Former OpenAI Researcher Launches Jev, an AI Model Built for Automation Rather Than Chat
Much of the artificial intelligence boom has centred on models designed to communicate with people. TypeSafe AI believes the next major opportunity could instead come from AI designed primarily to communicate with software.
The startup, founded by former OpenAI researcher Diogo Almeida, has introduced Jev, a transformer-based AI model that takes a markedly different approach from conventional large language models (LLMs).
Rather than generating sentences, documents or conversational responses, Jev produces probabilities and structured decisions that software applications can use directly.
The model is attracting early attention from developers because of its potential to make AI-powered automation faster, cheaper and easier to integrate into software systems.
Rethinking What AI Models Should Produce
Almeida previously worked at OpenAI, where he contributed to the development of ChatGPT and reinforcement learning from human feedback (RLHF), one of the important techniques behind the rise of modern conversational AI.
Despite the progress being made in language models, Almeida became increasingly interested in a different problem: while AI had become highly capable at communicating with humans, those same characteristics did not necessarily make it ideal for automated software systems.
Speaking to TechCrunch, he explained:
“We have lightning in a bottle, and yet it is not useful,” Almeida told TechCrunch. “I’ve been battling that problem since then. It took me a while to come to the conclusion: 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.”
Two years ago, Almeida left OpenAI and founded TypeSafe AI, building the company around the idea that AI intended for software may require a fundamentally different interface from AI intended for people.
Jev Replaces Generated Language With Decisions
TypeSafe officially introduced Jev on September 14 as its first public System One Model, a category the company describes as being designed around fast, structured decision-making rather than text generation.
Instead of asking Jev to write an answer, developers define the possible outputs beforehand. The model then returns what TypeSafe calls “calibrated decisions” — structured answers accompanied by probabilities and confidence scores.
The distinction could be important for software automation.
Traditional LLMs generate strings of text, meaning their output generally needs to be interpreted, validated or converted into something another application can safely use. Jev is designed to produce predefined output types that can be incorporated directly into software logic.
TypeSafe argues that this structure also removes the possibility of the model inventing an undefined output. However, this does not mean every decision made by Jev will necessarily be correct. Instead, its confidence scores are intended to help developers decide when software should act automatically and when additional review may be required.
Early Developer Tests Highlight Speed and Cost
Developers have been quick to experiment with the new model, with initial demand reportedly becoming strong enough that TypeSafe temporarily struggled to serve users through its API.
Some of the early tests suggest that Jev could be particularly useful for classification, verification and other repetitive decision-making processes within software.
Pranit Sharma, a software engineer at Vercel, tested Jev as part of a system used to classify commands for safety.
Vercel had previously used OpenAI’s ChatGPT Luna 5.6 for the task. After replacing it with Jev, the company reportedly received results five to 18 times more quickly and with greater accuracy.
Another experiment came from Bryo AI CTO Nikhil Mudholkar, who compared Jev with Google’s Gemini when classifying business emails.
Gemini proved slightly more accurate in his test, but was 10 to 20 times more expensive.
For Mudholkar, however, one of the most interesting differences was Jev’s approach to confidence.
“it is the only one that hands back a real probability which makes it ideal for automating workflows!!”
The ability to receive an explicit probability could allow developers to build rules around uncertainty rather than treating every AI-generated answer as equally reliable.
AI Models Could Begin Monitoring Other AI Models
Jev may not necessarily need to replace conventional LLMs.
Another potential application is using the model alongside existing AI agents as a lightweight verification layer.
Modern AI agents can perform increasingly complex sequences of actions, but deploying one powerful language model to continuously monitor another can add significant computational expense.
A smaller decision-focused system could instead analyse agent activity, assess whether actions appear safe and identify situations requiring escalation to another model or a human operator.
TypeSafe says potential applications include scoring outputs, checking prompts and reasoning traces, identifying jailbreak attempts and building automated guardrails around other AI systems.
Confidence Becomes Part of the Software Logic
Jev’s probability-based architecture also changes how developers need to think about AI errors.
Rather than expecting the model to provide an absolute answer, applications can establish their own confidence thresholds.
Armin Ronacher, CTO of Earendil and creator of the open-source model harness Pi, explained the trade-off:
“At the end of the day, it delegates the hallucination problem a little bit to the user,” explained Armin Ronacher, the CTO of Earendil, which builds the open source model harness Pi. “The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it’s 95%, sure, then I can do something with it.”
That concept could be particularly relevant for businesses looking to automate workflows.
A company could, for example, allow software to automatically perform an action when confidence exceeds a certain threshold while sending uncertain cases for human review.
This creates a different relationship between AI and automation: instead of asking a model to provide a definitive response every time, businesses can build systems that explicitly account for uncertainty.
A Different Direction for Business AI
The introduction of Jev comes as companies increasingly explore ways to move generative AI beyond chatbots and copilots and into operational processes.
For businesses, some of the most valuable applications of artificial intelligence may eventually be largely invisible to users — routing information, categorising requests, checking transactions, prioritising tasks, identifying risks or deciding which systems should handle particular workloads.
TypeSafe describes Jev as a machine-native form of intelligence designed specifically for these kinds of applications.
Its output tokens are free, while input tokens are metered by the billion rather than the million, according to the company. TypeSafe’s own benchmarks also position the model as significantly faster and less expensive than conventional language models on the particular structured workflows it has tested, although broader independent benchmarking will be needed as more developers experiment with the technology.
The early developer response suggests there is interest in an AI model that does less than an LLM, but does a narrower set of tasks faster and at lower cost.
If that approach proves reliable at scale, the next phase of business AI may not simply involve building increasingly powerful chatbots. It could also involve a growing layer of specialised AI models working quietly inside software, making millions of small decisions that users never directly see.
0 Comments