Jev: The AI That Decides Instead of Writing (And Why It Could Change Automation)

One of ChatGPT's co-inventors asked himself a troubling question: why haven't "superhuman" chat models yet massively automated work? His answer is a new type of model, presented by the company TypeSafe and called Jev, trained using a method called RLCD. His bet: the AI most useful to the economy isn't the one that speaks the best, but the one that decides the fastest. Analysis, with the important nuances.

The Problem: Discussing Isn't Deciding

ChatGPT, Claude, or Gemini are excellent at understanding, explaining, and creating text. But in software, we often need something much simpler: a decision. Is this email urgent? Is this transaction suspicious? Which service should this ticket be routed to? A chat model responds with a sentence that then needs to be interpreted; Jev responds directly with the decision.

What Jev Does in Practice

Whereas a classic LLM writes a free-text response, Jev directly selects from authorized values and returns a structured output that your code can use immediately.

// A classic LLM responds in free text:
// "After analysis, I think this email is
//   probably an urgent complaint..."

// Jev directly responds with a structured decision:
{
  "category": "reclamation_urgente",
  "probability": 0.94
}

The software can then act autonomously: forward the message, open a ticket, or request human validation if the probability is too low.

Why « 20 to 200× faster » and « 40 to 400× cheaper »

LLMs produce text word by word, and each generated word is billed as an output token. Jev does not write: it directly selects a value — a choice, a score, a probability — and can process multiple questions in parallel. This explains outputs announced as free and a very low entry cost. Note, however: these figures concern decision-making tasks suited to Jev, not a universal comparison with LLMs, and they currently come from TypeSafe.

Jargon decoded

• Co-inventing ChatGPT : having participated in the work that led to ChatGPT — not having invented it alone.
• Superhuman chat models : better than humans on certain specific tasks, not on everything.
• AGI : artificial general intelligence, capable of adapting to almost any problem like a human.
• In stealth : the company was working in secret, without a public product.
• RLCD : Reinforcement Learning for Calibrated Decisions — producing decisions accompanied by reliable probabilities.
• Frontier model : model presented as being at the technological forefront.
• Composable intelligence : combining several small decisions in a complex workflow.
• AFAICT : « as far as I can tell », as far as I can tell.

Beware of "zero hallucination" marketing

The formula is a bit misleading. Jev normally cannot invent a response outside the authorized choices, nor produce broken JSON or imaginary text. But it can perfectly choose the wrong valid answer: if it has to decide between "fraud", "normal" and "uncertain", it will always return one of these values — even if it means classifying a real fraud as "normal". Zero hallucination therefore does not mean zero errors.

Very concrete use cases

Jev shines when the possible answers are known in advance:

• Classify an email: urgent, normal or spam.
• Detect a potentially fraudulent transaction.
• Automatically route a ticket to the right department.
• Assess the severity of a cybersecurity alert.
• Classify a document: invoice, contract, resume or form.
• Give a risk score to a customer or an operation.
• Filter dangerous or inappropriate content.
• Decide which AI agent or which tool to call.
• Determine if an uncertain case should be sent to a human.

For real projects: a personal finance tool can categorize expenses and spot unusual transactions; a training platform can classify student questions or verify precise evaluation criteria.

Real example: compaction in Claude Code

An open source plugin, fast-jev-compaction, replaces Claude Code's compaction summary with Jev decisions: each previous tool call and its result are scored in a single fast request, obsolete traces are removed or shortened, and what is kept remains unchanged. Result: fewer tokens, less noise in the context, reduced cost, and potentially faster processing.

Two important clarifications to avoid misunderstandings. First, the plugin does not remove the tools, skills, or plugins themselves: it only removes their old traces deemed unnecessary in the history; recent messages and texts from the user and assistant are protected, and in case of failure it falls back to Claude Code's normal compaction. Next, although this plugin is written in TypeScript, Jev is not limited to TypeScript: it works via API, so its decisions can be used just as well from Python, Java, C# or other languages.

The Right Architecture: LLM + Jev

Jev is neither “a more powerful ChatGPT” nor an AGI. It is an extremely fast judgment engine. The most promising combination looks like this: an LLM to understand, explain and create, and Jev to quickly make repetitive decisions. The rule is simple: if you want to produce content, use ChatGPT or Grok; if you need to make thousands of quick small decisions, use Jev.

The idea is genuinely interesting for industrial automation. One major caveat remains: the announced performance figures come mainly from TypeSafe and will need to be confirmed by independent evaluations. Worth watching closely.