Eric Hartford

Hartford describes a software-engineering career spanning 25 years before his move into AI research. In 2026 he introduced With, a compiled systems language intended for software including AI infrastructure. Its design uses scoped access to data and aims to make native programming approachable to developers accustomed to Python.[1]

Publish the instruction data as well as the model

Dolphin began as an effort to reproduce the instruction-data approach described in Microsoft’s Orca paper. Hartford and collaborators collected responses to FLAN prompts, removed duplicates and released the resulting data. His original announcement describes filtering refusals and related responses so that users could apply their own alignment afterward. Dataset and base-model licenses must be checked separately; the dataset’s license does not automatically govern every checkpoint.[2]

Combine coding weights with visual inputs

For Devstral-Vision-Small-2507, Hartford describes replacing the language-model weights inside a Mistral multimodal model while retaining its vision encoder, projector and image-token handling. The resulting checkpoint targets screenshot-assisted coding. Its card also states that it was not specifically fine-tuned on vision-to-code tasks, an important limit when evaluating the combination.[3]