Goodbye JSON Parsing: TypeSafe Launches "Non-Conversational" AI Built for Code

Overview

As generative AI sees widespread adoption in industry, the friction between natural language processing and programmatic execution is becoming increasingly prominent. Over the past three years, the tech industry broadly assumed the future of computing would be conversational, forcing developers to adapt large language models (LLMs) into "JSON-only output" classifiers to fit existing APIs and databases. However, this autoregressive text generation approach brings high latency, formatting errors, and unreliable "hallucination" risks.

Recently, Diogo Almeida, a former OpenAI researcher and lead author of the InstructGPT paper, announced that his company TypeSafe AI has raised $40 million and launched Jev, an AI model designed specifically for code environments that is "non-conversational." This model aims to address the latency and cost pain points of semantic decisions in AI workflows, marking a technical paradigm shift from "System 2" reasoning to "System 1" rapid judgment.

Farewell to JSON Parsing: When Strings Become the Bottleneck

For any engineer deploying LLMs into production, waiting for the model to generate tokens one by one is fraught with uncertainty. Starting from {, passing through commas, quotes, and key names, to the final closing bracket, this process not only takes 2 to 5 seconds but also easily leads to parsing failures due to model "hallucinations" or formatting errors. Developers are forced to write regular expressions to strip out redundant replies like "Sure, here is your JSON," and build retry loops to handle unescaped quotes or KeyError exceptions.

TypeSafe AI believes that strings are a disastrous abstraction for software systems. The core design philosophy of the Jev model is precisely to abandon the ability to generate conversational text—it cannot write articles or play roles. Instead, it focuses on directly returning type-safe structured decisions, scores, and calibrated probability distributions within 70 to 300 milliseconds. This architecture eliminates JSON decoding loops, achieves a zero-pattern hallucination rate, and allows AI to be embedded reliably into software stacks just like traditional code logic.

"System 1" Models and the Jevons Paradox

TypeSafe's technology naming is deeply inspired by psychologist Daniel Kahneman's dual-process theory proposed in Thinking, Fast and Slow. Current frontier LLMs (such as OpenAI's o1/o3 series) focus on System 2 mode, which involves slow, deliberative reasoning through lengthy chains of thought. While this is crucial for proving theorems or designing complex architectures, 95% of software execution scenarios—such as routing, classification, scoring, and branching judgments—do not require this overhead.

Jev is designed as a semantic-reflective System 1 model. Additionally, the company founder cites economist William Stanley Jevons' paradox, pointing out that as computational efficiency improves, application boundaries will expand significantly. Currently, frontier models cost between $2 and $15 per million tokens, whereas Jev's price is only $0.042 per million input tokens, with output tokens completely free. When the cost of semantic decisions drops to $0.00004 and latency is controlled within 100 milliseconds, AI will no longer be limited to high-profit human-computer interaction scenarios but will penetrate every line of if statements, database triggers, and real-time network packet filters.

Three Core Primitives: Choice, Score, and Noul

To achieve programmatic interaction, Jev abandons open-ended prompt engineering in favor of three modular software primitives:

• Choice: Selects 1 result from a declared set of options (supporting up to 255 discrete options), returning the selected value, probability distribution of all candidates, and confidence score.
• Score: Evaluates state along ordered continuous criteria (e.g., sentiment 0-3, urgency 1-5), returning continuous numerical values, hierarchical distributions, and confidence.
• Noul: Evaluates the truth of binary propositions, returning a calibrated probability value between 0.0 and 1.0.

"Natural language is an extraordinary interface for humans, but for software, strings are a disastrous abstraction."

This architecture means developers no longer need to handle unstructured text garbage but can directly obtain deterministic data ready for programmatic consumption. As TypeSafe AI progresses, the role of AI in backend logic is shifting from "conversational partner" to "instant decision engine," laying the foundation for building cheaper and faster intelligent software systems.