Simón Fishman

Simón Fishman, also known professionally as Simón Posada Fishman, is an AI engineer and researcher whose contributions include the OpenAI Cookbook, GPT-4’s launch, and GDPval, an evaluation of artificial intelligence on realistic professional work. His work addresses two practical questions: how developers combine emerging model capabilities into useful applications, and how to determine whether those applications produce genuinely useful results.

Fishman attended Pomona College and was selected to speak at TEDx Claremont Colleges in 2016. At OpenAI, he worked as a solutions engineer supporting customers, contributed to GPT-4 launch partnerships and product operations, and joined the OpenAI Cookbook’s contributor community.

His 2023 multimodal demonstrations with Logan Kilpatrick explored how GPT-4 with Vision, DALL·E 3, and Whisper could cooperate before their capabilities were integrated into a single model. Three ideas shaped that work:

• Text-mediated multimodal architecture: Image descriptions and speech transcripts can serve as shared representations connecting otherwise separate vision, audio, language, and image-generation systems.
• Visual feedback loops: A vision model can compare generated images against their originals, identify discrepancies, and refine subsequent generation prompts, although imperfect fidelity still limits the results.
• Video understanding beyond transcription: Combining sampled video frames with speech transcripts preserves visual information that spoken narration and transcript-only summaries omit.

In 2025, Fishman co-authored the GDPval research paper, which evaluates model-generated professional deliverables across 44 occupations and nine major U.S. economic sectors. Its public release includes 220 tasks involving artifacts such as documents, spreadsheets, and diagrams; Fishman also contributed to the official GDPval dataset. The benchmark emphasizes useful workplace output while acknowledging that one-shot tests cannot capture the full complexity of iterative professional work. He has also highlighted his team’s public evaluation work, extending his focus from building AI applications to measuring their real-world performance.