AI Performance

Tests, comparisons and methods to improve the quality, accuracy, consistency and efficiency of responses.

3articles

Painel futurista horizontal com três interfaces de modelos de IA, controles de temperature, top_p e top_k e linhas ramificadas que representam a variabilidade das respostas.

AI Performance30/08/202613 min read

Temperature, top_p, and top_k in 2026: what still works in OpenAI, Claude, and Gemini

In 2026, classic sampling controls are no longer universal. See where they still work, where they cause errors, and which settings control variability now.

Read article

Estação de trabalho com painéis comparando um modelo anterior e um novo modelo, gráfico de regressão e etapas de diagnóstico de prompt, schema, ferramentas, contexto e parâmetros.

AI Performance16/08/202612 min read

Did your prompt get worse after switching models? How to diagnose regression before rewriting everything

If a workflow changed after a model change, preserve the prompt as a baseline, isolate variables and test format, tools, context and parameters before rewriting.

Read article

Mesa de trabalho com notebook exibindo um fluxo de avaliação de prompts, rubrica impressa, casos de teste e anotações sobre comparação A/B e regressão.

AI Performance16/08/202611 min read

How to know if a prompt really improved: test cases, rubric and regression without evaluating on feeling

A practical method for comparing prompt versions with representative cases, rubric defined before responses, blind A/B testing, and recording regressions.

Read article