Bespoke Labs, an AI startup focused on enhancing the post-training stage of artificial intelligence projects, announced it has raised $40 million in a two-part funding effort. The majority of the funds, approximately $31.75 million, came through a Series A round led by Wing VC with participation from Mayfield, The House Fund, and employees at leading tech companies like Anthropic PBC. Earlier contributions included $8.25 million from investors such as Google DeepMind chief scientist Jeff Dean.
The company targets the critical post-training phase where AI models develop advanced reasoning and task execution skills after the initial pre-training has equipped them with basic knowledge. This second stage often involves reinforcement learning, where models improve through trial, error, and rewards based on task performance in simulated environments tailored to the model’s intended use.
Bespoke Labs provides a platform designed to simplify and accelerate the creation of these reinforcement learning environments. Leveraging automation and insights from human experts, the platform generates simulations much faster than conventional manual methods. These environments are then managed within a specialized sandbox layer to reduce latency and increase processing efficiency during training.
In addition to managing simulations, the platform features automated optimization for AI output quality. It employs tools like GEPA, an open-source initiative released by Bespoke last year that automates prompt engineering—a crucial process that identifies the best prompts and formats to maximize AI performance.
Beyond reinforcement learning, Bespoke also prioritizes supervised fine-tuning (SFP), a method where AI models refine their outputs by learning from curated prompt-answer datasets. Acknowledging the resource-heavy nature of assembling such datasets, the company introduced OpenThoughts, a publicly available dataset containing over a million sample prompts and responses designed to assist AI fine-tuning efforts.