TypeSafe Exits Stealth With Jev, an AI Model Built for Programmatic Decisions

 

TypeSafe AI has emerged from stealth with Jev, a new AI model designed to make structured programmatic decisions directly inside production software without generating conversational text.

Founded by OpenAI veteran and ChatGPT co-inventor Diogo Almeida, TypeSafe spent two years developing Jev as an alternative to traditional large language models for applications where speed, predictable outputs, and strict data structures matter more than natural-language generation.

Instead of producing words token by token, Jev takes an unstructured input state and returns typed probabilistic decisions in a single parallel operation.

Almeida describes the system as a kind of frontier-intelligence function call.

“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”

The company positions Jev as a System One Model, optimized for rapid decision-making rather than extended conversational reasoning.

AI Without Text Generation

Traditional large language models typically generate outputs autoregressively, predicting one token after another.

That architecture works well for chatbots, writing assistants, coding systems, and reasoning tasks, but it can introduce unnecessary latency and complexity when software only needs a structured decision.

Jev removes text generation from the process entirely.

Rather than asking an AI model to generate a sentence and then converting that response into machine-readable data, developers define an output schema in advance.

The model then produces values that conform directly to that structure.

This approach is designed to reduce problems such as:

• Invalid output formats
• Parsing failures
• Unexpected text responses
• Hallucinated fields
• Additional validation pipelines
• Complex guardrail wrappers

For developers building automated workflows, Jev effectively turns AI inference into something closer to calling a probabilistic software function.

Reinforcement Learning for Calibrated Decisions

TypeSafe developed a training approach called Reinforcement Learning for Calibrated Decisions, or RLCD.

Traditional frontier AI systems commonly use techniques such as Reinforcement Learning with Human Feedback and Reinforcement Learning with Verifiable Rewards.

These approaches generally optimize models for conversational preferences, reasoning performance, or measurable task outcomes.

RLCD instead focuses on whether a model's reported confidence accurately reflects its likelihood of being correct.

For example, if Jev assigns a decision an 80 percent confidence score, TypeSafe aims for that prediction to actually be correct approximately 80 percent of the time across comparable cases.

This type of probability calibration could become particularly important for AI systems operating inside automated business processes, where software may need to determine whether to execute a decision automatically or escalate it for further review.

Parallel Sampling Instead of Token-by-Token Generation

One of Jev's biggest architectural differences is its hardware-aware parallel sampler.

Rather than generating outputs sequentially, Jev evaluates structured decision values simultaneously.

The system can handle selections involving as many as 255 discrete options.

TypeSafe says its two-stage process evaluates candidate options independently before making the final selection.

This architecture is intended to take greater advantage of parallel computing hardware while avoiding the latency created by long autoregressive generation sequences.

Response Times as Low as 70 Milliseconds

According to internal evaluations conducted by TypeSafe, Jev recorded end-to-end response latencies ranging from approximately 70 milliseconds to 500 milliseconds.

The company compared this with conversational frontier-model deployments that reportedly required between 3 seconds and 329 seconds for comparable workflows.

In multi-step workflow benchmarks involving fixed decision graphs, TypeSafe reported execution speeds reaching up to 193.6 times faster than selected conversational AI baselines.

The advantage comes largely from eliminating the requirement to generate and interpret natural language during each decision step.

For applications requiring hundreds or thousands of AI decisions every second, this architectural difference could significantly affect both latency and infrastructure costs.

AI Decisions at a Fraction of Traditional Token Costs

TypeSafe is also positioning Jev as a lower-cost alternative for high-volume inference.

The company lists input processing at approximately $0.042 per million tokens.

Because Jev does not generate traditional autoregressive output tokens, the company says structured outputs are not charged using conventional output-token pricing.

This pricing model could make the platform attractive for large-scale environments where AI decisions are executed continuously across massive datasets.

Potential applications include:

• Real-time feature extraction
• Automated classification
• Fraud and anomaly detection
• Recommendation systems
• Data-routing decisions
• Large-scale ETL workflows
• AI verification layers
• Automated software branching
• High-frequency operational decisions

Testing Jev Inside Doom

To demonstrate how the architecture performs under real-time conditions, TypeSafe tested Jev inside the classic video game Doom.

Engineers created an AI-controlled reactive bot capable of processing game states at approximately 10 queries per second.

The company estimates that maintaining this continuous inference workload costs roughly $7 per hour.

The experiment demonstrated Jev's ability to repeatedly interpret changing environmental states and make fast structured decisions without producing natural-language reasoning.

Wikiracing Tests Decision-Making Across the Web

TypeSafe also evaluated Jev through a Wikiracing experiment.

In Wikiracing, an AI system must navigate from one Wikipedia article to another by selecting links across multiple pages.

Instead of generating natural-language plans, Jev evaluates available links directly and chooses among them.

According to TypeSafe, the model completed navigation tasks using fewer steps than several external non-reasoning models while reducing failures caused by hallucinated or nonexistent links.

The experiment demonstrates the type of environment Jev is designed for: large decision spaces where software must continuously choose among structured alternatives.

From Chatbots to Decision Engines

Jev represents a broader shift emerging across the AI industry.

The first wave of generative AI focused heavily on creating content.

The next wave increasingly involves AI systems that make decisions and take actions inside software infrastructure.

Not every AI problem requires an intelligent chatbot.

A logistics platform may simply need to choose the best route.

A cybersecurity platform may need to classify an event.

A financial system may need to flag a transaction.

A recommendation engine may need to select one option from hundreds.

In these environments, generating paragraphs of natural language can become unnecessary computational overhead.

Models like Jev attempt to strip AI down to something much closer to its operational core:

State in. Decision out.

TypeSafe has now opened early developer access to Jev and begun onboarding engineering teams from its deployment waitlist.

If the company's architecture performs at scale as advertised, Jev could signal the emergence of a new AI category sitting somewhere between conventional machine learning models and massive conversational LLMs.

Instead of asking AI to explain everything, developers may increasingly ask it to simply make the right decision, quickly, cheaply, and in a format software can immediately execute.