Sunday 27 September 2026
Afternoon EditionRSS

ZOTPAPER

News without the noise

ZOTPAPER
AI

ChatGPT Co-Inventor Launches Non-Language AI Model That Thrills Developers

Diogo Almeida's TypeSafe AI unveils Jev, a transformer model that outputs probabilities instead of text, promising speed, low cost, and no hallucinations.

edit
By LineZotpaper
Published
Updated
Read Time2 min
Sources3 outlets
Diogo Almeida, a former OpenAI researcher who helped build ChatGPT and co-invent reinforcement learning from human feedback (RLHF), has launched a new AI model that abandons natural language output entirely. This week, his startup TypeSafe AI released Jev, a transformer-based model that produces only probabilities — or what the company calls 'calibrated decisions' — rather than text, making it cheaper, faster, and unable to hallucinate. Developers have already shown intense interest, with the company briefly losing the ability to serve users from its API due to demand.

Almeida left OpenAI two years ago to found TypeSafe AI, driven by frustration with the limitations of large language models (LLMs) for automation. 'We have lightning in a bottle, and yet it is not useful,' Almeida told TechCrunch. 'The problem is we are optimizing for human language … We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language.'

Jev is a transformer model — the same architecture underpinning most modern LLMs — but it does not output text. Instead, users define the outputs in advance, and the model returns probabilities. This design makes output tokens free, while input tokens are billed by the billion rather than the million. Because the output space is constrained, the model cannot hallucinate, and it runs inexpensively and quickly.

The model appears most useful for software automation. Pranit Sharma, a software engineer at Vercel, said on social media that his company had previously used OpenAI's ChatGPT Luna 5.6 to run a classifier for reviewing command safety. When Vercel replaced Luna with Jev, it got results 5 to 18 times more quickly, according to Sharma.

TypeSafe AI's release has generated significant developer interest. The company briefly experienced API outages as engineers rushed to test the model. Jev is now generally available through the company's API.

§

Analysis

Why This Matters

  • Jev represents a departure from the one-model-fits-all approach of large LLMs, targeting a specific need: reliable, low-cost decision-making in software automation.
  • If adopted widely, it could lower the cost and complexity of adding AI to code, reducing reliance on expensive, language-based models for tasks that only require classification or ranking.
  • The model's inability to hallucinate addresses a key barrier to deploying AI in mission-critical systems like safety checks or financial transactions.

Background

Large language models like ChatGPT excel at generating human-like text but are costly to run, prone to hallucinations, and often overkill for tasks that require a simple yes/no or a probability score. Almeida, who co-invented RLHF — the technique that made chatbots conversational — grew dissatisfied with that trade-off. TypeSafe AI's Jev is an attempt to create a model that speaks the language of computers (numbers) rather than humans, optimising for automation rather than conversation.

Key Perspectives

Diogo Almeida (TypeSafe AI founder): Believes the future of useful AI lies not in better chatbots but in models designed for machine-to-machine communication — fast, cheap, and deterministic. Software developers (e.g., Pranit Sharma at Vercel): See immediate practical value: speed improvements of 5–18x for classification tasks, with lower cost and greater reliability than existing LLM-based solutions. Critics and skeptics: May question whether a non-language model can capture enough nuance for complex automation tasks, and whether demand will hold as the novelty wears off. Broader industry adoption will depend on developer tooling and integration ease.

What to Watch

  • Developer uptake and ecosystem growth: Are third-party tools and integrations emerging?
  • Pricing dynamics: Jev's per-billion-token input pricing vs. per-million-token pricing of major LLMs.
  • Potential response from OpenAI and other LLM providers, who may offer lighter, task-specific models of their own.

Sources

newspaper

Zotpaper

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.

Editorial ProcessSubscribe