reinforcement learning, large language models, LLM

It is a daunting task facing most businesses: Even those with the best engineers in the world and bottomless resources, find it onerous to devote a year to building a unique LLM and putting it into production.

Fine-tuning and refining language model systems has been a major challenge, and a logistical headache. Until, perhaps, now.

Startup Automorphic Inc. could change the landscape. It lets programmers easily create personalized, fine-tuned models in a matter of minutes.

The company’s flagship product, Conduct, lets developers quickly transform raw data into a bespoke language model that can be used in production. According to Automorphic, its technology lets developers quickly load and stack adapters that have been fine-tuned, allowing them to concentrate on feedback and tweaking their models. It claims Conduct’s iterative process gets LLMs into production faster without touching current code, because it is compatible with the OpenAI API.

By comparison, engineers at Alphabet Inc.’s Google and Facebook parent Meta Platforms Inc. devote 12 to 18 months laboriously transforming a model from the research phase to full-on production.

“The startup we’re building now is called Automorphic, and our goal is to help developers iterate and improve custom language models cheaply and efficiently,” Govind Gnanakumar, co-founder of Automorphic, said in an interview with Business Insider. “Right now, people are using these huge models like GPT-4 that contain trillions of parameters. In the future, though, we imagine that people will want to run more task-specific models that are significantly smaller.”

In the interview, Gnanakumar acknowledged major obstacles, starting with the “dark arts knowledge” of how to train and fine-tune models. Most JavaScript developers usually lack understanding about what’s happening at the bleeding edge, he said. Automorphic, conversely, is attempting to stay on top of the issue by reading the research and presenting it in accessible form, Gnanakumar added.

In the end, Conduct allows developers to quickly reshape raw data into a bespoke language model that can be used in production, saving time and money, to make domain-specific LLMs, according to Automorphic.