AI engineering · September 2026
6 AI Classics Behind Better Prompts and Model Choices
Your next AI fix may be hiding in the original research. Revisit 6 landmark works for better prompts, useful retrieval, and the limits people leave out.
Adding examples, retrieving documents and asking for intermediate reasoning all have research behind them. They also have conditions that disappear when the advice becomes a prompting tip.
These 5 research papers and 1 essay help builders understand the choices underneath an AI application, from how a transformer handles text to how training data and model size compete for a fixed compute budget. The useful question is what each result establishes, and whether its conditions resemble your own task.
Chain-of-thought prompting, for instance, hurt some smaller models in the 2022 experiments. The RAG paper tested whether retrieval improved answers in specific comparisons. Neither gives you a guarantee for your own app.
GPT-3 examines examples supplied inside a prompt. Chinchilla compares training choices under a compute budget. The Bitter Lesson makes a historical argument about methods that can use increasing computing power. It is an essay, so its argument deserves a different reading from an experimental result.
Use the written and audio explanations below to find the work relevant to your decision. Each links to the original, including the comparisons and limitations behind the familiar headline.
In this playlist
Attention Is All You Need: Why Drop Recurrence?
Wondering why longer prompts cost more? Start with the transformer paper. Its design helps explain how today’s models handle your text.
5 min read10 min listen
GPT-3: What Examples in a Prompt Can Teach a Model
Your prompt works once, then fails on the next example. See how GPT-3 used examples inside a prompt, and where that approach struggled.
4 min read7 min listen
When Chain-of-Thought Prompting Helped, and When It Hurt
Intermediate reasoning helped some of the models tested in 2022 and hurt smaller ones. Read the task and model details before treating the result as a rule for today's models.
4 min read8 min listen
The RAG Paper: What Retrieval Added to Generation
Will giving your model documents make its answers better? Read the original retrieval experiment, including what its fact-checking comparisons actually measured.
4 min read7 min listen
Chinchilla: What a Smaller, Better-Trained Model Showed
A smaller model can win. See why the Chinchilla researchers trained a 70-billion-parameter model on far more text than the 280-billion-parameter Gopher model.
4 min read6 min listen
The Bitter Lesson: What Sutton Means by Scale
Before you spend weeks adding clever rules, read Rich Sutton’s 2019 essay. It asks why methods that use more computing power so often win over time.
3 min read4 min listen