Adaptive, a startup founded by the team that built the open-source, large-language Falcon model after which partnered with open-source artificial intelligence company Hugging Face, has emerged from stealth with $20 million in an initial round of enterprise capital.
The company is working on technology that will make it easier for companies to train large language models (LLM) tailored to their specific needs.
The seed investment is led by Index Ventures with participation of ICONIQ Capital, Motier Ventures, Databricks Ventures, HuggingFund by Factorial and several other individual business angels. The company’s valuation was not disclosed, although the tech has previously published The Information reported that the startup was valued at $100 million in the financing round.
Adaptive is working on a way to improve a process often known as reinforcement learning from human feedback (RLHF). This process has proven to be key to taking LLMs, that are initially trained on vast amounts of text to predict the next word in a sentence, and making them more useful as engines powering chatbots like OpenAI’s ChatGPT.
RLHF involves collecting feedback from assessors on the quality of LLM responses. The LLM is then further trained to give answers that more closely resemble those that rate high. However, RLHF typically involved hiring contractors to evaluate the model, often by rating possible responses. This method is pricey – for instance, the cost of knowledge annotation contracts accounts for a good portion of the cost of coaching LLM-based chatbots – and the quality of feedback is usually too stringent to ensure good results in many LLM business use cases.
“It’s hard to get a model to do what you want,” said Julien Launay, co-founder and CEO of Adaptive.
Adaptive wants to enable LLM managers to usually and constantly learn from how the company’s employees or customers actually use the software. The user’s subsequent actions and reactions in response to LLM results are, in many cases, a much richer training signal than the thumbs up or down given by a paid evaluator.
Launay said Adaptive plans to offer a suite of solutions that capture how people interact with LLM responses after which allow the model to be trained and fine-tuned based on that data. Adaptive also provides a platform for running reinforcement learning algorithms that adapts the model, as this process is difficult for a lot of non-expert teams to implement. It also allows the company to select exactly what data it wants to collect, what goal the model wants to achieve, and what reinforcement learning algorithm it wants to use to conduct that training. This control helps companies higher manage cost-performance trade-offs, Launay said.
The platform may even help companies perform a process called reinforcement learning from AI feedback (RLAIF), in which a separate AI model analyzes the responses of the trained AI model. This can reduce training costs and result in higher coverage of coaching data than using human raters.
Adaptive will enter a market that is becoming increasingly crowded. RLHF training platforms are also offered by some big data labeling companies that have traditionally provided evaluators. These include Appen and Scale AI. Surge AI, CarperAI and Encord also offer similar tools. However, most RLHF tools will not be designed to capture model user preference data after the model has been deployed.
The technology Adaptive builds will run on any open-source LLM model or any model the company has built in-house. Open source models have gotten increasingly popular amongst companies that are on the lookout for greater control over each the results of generative AI models and ways to reduce the costs of generative AI applications. However, the startup’s technology is not going to allow enterprises to refine proprietary third-party models reminiscent of those available from OpenAI, Google, Anthropic and Cohere. “We need access to the model weights,” Launay said.
The Adaptive platform is designed to help customers test the performance of various LLMs against one another and help them monitor the performance of those models after implementation. Adaptive develops dashboards and metrics that can link LLM results to key business metrics, reminiscent of the successful resolution of a customer inquiry.
He said Adaptive already has several customers using its platform, although he declined to name them. The company, which currently has just nine employees, said it plans to use the new enterprise capital funding to expand its teams each in Paris, where it relies, and in New York, with a particular give attention to go-to-market teams and sales .
Launay previously worked at a Paris-based AI hardware startup with Adaptive co-founder Daniel Hesslow, now the startup’s chief scientist. The two later ended up working with co-founder Baptiste Pannier, now Adaptive’s chief technology officer, as a part of the team that built the open-source Falcon LLM family of models at the Abu Dhabi Institute for Technology Innovation. The Falcon models impressed people with their performance considering their size and the modern training techniques their creators used. Falcon models usually top Hugging Face’s rankings for performance and model popularity.
The team then moved on to work at Hugging Face, which each builds its own open-source AI models and offers a popular repository of other open-source models.
Bryan Offutt, a partner at Index Ventures who led the investment in Adaptive, says he was impressed by the company’s founding team’s combination of technical knowledge and understanding of business needs, in addition to its energy, which he described as “contagious.” He said the problem the team is trying to solve – how to tune a generative AI model to user preferences – is a technical challenge that many companies struggle with.
He said the challenge for Adaptive in the future might be to work with clients to find ways to motivate LLM users to provide feedback that might be most useful to training. If a person fully explains why they find the model’s response helpful or unhelpful, this is amazingly priceless data for refining the model. However, having to provide such a detailed feedback for every model response is time-consuming and sure annoys users. Therefore, Adaptive will need to find ways to work with its clients to balance the need for feedback with the burden this places on LLM users, Offutt said.
Correction, March 11: An earlier version of this story mischaracterized the typical process used in RLHF. Evaluators typically evaluate the possible answers generated by the LLM, slightly than giving a easy thumbs up or thumbs down. An earlier version of this story also misidentified the city where Adaptive’s founders previously worked on their hardware startup. This was Paris, not Amsterdam.
Credit : fortune.com
