vs. domain-specific), giving us four categories:

• Foundation Models – big, general-purpose systems (e.g., ChatGPT-4o).
• Expert Models – heavyweight tools trained on specialised data (e.g., AlphaFold).
• Compact Models – smaller, fast, efficient models that run on edge devices (e.g., Mistral 7B).
• Vertical Models – lightweight models tailored to industries or publishers (e.g., BloombergGPT).

Plus, there are task-specific LLMs designed for things like transcription, translation, and summarisation.

3. How Marketers Are Using AI Today

This isn’t theoretical — 88% of marketers already use AI daily. The biggest uses cluster into four areas:

• Creative & Content (copywriting, asset creation, versioning)
• Media Planning (optimising spend, testing creative)
• Research, Insights & Testing (summaries, trends, competitor analysis)
• Marketing Strategy (customer insights, scenario exploration)

To get the most out of an LLM, Epoch recommends the CRAFT method: Context, Role, Action, Format, Tone — a simple way to turn vague prompts into sharp, useful ones.

4. Concerns and Challenges

The report doesn’t shy away from the risks. Key issues include:

• Hallucinations (confidently wrong facts)
• Model Collapse (models degrading by training on AI-generated data)
• Prompt Injection (malicious instructions hidden in user input)
• Copyright & Privacy
• Bias and Alignment
• Job Displacement Anxiety

But it also stresses a crucial point: AI won’t replace humans — humans who use AI will replace humans who don’t.

5. How to Think About AI

Epoch explores a set of useful mental models — from “jagged intelligence” (AI is brilliant at some things and terrible at others) to “work slop” (the coming wave of low-effort AI content). These help us develop a more balanced, realistic perspective on what AI can and can’t do.

6. The Future of LLMs

LLMs are becoming more capable across four big areas:

• Knowledge: Personal, private, and retrieved data.
• Actions: Automating tasks, scheduling, and interacting with apps.
• Multimodality: Understanding and generating text, images, audio, and video.
• Autonomous Workflows & Agents: AIs that can take steps toward a goal without being micromanaged.

They’ll also increasingly live in dedicated hardware — wearables that “see” and “hear” the world.

7. The Human Skills That Will Matter More

As AI takes on structured tasks, timeless human qualities become even more valuable:

• Taste, intuition, and discernment
• Leadership, influence, connection, and empathy
• Morality, judgement, conviction, and initiative

In short: AI makes the human parts of our work more important, not less.

About Epoch

We are Epoch. We are a global brand agency. We build brands that build bonds.