Horizon

What Horizon is

Plain English explanations of how AI works, why it works, where it fails, and how to think about it inside a real business. Written for the people who decide about AI — founders, CXOs, ops leaders, function heads — not for the engineers building it.

We wrote this because we needed it. Comet Lab puts AI to work inside operations every day, and the same questions come up in every client conversation: fine-tuning or RAG? Wait or move? How do we know it’s working? Horizon answers them once, at the right depth, in one place.

Free. No signup. No paywall. No mailing list. Read what you need. Come back for the rest.

What’s here

Four modules, twenty-two chapters. Read them in order if you’re new to all this. Dip in anywhere if you’re not.

01 · Fundamentals

The mental model: what a model is, why it makes things up, the context window, prompts, how models are built, what they "know," how the words are used — and a single request traced end to end.

1. 1.1 What an AI model actually is →
2. 1.2 Why models sometimes make things up →
3. 1.3 The context window →
4. 1.4 Why prompts matter →
5. 1.5 How models are built — training →
6. 1.6 What models actually "know" →
7. 1.7 AI, ML, LLM — what the words really mean →
8. 1.8 What happens when you hit enter →

02 · The Spectrum of AI Tools

Models, products, and platforms — three different things people mean when they say AI. Who makes the big models. Open vs. closed. What changes beyond text. And how the four axes actually combine in a real operation.

1. 2.1 Models, products, and platforms →
2. 2.2 The big model families →
3. 2.3 Open vs. closed →
4. 2.4 Beyond text — image, audio, and video AI →
5. 2.5 How the layers actually meet →

03 · What AI Can and Can't Do

The honest list of what models are good at, what they're bad at, what tools and memory add, where agents fit in, and how you make a model better — prompting, RAG, fine-tuning.

1. 3.1 What AI is good at →
2. 3.2 What AI is bad at →
3. 3.3 When a model needs help — tools and memory →
4. 3.4 Agents — when AI takes action →
5. 3.5 Why code AI is unexpectedly good →
6. 3.6 The three knobs — prompting, RAG, and fine-tuning →

04 · Risk, Trust & Governance

The questions every CXO asks before approving AI: where your data goes, who owns the output, what can quietly go wrong, and how to keep humans in the loop.

1. 4.1 The four worries — what can go wrong →
2. 4.2 Where your data goes →
3. 4.3 IP, copyright, and provenance →
4. 4.4 Building trust — humans, audits, gates →

05 · Putting AI to Work

A method for scanning the operation to find candidates. The anatomy of a good fit. Build vs. buy. The shape of a good project. How you know it is working. What to watch for in production.

1. 5.1 Scanning your operation →
2. 5.2 Where AI fits →
3. 5.3 Build vs. buy →
4. 5.4 The shape of a good AI project →
5. 5.5 How you know it's working — measuring AI →
6. 5.6 What to watch for →

Who writes this

The people at Comet Lab. We deploy AI inside our own work. We built Horizon to help others think clearly about deploying it inside theirs.

Chapters are added as the field moves and as gaps surface. If something here is wrong, or could be plainer, write to us at hello@cometlab.in.