startup-stack

Infographics — what AI is

Eight pictures covering the foundations: what a language model actually is, how one gets built, what people mean by weights, why the models keep improving, and what changes when the same model is put in a different place. Nothing here is specific to this repository. Take any of it for a workshop.

Click an image to open it at full size.

What this thing actually is

The plain description, with no metaphor that misleads. A model predicts the next piece of text, repeatedly. Everything useful about it and everything dangerous about it follow from that one sentence.

A plain description of a large language model: it predicts the next piece of text, one piece at a time. What follows from that — fluency, format, patience — and what also follows: it produces the plausible continuation rather than the true one.
A plain description of a large language model: it predicts the next piece of text, one piece at a time. What follows from that — fluency, format, patience — and what also follows: it produces the plausible continuation rather than the true one.

How a model gets made

Four stages — gather, pre-train, fine-tune, align — and the reason a new version arrives faster than the cost of the first one would suggest.

The four stages of building a language model, from an enormous cleaned corpus through pre-training, fine-tuning on good conversations, and alignment against human preference.
The four stages of building a language model, from an enormous cleaned corpus through pre-training, fine-tuning on good conversations, and alignment against human preference.

What weights actually mean

The part of a model that is the model: a very large list of numbers, none of them written by a person, none of them readable afterwards. This is also where open weights and closed weights get explained, which is the choice that matters if you work under a confidentiality obligation.

Weights explained as the numbers on the connections in a network, adjusted trillions of times during training. Open weights versus closed weights, and why the trade is capability against control.
Weights explained as the numbers on the connections in a network, adjusted trillions of times during training. Open weights versus closed weights, and why the trade is capability against control.

The loops that make models improve

Four feedback cycles — human preference, AI feedback, verifiable rewards, synthetic data — and an honest note on where each one stops working.

Four self-improvement loops in model training, and the limits of each: no compiler exists for 'is this good strategic advice'.
Four self-improvement loops in model training, and the limits of each: no compiler exists for 'is this good strategic advice'.

The same model in four different places

Website, desktop app with connectors, editor, API. The model is identical in all four. What differs is what it can see and what it is allowed to do — and that difference is most of the reason two people get different results from the same question.

One model shown across four surfaces — web chat, desktop app with connectors, an IDE, and the API — with what each one can see and what it may do.
One model shown across four surfaces — web chat, desktop app with connectors, an IDE, and the API — with what each one can see and what it may do.

What agentic actually means

The difference between asking for text and delegating a task, and the inner loop — plan, act, observe, check — that makes the second one possible.

Assistant versus agent, and the plan-act-observe-check loop at the centre of agentic work, with the three things it requires: tools, visibility of results, and a stopping condition.
Assistant versus agent, and the plan-act-observe-check loop at the centre of agentic work, with the three things it requires: tools, visibility of results, and a stopping condition.

Which model should I use

Three tiers, a separate dial for thinking effort, and the rule of thumb that saves the most money: start in the middle and move for a reason.

Model tiers from small and fast to largest, the separate effort dial, and the context window explained as the desk the model works at.
Model tiers from small and fast to largest, the separate effort dial, and the context window explained as the desk the model works at.

Where your words actually go

What happens between pressing enter and getting an answer, and why it explains the three things that puzzle people most: rising cost in long threads, forgetting the start of a conversation, and why an index is a cost decision rather than a tidiness preference.

The path from a typed question through context assembly, tokenisation, the model, and streamed output — and why a long conversation gets slower and dearer with every turn.
The path from a typed question through context assembly, tokenisation, the model, and streamed output — and why a long conversation gets slower and dearer with every turn.

Next: the tools — terminal, Git, Node, databases and the rest of the words.

Generated from docs/infographics-ai.md in the repository. Edit the markdown, not this page.