In a significant development for the artificial intelligence landscape, Diogo Almeida, a former OpenAI researcher instrumental in the creation of ChatGPT and the pioneering technique of Reinforcement Learning from Human Feedback (RLHF), has introduced a groundbreaking new AI model named Jev. Launched by his startup, TypeSafe AI, Jev represents a radical departure from the Large Language Models (LLMs) that currently dominate the AI discourse. Almeida’s journey from a foundational architect of conversational AI to a vocal critic of its limitations for automation, culminating in the development of Jev, underscores a growing recognition within the industry that the path to truly pervasive and reliable AI might lie beyond human language.
The Genesis of Disillusionment: A Visionary’s Pivot from LLMs
Diogo Almeida’s name is inextricably linked with some of the most pivotal advancements in modern AI. As a researcher at OpenAI, he played a crucial role in the initial development of the chatbot that would eventually become ChatGPT, a phenomenon that redefined public perception of AI capabilities. More profoundly, he is credited with inventing Reinforcement Learning from Human Feedback (RLHF), a model-training technique that has since become the gold standard for aligning powerful AI models with human intentions and preferences. RLHF involves using human preferences to fine-tune a language model, significantly improving its ability to generate coherent, relevant, and helpful responses, thereby mitigating issues like toxic output or irrelevant information. This innovation was arguably one of the most critical factors in making conversational AI models like ChatGPT not just powerful, but also genuinely usable and widely adopted.
Despite his integral role in unleashing the current age of AI, Almeida found himself increasingly disillusioned with the very technology he helped champion. He perceived a fundamental disconnect between the extraordinary linguistic capabilities of LLMs and their practical utility for robust automation. "We have lightning in a bottle, and yet it is not useful," Almeida candidly shared with TechCrunch, articulating the core of his frustration. He elaborated, "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."
This realization crystallized a critical challenge in the AI development paradigm. While LLMs excel at generating human-like text, their inherent probabilistic nature and tendency to "hallucinate" (generate factually incorrect but plausible-sounding information) make them less than ideal for tasks requiring absolute precision, reliability, and deterministic outcomes – qualities paramount in automated systems. The cost and computational overhead associated with running LLMs for repetitive, high-volume tasks further compounded this issue, creating a bottleneck for widespread intelligent automation.
TypeSafe AI and the Birth of Jev: A New Paradigm for Machine Intelligence
Driven by this profound dissatisfaction and a vision for a more practical form of AI, Almeida departed from OpenAI two years ago to establish TypeSafe AI. His mission: to build AI systems that speak the "language of computers" rather than mimicking human conversation. This week, TypeSafe AI publicly unveiled Jev, a transformer-based model designed specifically to address the limitations Almeida identified in LLMs. Crucially, Jev is not a Large Language Model; it does not generate text. Instead, it produces probabilities, which TypeSafe AI terms "calibrated decisions."
This fundamental design choice carries several profound implications and advantages for automation. By eschewing human language as its output, Jev becomes:
- Incredibly Cheap and Fast: The computational resources required to generate probabilities are significantly lower than those needed for complex language generation. This translates directly into dramatically reduced operational costs and increased processing speed. TypeSafe AI states that Jev’s output tokens are free, and input tokens are metered by the billion, a stark contrast to LLMs typically metered by the million, indicating a potential cost reduction by orders of magnitude.
- Incapable of Hallucination: One of the most persistent challenges with LLMs is their propensity to "hallucinate." Because Jev’s outputs are predefined probabilities for specific outcomes, and not free-form text, it cannot invent information or generate factually incorrect responses. Users define the possible outputs in advance, ensuring the model’s responses remain within a controlled, verifiable framework.
- Deterministic and Reliable: The output of calibrated decisions provides a clear, quantifiable measure of confidence, making Jev ideal for critical automation workflows where certainty is paramount. This allows developers to build systems that can reliably act on high-confidence decisions or flag low-confidence decisions for human review.
Real-World Validation: Jev’s Performance in Practice
The launch of Jev has been met with significant enthusiasm from the developer community, so much so that TypeSafe AI briefly experienced an overload on its API due to unprecedented demand. Early adopters are quickly discovering Jev’s superior performance in various software automation tasks, often outperforming established LLMs in speed, cost-efficiency, and accuracy for specific applications.
One compelling example comes from Pranit Sharma, a software engineer at Vercel, a company specializing in agentic infrastructure. Vercel had been utilizing OpenAI’s ChatGPT Luna 5.6 to power a classifier for reviewing commands, primarily to ensure safety and compliance. Upon replacing OpenAI’s Luna with Jev, Sharma reported astonishing improvements: Vercel achieved results five to 18 times more quickly, coupled with greater overall accuracy. This quantitative leap demonstrates Jev’s capacity to significantly enhance the efficiency and reliability of critical safety protocols within automated systems.
Another independent test was conducted by Nikhil Mudholkar, CTO of Bryo AI. Mudholkar benchmarked Jev against Google’s Gemini for classifying business emails. While Gemini showed a slight edge in raw accuracy in his specific test, Jev’s performance was remarkable in other areas. Most notably, Jev proved to be 10 to 20 times more cost-effective. Beyond cost, Mudholkar highlighted Jev’s unique ability to provide "confidence scores." He stated, "it is the only one that hands back a real probability which makes it ideal for automating workflows!!" This feature allows developers to programmatically decide whether to act on a Jev output based on its probability score, introducing a level of control and reliability absent in most LLM integrations.
Beyond Replacement: Augmentation and New Use Cases
Jev’s utility extends beyond merely replacing LLMs in certain applications. It also presents exciting opportunities for augmenting existing AI systems and enabling entirely new automation paradigms. Diogo Almeida envisions Jev playing a crucial role in monitoring and managing the behavior of LLM-powered agents. The cost of using one LLM to monitor another can quickly become prohibitive, but Jev’s low operational cost makes this a viable strategy. Almeida suggests deploying Jev to track LLM agent traces, identifying and preventing undesirable behaviors such as "jailbreaks" – attempts to bypass safety filters or ethical guidelines.
Armin Ronacher, CTO of Earendil, a company building the open-source model harness Pi, elaborated on the practical implications of Jev’s calibrated decisions. He explained that Jev effectively "delegates the hallucination problem a little bit to the user." Instead of an LLM generating potentially false information, Jev provides a probability score. "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," Ronacher noted. This empowers developers to build more resilient and trustworthy AI applications by incorporating explicit risk assessment into their automation logic.
Ronacher also highlighted Jev’s potential for "model routing." In complex AI systems, determining which specific model is best suited for a given workload can be a significant challenge. Using an LLM for this decision-making process would be overly expensive and slow. Jev’s combination of low cost and high speed makes real-time, intelligent model sorting a practical reality, optimizing resource allocation and improving overall system efficiency.
The Economic Imperative: Jevons Paradox and Pervasive Intelligence
The name "Jev" is a deliberate homage to William Stanley Jevons, the 19th-century economist whose eponymous paradox describes how an increase in efficiency or a fall in the cost of a resource can lead not to a reduction in its consumption, but rather to an increase. In the context of Jev, Almeida posits that the dramatically falling cost of "intelligence" – specifically, reliable, non-linguistic decision-making – will lead to its widespread and emergent deployment across countless software applications.
"We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed… much more like the early internet than you know like the mega apps that people are trying to build right now," Almeida articulated. This vision suggests a future where discrete, highly efficient AI components like Jev are integrated into the fabric of software, enabling a pervasive layer of intelligence without the need for monolithic, resource-intensive LLMs. This distributed intelligence could unlock unprecedented levels of automation and functionality across industries.
Architectural Innovations and Training Philosophy
While Diogo Almeida remains tight-lipped about the precise architectural details of Jev, outside observers speculate that it might be built upon an open-weight LLM, albeit heavily modified. TypeSafe AI refers to Jev as a "System One model," a term borrowed from cognitive psychology, which distinguishes between intuitive, fast thinking (System One) and deliberate, slow reasoning (System Two). Jev’s design prioritizes rapid, intuitive decision-making tailored for specific tasks, rather than broad-ranging reasoning.
A cornerstone of TypeSafe AI’s development philosophy, and a key differentiator for Jev, is its training methodology. Almeida revealed that Jev is trained exclusively on synthetic data, utilizing a novel technique he terms "reinforcement learning from calibrated decisions." This approach stands in contrast to RLHF, which relies on human-generated feedback. The reliance on synthetic data is a strategic bet that Almeida considers one of the best decisions his company has made. "Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data, and that is now my life joy," he shared with TechCrunch. This focus on carefully constructed synthetic datasets allows for greater control over the training process, potentially leading to more robust and less biased models.
Market Reception and the Future of AI
Jev currently occupies a unique niche in the AI landscape, standing as a distinct category of model. However, industry observers like Armin Ronacher anticipate that its demonstrated utility will inevitably spur competition. "We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don’t have to be creative yet," Ronacher commented, suggesting that the current abundance of LLM solutions may have stifled innovation in alternative AI architectures.
TypeSafe AI itself is not resting on its laurels, with plans to develop further versions of the model, exploring new modalities beyond its current capabilities. When questioned about TypeSafe’s ambition to be a "frontier lab," Almeida offered a philosophical distinction. He stated, "the main product of frontier labs is fear or hype. I would like our main product to be intelligence… [but we are] not a lab in the sense of, you know, like bet on infinite wealth, or a religion, or building God in a data center, or whatever is the thing of today." This statement underscores TypeSafe AI’s pragmatic, utility-driven approach, prioritizing tangible intelligence and practical application over grand, often speculative, AI aspirations.
The emergence of Jev signals a potentially significant re-evaluation within the AI community regarding the optimal architecture for various applications. While Large Language Models will undoubtedly continue to evolve and find profound uses in human-computer interaction and content generation, Diogo Almeida’s work with TypeSafe AI suggests a powerful alternative for the bedrock of intelligent automation. By focusing on cost-effective, hallucination-free, and probabilistically calibrated decision-making, Jev offers a compelling vision for a future where AI intelligence is not just powerful, but truly pervasive, reliable, and deeply integrated into the operational fabric of our digital world.
About the Author
