CHAPTER 01 / 6 MIN READ
Beyond the answer
Start with the boundary between saying something and doing something.
START HERE
You ask a model to check whether a train is running. It gives a plausible timetable. But no timetable was opened. Did it answer the question—or perform the task?
A model generates a continuation. A harness decides how that continuation becomes part of a task.
1.The model sees an input, not your whole world.
At a particular call, a language model receives a bounded input and produces output. Your files, a browser tab, yesterday’s decisions, and the result of a command are available only if the surrounding system makes them available.
A harness is that surrounding machinery: it constructs the input, exposes actions, runs them, returns observations, and decides whether to continue. The same underlying model can therefore encounter very different information and very different opportunities to act.
2.The first control surface is context.
GPT-3’s few-shot setting adapts behaviour with examples in the prompt, without changing weights at inference. Chain-of-Thought studies examples that include intermediate steps. In both cases, the intervention is in what the model sees.
These techniques do not, by themselves, open a browser or validate a result. A detailed explanation can still be detached from the world. More intermediate text is not automatically more evidence.
3.A tool call crosses the boundary.
With tools, a generated action becomes an operation: search, open a page, run a calculation. The tool returns an observation that can enter the next model call. WebGPT uses a text browser and collected references; Toolformer studies learning when and how to call APIs.
There are now two questions to ask about any answer: what did the model infer, and what did the environment establish? Keep that distinction throughout this book.
The train question
BEFORE
“The 09:10 train should be running.” The answer comes from available context and learned patterns.
AFTER
Search the operator → open the service notice → return a time-stamped status with its source.
Tools can ground a response in external observations. They do not guarantee that the model searches the right source or interprets it correctly.
PAUSE & PREDICT
Which change requires a tool boundary?
The title of the source talk is a provocation, not a universal ranking. Model capability and harness design interact; this corpus does not isolate which always matters more.
GO TO THE SOURCE
GPT-3 / Few-Shot Learners2020 ↗Chain-of-Thought2022 ↗WebGPT2021 ↗Toolformer2023 ↗This chapter is an original teaching synthesis. Open each research note for source links, findings, and classification caveats.
TAKE IT INTO YOUR OWN WORK
Pick a task you delegated recently. List which facts were in the prompt and which facts had to be observed.