How context shapes
a model’s response
Why the same request can produce different results, and how to make the task clearer.
Read the articleAN OPEN NOTEBOOK ON ARTIFICIAL INTELLIGENCE
Exploring
the potential
of AI.
From compelling answers to deeper understanding.
Explore how models work, where they fall short,
and what we can build with them.
FIG. 01 / LATENT CONNECTIONS+
IN FOCUSLanguage models ↗Knowledge retrieval ↗Answer quality ↗Humans and AI ↗
01 / A CLOSER LOOK
Notes on intelligence
One idea. A clear explanation.
An experiment you can repeat.
How context shapes
From document search
to answers with sources
Connect generation to your own knowledge base, and understand what still needs checking.
Read the articleA good answer.
But how do you know?
Move beyond first impressions to repeatable comparisons on the tasks that matter to you.
Read the article02 / A SMALL EXPERIMENT
A little more
unpredictability.
Temperature changes the probabilities of what comes next. A lower value favors the leading option; a higher value gives other possibilities more room.
Move the slider to see how the distribution changes.
Read about text generationTHE NEXT WORDINTERACTIVE EXAMPLE
“Research begins with …”
a question
a hypothesis
observation
wonder
More predictableMore varied
A teaching example: softmax over four fixed scores. This illustrates the principle; it is not output from a neural network.
03 / OUR APPROACH
Start curious. Then test.
A convincing answer starts a conversation.
A repeatable result builds a case.
Ask a question
Choose a specific task and decide what a good result looks like before you begin.
Change one thing
Compare prompts, context or model settings while keeping the other conditions the same.
Check the conclusion
Repeat the experiment, examine the errors and report the limitations alongside the results.
ABOUT BLD AI
Complex technology.
Human curiosity.
BLD AI is a journal for anyone curious about how artificial intelligence works and where it can help. It brings together explanations, original research and ideas for hands-on experiments.
Our principle is simple: distinguish observations from assumptions, explain the limitations and give readers a way to check the conclusions.
Start with an article