A generative artificial intelligence startup called OpenPipe Inc. is hoping to make the power of large language models more accessible after closing on a $6.7 million seed funding round.

Today’s round was led by Costanoa Ventures and saw participation from Y Combinator and angel investors, such as Logan Kilpatrick, the former head of developer relations at OpenAI, Alex Graveley, who created GitHub Inc.’s powerful Copilot tool, and Tom Preston-Werner, who founded GitHub.

Those prominent angel investors have apparently been sold on OpenPipe’s potential to help smaller developer teams without extensive generative AI skills build more capable generative AI applications than they could do by themselves. The startup explains that its platform is based on the simple and credible theory that smaller LLMs are just as powerful as much larger ones when they’re trained to specialize in a very specific task. Developers can create these specialized models by fine-tuning lightweight, general purpose LLMs on their own datasets, and that’s exactly what OpenPipe’s platform does.

OpenPipe explains that the fine-tuning process can dramatically improve the quality and correctness of generative AI responses. The challenge is that most developers lack the complex skills to do this fine-tuning effectively, as existing platforms are aimed at specialized engineers with extensive knowledge of machine learning practices. As such, companies without these specialists struggle to create lightweight LLMs.

OpenPipe’s platform is designed to simplify and speed up the process of fine-tuning AI models, and it does this in a cost-effective way, the company insists. It works by integrating into a customer’s codebase to collect their existing prompts, which are then used as the basis of its training data. The platform uses this data to fine-tune much smaller, more specialized LLMs that can only perform one, very specific task, but do it in a way that matches the performance of the biggest LLMs.

OpenPipe’s co-founder and Chief Executive Kyle Corbitt said one of the biggest challenges for developers is moving from proof-of-concept to usage at scale. “OpenPipe enables teams to easily convert their prompts into production-ready, fine-tuned models,” he said. “These fine-tuned models allow developers to deliver a much snappier experience at a much lower price point, all without sacrificing quality.”

OpenPipe collaborates with a number of established AI players, including the LLM debugging platform provider Langfuse, and the data collection startup Athena AI Inc. Other partners include OctoAIML Inc. and Fireworks AI Inc., which help to serve its fine-tuned models.

Andy Thurai, principal analyst and vice president at Constellation Research Inc., said OpenPipe is entering a very crowded marketplace for LLM fine-tuning tools that includes the likes of Predibase, Labellerr, Kili, Label Studio, Databricks Lakehouse, plus open-source versions such as LLaMa Factory, Ludwig and H20 LLM Studio. He said each of these tools tries to differentiate itself with additional capabilities such as custom workflow setups, multi-data format support, collaborative annotation capabilities and active learning workflows to improve the precision and efficiency of LLMs.

In addition, the more mature versions such as Databricks Lakehouse also provide integration with frameworks such as Ray and AI model cataloging platforms like HuggingFace.

“OpenPipe doesn’t have any of these bells and whistles but it will hopefully build them in due time,” Thurai said. “What it does have is a novel feature that allows it to collect the history of prompts which can be used as training data to fine-tune much smaller and more specialized models, based on the original larger LLM. It’s almost as if the user can identify a subset of the larger LLM and separate it from the rest of the model.”

Thurai said he believes OpenPipe can carve out a niche for itself despite the competitive nature of the fine-tuning marketplace, and the startup itself boasts an “active customer base” that includes both other startups, and also some big enterprises, which are using its fine-tuned LLMs to perform critical business functions. It gave examples such as translating and extracting information from government regulations, categorizing and filtering YouTube video based on transcripts, and routing inbound customer services requests to the most appropriate teams. All told, it reckons its smaller, fine-tuned models have saved customers over $3 million in inference costs since launching its platform in September.

Costanoa Venture’s Partner Tony Liu said OpenPipe promises to increase the accessibility of LLMs and make them available and affordable for companies of all sizes. “This investment underscores our belief in the profound impact OpenPipe will have in enhancing the productivity of developer teams and boosting their ability to launch competitive go-to-market applications with more precision,” he said.

Image: Wirestock/Freepik