Goodfire raises $150M in funding to enhance its AI interpretability platform

Goodfire Inc., a startup working to uncover how artificial intelligence models make decisions, has raised $150 million in funding.

B Capital led the Series B round. Goodfire stated in its funding announcement today that the deal also drew contributions from Salesforce Inc., former Google Chief Executive Eric Schmidt and more than a half-dozen others. The company is now valued at $1.25 billion.

A large language model consists of code snippets called artificial neurons. Those code snippets often have a simple design, but they interact in complex ways: Upwards of tens of thousands of neurons are involved in generating a prompt response. LLMs’ complexity makes it difficult to determine how they make decisions.

San Francisco-based Goodfire is working to ease the task. The company has built a platform that it calls a model design environment to map out LLMs’ internal components. Understanding how a model goes about processing data should make it easier to identify and fix flaws in its design.

The platform’s first component focuses on the LLM training phase. Goodfire says that researchers have often limited visibility into how a neural network learns new skills from its training dataset. The company’s platform maps out the training workflow and identifies flaws, which enables researchers to boost LLM output quality.

The second component of Goodfire’s platform monitors models’ performance once development is complete and they’re running in production. The company says that it reduced AI hallucinations by half in one recent project.

One of Goodfire’s first customers is healthcare AI startup Prima Mente Inc. The latter company has developed an AI model that analyzes particles called cfDNA fragments to detect Alzheimer’s disease. According to Goodfire, its researchers analyzed the algorithm and discovered that it mainly considers the length of cfDNA fragments when diagnosing patients. Existing scientific literature didn’t contain data on the diagnostic significance of cfDNA fragment length.

Last year, Goodfire developed a method called SPD to understand how LLMs process data. It works by identifying model components that may be involved in generating a prompt response and removing them one by one. If the removal of a component doesn’t affect an LLM’s output, researchers can conclude that it’s not involved in the processing workflow.

“Interpretability, for us, is the toolset for a new domain of science: a way to form hypotheses, run experiments and ultimately design intelligence rather than stumbling into it,” said Goodfire CEO Eric Ho.

Goodfire will use the proceeds from its funding round to enhance its platform and finance AI interoperability research projects. 

Image: Unsplash