Trends · 2026-09-16 · 1:42

Programming language models instead of prompting them

Name: Programming language models instead of prompting them
Uploaded: 2026-09-16
Duration: 1 min 42 s
Description: For anyone who prefers measuring to guessing, this is the tool that formalizes it — best aimed at one task that already runs on a brittle hand-written prompt.

DSPy treats prompts as parameters to be compiled, not hand-tuned text. You declare inputs and outputs, and optimizers tune the prompt (and examples) against a metric and a dataset.

In favour

Backend-agnostic: runs against local models via Ollama or vLLM, so optimization stays on your own hardware.
Turns prompt engineering from craft into something measurable and repeatable.

Worth watching

Not a quick win: you need a metric, an evaluation set, and some patience — it is a framework, not a chat box.
Value shows up on recurring, measurable tasks, not one-off prompts.

Our takeFor anyone who prefers measuring to guessing, this is the tool that formalizes it — best aimed at one task that already runs on a brittle hand-written prompt.

Source: DSPy (Stanford NLP, GitHub)

← All Focus posts

AI-narrated · AI-generated content

Transcript

A prompt is usually hand-tuned text. Written, rewritten, and judged by feel. So here is the question. What if you programmed the language model, instead of prompting it? That is the idea behind one framework. It treats prompts as parameters to be compiled, not as hand-tuned text. Less like writing copy. More like writing code. Here is how it works. You declare the inputs and the outputs. Then optimizers tune the prompt, and its examples, against a metric and a dataset. First, it is backend-agnostic. It can run against local models, served on your own machines. So the optimization stays on your own hardware, not on someone else's. Second, it turns prompt engineering from a craft into something measurable, and repeatable. Not a feeling that the new wording is better. A result you can run again. Two things to watch. It is not a quick win. You need a metric, an evaluation set, and some patience. This is a framework, not a chat box. And the value shows up on recurring, measurable tasks. Not on one-off prompts. A prompt you write once, for one question, is not where the effort pays back. So, our take. For anyone who prefers measuring to guessing, this is the tool that formalizes it. Best aimed at one task, one that already runs on a brittle, hand-written prompt.