Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

Launch HN: Augento (YC W25) – Fine-tune your agents with reinforcement learning

Mar 31, 2025 - news.ycombinator.com
Augento is a platform offering fine-tuning as a service for language models, inspired by the DeepSeek R1 paper's approach of using reinforcement learning for post-training. Users can connect their agents to Augento, provide feedback on when the agent's outputs are right or wrong, and receive a fine-tuned open-source model optimized for specific tasks. Instead of requiring large datasets for supervised fine-tuning, Augento uses a reward function to evaluate the model's outputs, making it suitable for verifiable domains like coding, tool specialization, browser navigation, and robot control. The platform also plans to introduce an "alignment mode" that allows users to provide high-level feedback without needing formal reward functions.

Augento's service is not open source, but it fine-tunes open-source language models, offering an alternative to OpenAI's reinforcement fine-tuning API with more customization options. Users are charged a monthly fee plus training and inference costs, and the platform is accessible for self-service with an initial $20 in training credits. The company aims to improve agent performance in practical applications by leveraging reinforcement learning techniques and seeks user feedback to refine their offering.

Key takeaways:

  • Augento offers a fine-tuning service for language models using reinforcement learning, allowing users to optimize models for specific tasks by providing a reward function.
  • The platform supports various use cases, such as coding agents, tool specialization, browser navigation, and robot control, by fine-tuning models based on task-specific criteria.
  • Augento plans to introduce an "alignment mode" that allows users to provide high-level feedback on agent failures without needing to write formal reward functions.
  • The service is not open source but fine-tunes open-source language models, offering a customizable alternative to OpenAI's reinforcement fine-tuning API, with a pricing model based on training and inference costs.
View Full Article

Comments (0)

Be the first to comment!