Jev inventor TypeSafe closes $870 million round at $7.5 billion valuation

TypeSafe Inc., the inventor of the Jev artificial intelligence model, today announced that it has raised $870 million in funding at a valuation of $7.5 billion.

Andreessen Horowitz led the round with contributions from Sequoia Capital, DCVC and unnamed angel investors. The cash injection comes less than a month after TypeSafe launched Jev. The company’s funding momentum reflects the growing popularity of its model: It has already been adopted by about a third of the Fortune 500.

When an enterprise application needs to perform a task using a large language model, it automatically sends a description of the task to the model. The LLM then responds with its output. This output is often in the form of natural language text. Applications must condense the natural language text into a structured, standardized format before they can use it.

Jev owes its popularity in part to the fact that it skips the final step. When an application asks Jev to perform a task, the model generates structured output instead of natural language text. This eliminates the need for applications to reformat and compress Jev’s output before use.

It only supports three types of requests. Applications can ask the model to answer a question “yes” or “no,” select an item from a list, or generate a score. Developers can customize what this score measures. For example, Jev can be configured to rate cybersecurity alerts based on their severity or quantify the urgency of a support ticket.

The fact that the model’s sparse output is easy for applications to process reduces the amount of data preparation code that developers need to write. This means software projects can be completed more quickly. Additionally, simplifying an application’s code base reduces the risk of errors.

Jev also increases application reliability in other ways. When the model generates a score or selects from a list of options provided by the user, it outputs a number indicating its confidence in the accuracy of the answer. Applications can use this number to mitigate the effects of hallucinations.

TypeSafe says it developed Jev for calibrated decisions using a new training approach called reinforcement learning. It is a variant of reinforcement learning, a popular LLM training method. TypeSafe has also developed a new model architecture.

The company claims that its customized technologies enable Jev to process user requests in less than 700 milliseconds. This makes the model up to 200 times faster than some frontier LLMs. Additionally, according to TypeSafe, it is up to 100 times more cost-effective.

Jev is part of a planned model series called System One. TypeSafe will use the proceeds from its funding round to further expand its range. Additionally, the company plans to introduce unspecified “enterprise features” designed to make it easier for large organizations to use its models.

Picture: Pixabay