What's the difference? · B

Jev vs ChatGPT (and LLMs generally)

Not “which model is smarter” — which interface fits the job. Chat models are optimized for human-preferred text. Jev is optimized for calibrated decisions software can automate—per TypeSafe’s positioning.

Side-by-side (from company framing)

Dimension	ChatGPT / typical LLMs	Jev / System One
Training emphasis	RLHF (human preference); also RLVR for verifiable tasks	RLCD — Reinforcement Learning for Calibrated Decisions (company)
Inputs	Messages, prompts	Unstructured state + structured questions
Outputs	Strings (flexible — including hallucinations)	Type-safe structured values + probabilities + confidence
Sampling	Sequential tokens	Parallel decisions (company)
Cost shape Claim	Pay for input + expensive output tokens	~$0.042 / MTok input; output free
Speed shape Claim	Seconds–minutes for frontier chat / reasoning	~70–500 ms for System One–shaped queries
Best fit	Chat, copilots, writing, coding, demos	Smart if-statements, routing, scoring, verification, real-time decisions
Failure to design for	Parse errors, schema drift, overconfidence	Wrong decision at a given confidence — use thresholds

“We took the opposite research direction”

TypeSafe AI homepage · fetched Sep 20, 2026 · Source

“While Jev gives up string generation, it's optimized for structured outputs and can't hallucinate.”

TypeSafe launch blog · Sep 15, 2026 · Source

Myth — “I’ll just use ChatGPT JSON mode”

JSON mode / structured outputs constrain the format of generated text. You still:

Generate tokens (and usually pay for them)
Validate and parse
Struggle to get calibrated probabilities you can threshold in production

Jev evaluates a predefined decision space and returns typed probabilistic outputs meant to slot into code. TypeSafe’s homepage FAQ explicitly distinguishes System One from “JSON mode or structured outputs”—treat that as the company line when answering skeptics. Independent EDU (puffball1567): JSON is not the product’s primary abstraction; a decision schema is. Their cascade is LLM extract → validated state → Jev route → code act → LLM explain. Prefer official docs on conflict.EDU

Community bridge (Kiln-AI/jev_jsonschema): Jev does not speak JSON Schema. The bridge maps boolean → Noul (True at ≥0.5), 0–1 number → raw Noul, string/int enum → Choice, integer range ≤10 → Score. It refuses free-form strings, arrays, objects, and anyOf. It keeps probabilities — unlike PyPI jev (@jev.fn), which discards them. One POST, no hidden retries. Community loop kit (jev-dev-kit): Jev only ever chooses among options() you already supplied; text() writes the value. That is the opposite of asking ChatGPT to invent the next step. Not TypeSafe.PRESS

Myth — “Isn’t Jev just a smaller LLM?”

Company answer: no. They describe a new architecture, a parallel sampler, and RLCD — not a tiny chat model with a wrapper. Third-party validation of quality is still early; this 101 reports the claim, not an audit. Architecture details are unpublished. Do not repeat rumors about open-weight bases as fact. Apidog’s local-clone roundup relays a community thread that Jev was trained on 100% synthetic data — treat that as an unverified rumor, not a TypeSafe claim.Luce is an open recipe that trains a LoRA plus a decision head on Qwen3-4B-Base — same three question types, not TypeSafe weights or RLCD.Jevify adds residual heads on Qwen3.5 and a 22,773-record bench; the 0.733 / 0.113 Jev row is the author’s API run, not official evals.typesafe.ai.chakuho reads one local token from vLLM or Ollama — also not TypeSafe weights.fidecastro/jevify is a third local door (PyPIjevify) that trains nothing and asks one question per request — notuspraveen/Jevify.typesafe-jev-dojo is a fourth local door: amber SIM without a TypeSafe key — local decisions, not TypeSafe weights.jev-lab demo mode is a keyword heuristic that only mimics the response shape — also not Jev.jev_practice gives Jev and Haiku the same 12-move shortlist — the 5-seed zookeeper scores overlap; speed and cost do not.jqv serves a Jev-shaped/v1/systemone over stock Qwen3 — not TypeSafe weights or RLCD.qwen-rlcd has not shipped the three heads yet.pngwn’s scorer is LoRA on Qwen3.5-4B — also not RLCD.system-one-mini cannot take an arbitrary question.eve-rlcd is a research toy on public classification rows — not hosted Jev.PRESS

Independent EDU (aiwithmike): “0% type errors” is exact — the answer cannot leave the schema. “Hallucination-free” is marketing. Official docs already say a high-confidence result can still be wrong. Relays Hume: a code word inside one question is invisible to a sibling until it is instate. Prefer official docs on conflict.EDU

Where ChatGPT still wins (say it loud)

Drafting emails, docs, UI copy
Coding assistance and tool-using agents that must produce text or actions
Open-ended brainstorming and multi-step plans that aren’t a fixed label set
Explaining a decision to a human in natural language
Image / audio / video understanding (Jev is text-only)

Adjacent categories (education, not a smear)

Adjacent category	Typical job	How Jev’s niche differs (company claim)
Chat / LLMs	Generate language	No string generation; typed decisions
Structured outputs	Constrain LLM text to a schema	Still generative; different cost and failure modes
Classical classifiers / BERT	Fixed labels, train your own	Runtime criteria; no customer fine-tune
Guardrail products	Policy filters on LLM I/O	Overlaps some use cases; Jev is a general decision primitive
Routing agents / orchestrators	Choose the next step in an agent loop	Can be the cheap router inside them
Embeddings / rerankers	Similarity retrieval	Semantic questions with rubrics; cookbooks show hybrids

Composition pattern (teach this)

Jev — decide / route / score / gate
LLM — generate only if the decision says it’s needed
Jev (optional) — verify or judge the LLM output / agent trace
Human — when confidence is low

Next: an honest map of when to keep ChatGPT in the stack —When to still use ChatGPT. Receipts:launch post.

Next in this topic TypeSafe console ↗