Jev Neural Network by TypeSafe AI

TypeSafe AI raised $870 million in funding at a company valuation of $7.5 billion. The Jev model was launched on September 15 and represents a fundamentally new approach to artificial intelligence.

The model's architecture is based on transformers, but Jev is not a large language model (LLM). It does not generate text: the system outputs only probabilities, which the company calls calibrated decisions.

How it works without hallucinations

Its mechanism differs from traditional chatbots. Developers predefine answer options, and the model selects the most suitable one. This eliminates the possibility of fabricated information or hallucinations.

• Training was conducted exclusively on synthetic data
• Output tokens are provided for free
• Input tokens are billed per billion

The exact model architecture remains a trade secret and is not disclosed publicly. A link to the technical documentation is available on the company's official website.

Typical tasks for Jev include sorting incoming emails, checking commands for security, detecting jailbreak attempts, and automatically selecting the optimal model for task execution.

One of the startup's co-founders is Diogo Almeida, who previously worked at OpenAI and contributed to the creation of ChatGPT. According to the company, a third of Fortune 500 firms already use the service.