Adaptive, a startup established by the team behind the creation of the open-source large language model Falcon, who previously collaborated at the open-source AI firm Hugging Face, has come out of stealth mode with an initial $20 million in venture capital funding.
The company’s primary focus is on developing technology that simplifies the process for businesses to train customized large language models (LLMs) to suit their specific requirements.
Index Ventures is spearheading the seed investment, with contributions from ICONIQ Capital, Motier Ventures, Databricks Ventures, HuggingFund by Factorial, and various individual angel investors. While the company’s valuation remains undisclosed, reports from The Information previously indicated that the funding round placed the startup’s value at $100 million.
Adaptive is dedicated to enhancing a technique known as reinforcement learning from human feedback (RLHF). This method plays a crucial role in advancing LLMs, which are initially trained on vast amounts of text to predict the next word in a sentence, transforming them into more effective engines that power chatbots like OpenAI’s ChatGPT.
Unlike the conventional RLHF approach, which typically involves hiring contractors to evaluate models using simplistic ratings like thumbs up or thumbs down, Adaptive aims to revolutionize this process. By enabling LLMs to continuously learn from real interactions with a company’s employees or customers, Adaptive believes that this organic feedback loop offers a richer training signal compared to basic human evaluations.
Adaptive’s strategy involves providing a comprehensive solution that captures user interactions with LLM responses, facilitating model training and refinement based on this data. The platform also supports the implementation of reinforcement learning algorithms tailored to the specific objectives of businesses, allowing them to make informed decisions regarding performance and cost trade-offs.
Moreover, Adaptive plans to introduce a process called reinforcement learning from AI feedback (RLAIF), where an AI model critiques the responses of the model undergoing training. This innovative approach not only reduces training costs but also diversifies the range of training data beyond human evaluation.
While entering a competitive market, Adaptive’s technology is designed to complement any open-source LLM model or proprietary models developed by businesses. The platform will enable customers to compare LLM performance, monitor post-deployment metrics, and assess the impact on key business outcomes.
With a focus on expanding its teams in Paris and New York, Adaptive intends to leverage the recent funding to enhance its “go-to-market” and sales strategies. The company’s founders, with a background in developing the Falcon LLM models, have garnered recognition for their technical prowess and understanding of business requirements, attracting investments from Index Ventures.
Moving forward, Adaptive faces the challenge of incentivizing users to provide valuable feedback without burdening them excessively. Balancing the need for detailed feedback with user convenience will be crucial in refining generative AI models effectively.
