TechniqueFrameworkArtificial intelligenceSoftware engineering
Without evaluation you have a demo, not a system
Dave Ebbelaar · How to Systematically Set Up LLM Evals
The idea
A model-based feature that works on the examples its builder tried is a demo. It becomes a system when there is a test set, a metric and a way of running one against the other every time something changes. Dave Ebbelaar defines what an eval is, the levels at which they operate, from unit-style checks on single outputs to end-to-end judgements, and where a human still has to sit, because a model judging another model has failure modes of its own. The discipline is the same one that makes software testable, applied to a component whose behaviour is statistical rather than deterministic.
Build the test set before the feature, measure every change against it, and keep a person in the loop where the metric cannot be trusted.
When it works
Any AI product discussion, any system design that includes a model, and the question of how you know an AI feature works.
When it may fail
Evals measure what you thought to test, and a model can pass them while failing users. They are necessary rather than sufficient, which is also true of every other kind of test.
How to use it
For an AI feature, write twenty inputs with the answers you expect, and run them after every prompt change. If you cannot write the expected answers, you do not yet know what the feature is for.
Where it came from
Video How to Systematically Set Up LLM Evals · Dave Ebbelaar
What the source says. Dave Ebbelaar defines what an eval is, the levels of evaluation, and where a human still has to sit in the loop.
Connected ideas
Puts it to work
A system optimises the objective you wrote, not the one you meantWhatever you can measure and reward will be maximised literally, including the gaps between the measure and the goal.
Says something similar
The context window is not the constraint people think it is; attention across it isA long window degrades before it fills; curate and order what goes in rather than relying on size.
Says something similar
Precision against recall: which error you can afford is a business decisionDecide which mistake is more expensive before tuning a model; the threshold is a business choice the statistics only express.
Concepts
Put it to work on L3VLUP
More like this
- FrameworkArtificial intelligenceVideo located
A model never sees words, only tokens, which explains most of its odd behaviour
Reason about a model in tokens: that is what it reads, what fills its window and what you pay for.
From Most Devs Don’t Understand How LLM Tokens Work · Matt Pocock
- ArgumentArtificial intelligenceBook
A model trained on a biased outcome will reproduce it at scale
A model predicts the past decisions in its training data; if those were biased, the model is a bias machine with a straight face.
From Weapons of Math Destruction · Cathy O’Neil
- TechniqueArtificial intelligenceVideo located
A prompt is a brief: task, role, background and output format
Write to a model as you would brief a capable junior: the task, their role, the context they lack, and what the output should look like.
From Prompting 101 · Anthropic
- Career lessonCareersDocumentary
A skill is only worth what the market structure around it allows
Separate what you can do from the structure that makes it valuable, and watch the structure.
From Floored · 2009