Adaptive’s co-founder, Baptiste Pannier, is the company’s chief technology officer. CEO Julian Launay. Daniel Hesslow, its Founding Research Scientist; Photo courtesy: Adaptive
Adaptive, a startup founded by a team that built the open source large-scale language model Falcon and then collaborated on the open source AI company Hugging Face, emerges from stealth with $20 million in its first venture capital round Did.
The company works on technology that allows enterprises to easily train large-scale language models (LLMs) tailored to their specific needs.
The seed investment was led by Index Ventures with participation from ICONIQ Capital, Motier Ventures, Databricks Ventures, Factorial’s HuggingFund, and select private angel investors. The company’s valuation was not disclosed, but technology publication The Information previously reported that this funding round valued the company at $100 million.
Adaptive is working on ways to improve the process known as reinforcement learning from human feedback (RLHF). This process employs an LLM that is first trained from vast amounts of text to predict the next word in a sentence, and uses his LLM as the engine that powers chatbots such as OpenAI’s ChatGPT. It’s the key to making it useful.
RLHF involves collecting feedback from human raters regarding the quality of LLM responses. The LLM is then further trained to provide answers that are close to those rated highly by the raters. However, RLHF typically hired contractors to evaluate the models, often using a simple thumbs-up or thumbs-up to evaluate the answers. This method is expensive, with the cost of data annotation contracts accounting for a significant portion of the training cost of his LLM-based chatbot, for example. Additionally, the quality of the feedback may be too coarse to produce good results in many business use cases for LLM. .
“Getting a model to behave the way you want it to behave is difficult,” says Julien Launay, co-founder and CEO of Adaptive.
Adaptive wants LLMs to be able to regularly and continuously learn from how their own employees and customers actually interact with the software. The next actions and responses a user takes in response to an LLM’s output are often a much richer training signal than the ratings and ratings given by paid raters.
Launay said Adaptive plans to offer a solution package that captures the way people interact with LLM responses and allows them to train and fine-tune models from this data. Adaptive also provides a platform for running reinforcement learning algorithms that adjust models. This is because this process is difficult to implement for many non-specialist teams. It also allows companies to choose exactly what data they want to collect, what goals they want to achieve with their models, and which reinforcement learning algorithms they want to use for this training. This control allows companies to better manage cost and performance trade-offs, Roney said.
The platform also helps companies perform a process called Reinforcement Learning from AI Feedback (RLAIF). In this process, another AI model critiques the responses of the AI model being trained. This reduces training costs and provides a wider range of training data than using human raters.
Adaptive will be entering an increasingly crowded market. Platforms for RLHF training are also provided by some of the big data labeling companies that traditionally provided human raters. These include Appen and Scale AI. Similar tools are also provided by Surge AI, CarperAI, and Encord. However, most of these RLHF tools are not designed to retrieve configuration data from model users after the model is deployed.
The technology Adaptive is building works on top of open source LLM models as well as models that companies build themselves. Open source models are becoming increasingly popular among companies that want more control over both the output of their generative AI models and how to reduce the cost of their generative AI applications. However, the startup’s technology does not allow companies to fine-tune third-party proprietary models available from OpenAI, Google, Anthropic, and Cohere. “We need access to the weights of the model,” says Launay.
Adaptive’s platform is designed to allow customers to test the performance of different LLMs against each other and monitor how these models perform after deployment. Adaptive develops dashboards and metrics that can correlate LLM output to key business metrics, such as whether a customer query was successfully resolved.
He said Adaptive already has customers using its platform, although he declined to name them. The company, which currently has just nine employees, will use new venture capital funding to expand its team in both its home base of Paris and New York, with a focus on ‘go-to-market’ and sales teams. He said he plans to do so. .
Roney previously worked at an AI hardware startup in Amsterdam with Daniel Hesslow, co-founder of Adaptive and current principal researcher at the startup. The two then went on to work with co-founder Baptiste Pannier, now Adaptive’s chief technology officer, as part of the team that built the Falcon LLM family of open source models at Abu Dhabi’s Technology Innovation Institute. became. The Falcon model impressed people with its performance for its size and the innovative training techniques used by its developers. Falcon models consistently top the leaderboards where Hugging Face maintains the model’s performance and popularity.
The team then went to work at Hugging Face, which builds its own open source AI models and provides a popular repository for other open source models.
Brian Offutt, a partner at Index Ventures who led the investment in Adaptive, described the combination of technical expertise and understanding of business needs, and the energy shown by the company’s founding team, as “infectious.” He said he was impressed with what he had done. He said the problem the team is trying to solve – how to tailor generative AI models to user preferences – is a technical challenge that many companies struggle with.
He said Adaptive’s challenge going forward will be to work with customers to find ways to ensure those using LLMs can provide the most valuable feedback to training. If you fully explain why you think a model’s response is helpful or unhelpful, that’s invaluable data for improving your model. However, having to provide this kind of detailed feedback for every model response can be time consuming and frustrating for users. As a result, Adaptive needs to work with customers to find ways to balance the need for feedback with the burden placed on LLM users, Offutt said.
