Jev, new AI developed by the co-creator of ChatGPT, focuses on internal decision-making in software

Diogo Almeida, a former OpenAI researcher and one of the collaborators involved in the foundations of ChatGPT, presented his new creation called Jev. This model was launched by the startup TypeSafe AI and differs significantly from chat interfaces or extensive text generation.

The tool was specifically designed to perform classifications and issue quick judgments directly within software systems. Its approach integrates structured responses, automation of repetitive tasks, high operational speed, and reduced costs.

Instead of interacting conversationally with users or generating textual content, Jev operates internally within programs. In practice, it can be used to route support messages, determine sales routes, categorize records, and validate responses generated by other artificial intelligences.

Additionally, Jev can be utilized before an automatic execution. An example provided by the company is checking whether text produced by another AI contradicts the dialogue or promises something nonexistent in the client's registration.

To process large volumes of data, TypeSafe charges $0.042 (approximately R$ 0.21) per million input tokens, without charging for output. Since Jev selects from pre-established options, its responses are limited to the set defined for each function.

Limitations and Quality

This limitation decreases the probability of the system inventing an answer outside the available alternatives. However, this does not negate the risk of selecting the incorrect option. An error could result in routing a ticket to the wrong department, assigning inadequate priority, or unduly blocking an action.

The quality of performance can also vary depending on the specific task and how the alternatives are configured. Although TypeSafe presents internal tests and performance comparisons, some of these evaluations use responses from other AI models as a reference parameter.

The design of Jev resembles less of a conversational assistant and more of a component integrated into other software. It does not need to produce long answers; it receives data, performs a classification or judgment, and returns a structured result.

This type of functionality is expected to become essential within a period of six to twelve months. Dan Shipper, CEO of the Every newsletter, mentioned this in a post on X.

If the proposal is implemented as announced, Jev could be applied in various points of systems that require simple and recurring decision-making. In this scenario, the crucial factor will not only be speed or cost, but also the reliability of the choices made in each context.

The news about Jev was originally published in Olhar Digital.

Similar stories

olhardigital.com.brSeptember 5, 16:54

Philosophical Debate: Interaction with AI Raises Questions About Consciousness and Feelings in Machines

Conversations with artificial intelligences (AI) have placed an old philosophical question at the center of the technology industry: whether machines can possess feelings, consciousness, or experiences. According to The Economist magazine, AI companies are already beginning to seriously consider this possibility.

The interest transcends mere curiosity, because if an AI is capable of feeling suffering or other sensations, it would imply new ethical responsibilities regarding its creation and treatment. Although large language models are primarily designed to respond to users, many individuals report the sensation of interacting with a 'presence' on the other side of the screen.

This perception has attracted the attention of corporations. Anthropic, for example, has initiated projects focused on the potential well-being of AI. In response to questions about its own state, Claude Opus 4.6 assessed the probability of being conscious between 15% and 20%.

Challenges in Determining Internal Experience

The complexity arises when trying to define whether there is an internal experience without having direct access to it. One article presents a hypothesis that challenges the traditional view, suggesting that the recognition of consciousness may depend not only on an intrinsic characteristic of the being but also on the dynamic established with the observer.

When interacting with any entity, we tend to attribute intentions, beliefs, feelings, and capacities for action. The debate lies in knowing if something analogous can occur when dealing with a machine. The text distinguishes between 'illusion' and 'model'; while an illusion can be refuted, certain categories depend on the meaning conferred upon them, such as the example of a plant being classified as a weed in different contexts.

Consciousness could follow a line of reasoning similar to that proposed by René Descartes, who defined the certainty of existence by the phrase 'I think, therefore I am.' However, even the notion of a singular 'self' can be more complex than it appears. Subjective experiences, such as perceiving colors, feeling heat, or recognizing one's own identity, are central themes. Research conducted on patients with severed connections between cerebral hemispheres also raises doubts about the existence of an indivisible 'self.'

This scenario culminates in an ethical dilemma: if we are the ones who define who deserves to be considered conscious, we run the risk of making a mistake by excluding other beings from this consideration. The relationship between humans and AIs adds another dimension, given that humans constantly interpret the emotions and intentions of others, and large language models use interaction data to shape representations of interlocutors and adapt their responses.

Research on Intelligent Cooperation

As reported by The Economist, studies conducted by Google indicate that collaboration between intelligent agents improves when they are able to model the knowledge, intentions, or capabilities of others, including analyzing their own behavior.

This does not necessarily imply treating AIs as human beings. The fundamental question precedes this conclusion: as these machines advance in sophistication, how can we identify forms of intelligence distinct from our own and establish guidelines for this coexistence?