Ex-ChatGPT Developer Unveils New Class of Models That Don't Generate Text
Ex-ChatGPT Developer Unveils New Class of Models That Don't Generate Text
Diogo Almeida, co-author of RLHF, introduced Jev by TypeSafe AI — a new class of AI models that bypass text generation to output structured data, running up to 200x faster than standard LLMs.
Diogo Almeida, a former ChatGPT developer, co-author of RLHF, and one of the authors of the InstructGPT paper (the technique that laid the foundation for ChatGPT in 2022), has introduced a new class of models called System One Models.
In 2024, Almeida founded TypeSafe AI. After two years of working in stealth mode, the company has officially unveiled its first model — Jev.
The Core Idea
The concept is that for a vast number of tasks, text is simply unnecessary. If a model needs to classify a request, assess risk, or choose the next step for an agent, a standard LLM still generates the answer token by token, which then has to be parsed.
Jev, however, immediately returns a structured response along with the probability of each option. For example, instead of writing text about where to route a user's request, it directly returns the required category.
Performance and Pricing
Thanks to skipping text generation, Jev operates 20–200 times faster than standard LLMs on such tasks (70–500 ms).
The pricing is also unusual:
- $0.042 per 1 million input tokens.
- Output tokens are free (since the model doesn't generate text, it outputs structured data).
Jev is currently available in early access.
Why it matters
AnalysisThis matters because eliminating text generation for classification tasks drastically reduces latency and infrastructure costs, making AI agents significantly more efficient and scalable.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Jev and System One AI: A Complete Practical Guide to the Decision Model from TypeSafe
A deep engineering breakdown of the flagship Jev model: System One architecture versus LLMs, Choice/Score/Noul primitives, request parallelism, confidence mathematics, video use cases, code, and an overview of local open-weight alternatives (Laya, GLiNER2.5, CLM-8B).
Ghostty developer adds playful troll rule for AI agents in AGENTS.md
Ghostty's AGENTS.md tells AI agents to add a mocking file when opening pull requests, while Turborepo uses a similar 'I didn't check my work' marker.
Six DeepSeek Harness Plugins to Cut Token Usage and Organize Context
A review of six plugins for DeepSeek Harness, covering free model access, memory organization, agent routing, and context monitoring.
AI in Business Processes: From Diagnostics to a Production-Ready Hermes Agent
A comprehensive practical B2B playbook covering process diagnostics, single-agent workflows, tooling, memory, security, evaluation, observability, and runbooks.