TypeSafe AI has launched with $40 million in funding and released Jev, a non-generative AI model designed to return typed probabilistic decisions for machine-to-machine interaction. Founded by former OpenAI researcher Diogo Almeida, the startup claims Jev operates at speeds of 70 to 500 milliseconds. Jev is highly cost-effective at $0.042 per million input tokens, though it faces competition from OpenAI's Structured Outputs and can still make incorrect choices within its preset options.
Jev model architecture
- ▪Jev is a System One model that utilizes an architecture called Reinforcement Learning for Calibrated Decisions (RLCD).
- ▪TypeSafe AI co-founder and CEO Diogo Almeida, a former OpenAI researcher, is a co-inventor of Reinforcement Learning from Human Feedback (RLHF) and ChatGPT.
- ▪Jev's architecture computes several outputs in parallel and returns all outputs to a query at once, skipping sequential next-token prediction.
- ▪TypeSafe AI's Jev model is designed to return typed probabilistic decisions for use by other software or AI models instead of natural language.
Response speed advantages
- ▪TypeSafe AI states that adding more questions in the same call to Jev, its non-generative AI model, barely increases the model's response time
- ▪A demo on the TypeSafe AI website shows Jev, TypeSafe AI's non-generative AI model, returning a response in 0.114 seconds, compared to 8.566 seconds for OpenAI's GPT-5.6 Terra
- ▪TypeSafe AI claims Jev delivers response times ranging from 70 to 500 milliseconds, which is 40 to 200 times faster than traditional large language models.
Cost efficiency comparison
- ▪OpenAI's GPT-5.6 Terra costs $2.00 per million input tokens and $12.00 per million output tokens, making Jev, TypeSafe AI's non-generative AI model, significantly cheaper
- ▪TypeSafe AI charges $0.042 per million input tokens for Jev, TypeSafe AI's non-generative AI model, and charges nothing for outputs
- ▪TypeSafe AI claims Jev, its non-generative AI model, costs 238 times less than the top-tier Fable 5.1 model
Business workflow applications
- ▪Jev processes state information through question primitives called Choice, Score, and Noul, which return structured responses with probabilities.
- ▪Jev, TypeSafe AI's non-generative AI model, can play the video game Doom when provided with structured data describing the player's game state
- ▪Jev is designed to handle tasks such as sorting customer service requests, identifying buying intent, and determining when a human agent should take over.
- ▪TypeSafe AI claims Jev, its non-generative AI model, is suited for AI automation software, real-time applications, AI map-reduce jobs for data classification, input verification, and AI model harnesses
TypeSafe AI funding
- ▪TypeSafe AI was founded to pursue machine-native AI focused on interactions with other machines rather than humans.
- ▪TypeSafe AI announced its launch out of stealth with $40 million in funding.
Structured output limitations
- ▪TypeSafe AI's published performance tests compare Jev against other AI models' responses rather than independently verified correct solutions, and exclude GPT-6 Astra.
- ▪Other AI models, such as OpenAI's Structured Outputs, can also output preset categories and structured data, meaning Jev must compete on speed, cost, and quality.
- ▪While TypeSafe AI markets Jev, its non-generative AI model, as hallucination-free, this guarantee only covers the output structure, and the model can still make factually incorrect choices within its preset options
Story comments
Loading comments…