Weights and biases raises $50M to advance LLMOps efforts for generative AI

Weights and Biases is looking to grow its generative AI and LLMops efforts with a new $50 million funding raise announced today.

The round was led by Daniel Gross and the former GitHub CEO Nat Friedman, with participation from Sapphire Ventures, Coatue, Insight Partners, Felicis, BOND and BloombergBeta. The new funding gives the San Francisco based startup a valuation of $1.25 billion.

Weights and Biases has been building out developer tools to enable what the company increasingly refers to as LLMops, that is, operations for effectively using and scaling gen AI large language models (LLMs). Back in April, the company announced a big rollout of LLMops tools, including an initial release of W&B Prompts, a tool to help organizations build and manage better prompt for LLMs. In June, Weights and Biases expanded its platform with new LLMOps capabilities for AI development and model monitoring.

The goal of the new funding is to continue expanding platform as the company continues to see growth opportunities.

“There’s a lot of enthusiasm around LLMs right now, but not a ton of people have really gotten products into production and I think that is going to take longer,” Lukas Biewald, CEO and cofounder of Weights and Biases, told VentureBeat. “I guess I’m more short term pessimistic on LLM products but I’m long term very optimistic, so I want to make sure that we are set up to build for the long term.”

What LLMOps is all about

According to Biewald, traditional software development is not about exploration, but rather, iterative development where new code features are added over time.

Machine learning (ML) and LLM development doesn’t necessarily follow the same sort of deterministic pattern, he said, as it’s not always clear how well things will work. As such, there is more exploration with LLM operations development today.

With ML operations (MLOps) workflows, he said, experiments by data scientists are about training a model and looking at lots of data. In contrast, experiments with LLMs are about looking at different data to fine tune a model and prompt engineering to figure out the best prompt to get the ideal output, as well as chaining multiple steps and even models together.

One of the first tools that Weights and Biases announced for LLMOps is the company’s W&B Prompts tool, which is now being enhanced to expand the scope of what it can do. The latest update to the prompt tool can log all of a user’s LLM requests and helps to enable larger production scale use cases.

What’s next for Weights and Biases

Weights and Biases also has a model registry and CI/CD (continuous integration/continuous deployment products to help developers build and iterate models.

Biewald noted that the two products today work well together. For example, every time a model gets promoted as a candidate for production deployment, a suite of tests can be run against the model candidate in an automated way.

“Now we’re in that phase of just improving it with all the feedback coming in and all the changes that the customers asked for,” he said.