Diogo Almeida

At TypeSafe, Almeida argues that fluent chat does not automatically produce dependable automation. In his launch essay, he introduces Jev as a model that accepts unstructured state but returns values within a predefined output structure, accompanied by probabilities. TypeSafe calls its training approach Reinforcement Learning for Calibrated Decisions. The design prioritizes composing small decisions inside ordinary programs, such as routing, extraction and classification. Its guarantee about matching an output schema is distinct from whether a judgment about the world is correct.[2]

Almeida is a co-author of the InstructGPT paper, which describes a training pipeline beginning with human demonstrations, followed by a model of human preferences and reinforcement learning against that feedback. The study found that people preferred outputs from the smaller instruction-tuned model to those from a much larger base GPT-3 model on its test distribution. It also documents remaining mistakes and limitations. Almeida later appears among the authors of the GPT-4 technical report. These were team research efforts that helped make language-model interfaces more responsive to instructions.[3][4]

In his Latent Space conversation, Almeida traces his shift from instruction-following assistants toward programmable decision systems. He emphasizes uncertainty: software needs to distinguish a confident answer from an ambiguous one, rather than receive the same fluent response in both cases. He also argues that choosing training tasks and data determines what a model becomes useful for. His public education record lists Georgia Institute of Technology studies from 2014 to 2016.[5][6]