Learn
Learn how AI decision models work, in plain English
Plain-language, sourced explainers on Laya, Jev and the ideas underneath them: typed questions, calibration, proper scoring rules and one-pass inference.
Swiss-hosted inference. Nothing you send is ever stored. Swiss data residency
23 pages · press / to search
Start here
- ExplainerFoundationsWhat is Laya? The open-source AI model for fast, calibrated decisionsLaya is an open-source AI model that answers typed questions about any text with calibrated probabilities in one fast pass. How it works, limits and use.
- ExplainerFoundationsWhat is Jev? TypeSafe's decision model, explained in plain EnglishJev is TypeSafe AI's hosted System One model: typed choice, score and yes/no decisions with probabilities. Its API, pricing, limits and independent benchmarks.
- ExplainerFoundationsSystem 1 vs System 2 AI: when to decide fast and when to reasonSystem 1 vs System 2 AI in plain English: when a fast, calibrated decision model beats an LLM, when it does not, and how to combine the two safely in code.
- ExplainerModelsNon-autoregressive models: how AI can decide in a single passWhat non-autoregressive models are and why they answer in one pass instead of token by token: how they work, how fast they are, and where they fall short.
- ExplainerCalibrationCalibrated probabilities: when can you trust an AI model's confidence?What calibrated probabilities are, how to measure them with ECE and reliability diagrams, how temperature scaling fixes them, and how to automate on them.
Deep dives
- ExplainerFoundationsTyped decisions: asking a model for a value, not a paragraphWhat is a typed decision? You list the allowed answers up front and the AI returns one of them with a probability. No free text to parse, no invented labels.
- ExplainerFoundationsChoice, score and noul: three primitives for typed decisionsChoice picks a label, score places text on a scale, noul gives the probability a statement is true. How each works in Laya, when to use it and where it is weak.
- ExplainerCalibrationProper scoring rules: rewarding a model for honest probabilitiesWhat a proper scoring rule is, why accuracy is not one, and how log, Brier, spherical and RPS scores reward honest probabilities, as in Laya's training.
- ExplainerCalibrationRLCD: reinforcement learning for calibrated decisionsRLCD trains Laya with reinforcement learning whose reward is a strictly proper scoring rule, so honest probabilities score best. What it fixes and what not.
- ExplainerModelsEncoder vs decoder models: which architecture should make your decisions?Encoder vs decoder transformers: encoders read text in one pass, decoders write it token by token. What that means for classification cost, speed and output.
- ExplainerModelsModernBERT: a modern encoder for fast classificationWhat ModernBERT is, how it improves on BERT with 8,192-token context and faster inference, what it is good at, and how Laya builds on ModernBERT-large.
- ExplainerModelsmmBERT: the multilingual encoder behind Laya's global checkpointWhat mmBERT is: a multilingual ModernBERT-style encoder from Johns Hopkins, trained on 1,800+ languages. How it compares with XLM-R and how Laya uses it.
- Deep diveCalibrationOption-marker scoring: how Laya reads a question and its answers togetherHow option-marker scoring works: Laya writes your answer options into the input, marks each one and scores them all in one pass. New labels need no retraining.
- GuideIn productionAct or escalate: routing decisions by calibrated confidenceWhen should AI act and when should it hand off to a human? How to set a confidence threshold from your costs, and how to gate Laya's answers on confidence.
- GuideModelsMultilingual classification: one decision, many languagesHow to classify text in many languages: translation, per-language models, multilingual encoders or LLMs. Why English-only models fail confidently.
- GuideIn productionZero-shot vs fine-tuned: when is a general model good enough?Zero-shot vs few-shot vs fine-tuned text classification: data needed, accuracy and how fast you can change labels, with Laya's published numbers as an example.
- AnalysisIn productionThe real cost of using an LLM as a classifierThe real cost of using an LLM as a text classifier: prompt and output tokens, latency, retries and parsing, and where a one-pass model like Laya is cheaper.
- ExplainerIn productionHallucination-free decisions: what a non-generative model rules outCan an AI classifier hallucinate? What a non-generative model like Laya rules out (invented labels, fake fields, unparseable text) and what it can get wrong.
- GuideIn productionLatency budgets for AI agentsHow to set a latency budget for an AI agent: where the time goes in a turn, serial vs parallel steps, tail latency, and where a fast decision model fits.
- GuideCalibrationTemperature scaling: the one-parameter fix for over-confidenceWhat temperature scaling is, how one fitted number fixes over-confident probabilities without changing predictions, and how to fit it on Laya outputs.
- ExplainerCalibrationExpected calibration error: measuring whether 80% means 80%What expected calibration error (ECE) measures, how to compute it with bins, a worked example, common pitfalls, and what Laya's ECE figures mean.
- Deep diveModelsLanguage routing: picking the checkpoint that can read the inputWhat language routing is: detect the script and language of each input, then send it to a model that can read it. How Laya's router decides in under 1 ms.
- GuideIn productionDecision models for AI agents: fast judgements around a slow plannerWhy AI agents need a separate decision model for guardrails, intent, tool choice and escalation, how it fits beside the LLM, and where Laya's limits are.
Get started
Try it on your own text
Send any text with your own typed questions and get a calibrated probability for every answer. 6k free tokens, no card needed; then $0.0357 per 1M input tokens — 15% below Jev.
Swiss-hosted inference. Nothing you send is ever stored.