In the rapidly evolving landscape of artificial intelligence, the industry has been singularly focused on a specific pursuit: making machines sound more human. From the meteoric rise of ChatGPT to the development of increasingly complex Large Language Models (LLMs), the goal has been to master the nuance, wit, and structure of human discourse. Yet, for Diogo Almeida, a pivotal figure in the creation of Reinforcement Learning from Human Feedback (RLHF)—the very technique that unlocked the current era of generative AI—this obsession has been a costly distraction.
Almeida, who spent years at the heart of OpenAI, has reached a sobering conclusion: by optimizing AI for human conversation, we have inadvertently made it ill-suited for the mechanical precision required for modern automation. His new venture, TypeSafe AI, seeks to break this paradigm with the release of Jev, a transformer-based model that abandons the chatbot interface entirely in favor of "calibrated decisions."
The Core Problem: A Mismatch of Languages
The fundamental disconnect, according to Almeida, lies in the medium of exchange. "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."
Large Language Models are designed to predict the next token in a sequence to construct a sentence. They are probabilistic engines of prose. However, software automation requires deterministic or statistically reliable outputs—binary choices, confidence scores, and structured data—not a conversational summary. By forcing AI to "speak," developers have been saddled with models that hallucinate, are prohibitively expensive to run at scale, and suffer from high latency.
A Chronology of Disillusionment and Innovation
The journey toward Jev began long before the launch of TypeSafe AI. As an early researcher at OpenAI, Almeida was instrumental in building the scaffolding of modern AI. He witnessed firsthand how RLHF transformed models from experimental toys into world-changing utilities. Yet, as the technology matured, he observed a growing gap between what the models could do (write poetry, draft emails) and what businesses needed them to do (automate complex software workflows without errors).
Two years ago, Almeida made the strategic decision to exit OpenAI. He recognized that the architectural constraints of LLMs—specifically the overhead of text generation—were a permanent bottleneck for high-frequency, low-latency automation.
- 2022-2023: Almeida begins deep-diving into the limitations of LLM deployment for industrial software applications.
- 2024: TypeSafe AI is founded with a mission to move away from natural language generation.
- September 2026: TypeSafe AI officially unveils Jev, a model focused on "System One" thinking—fast, intuitive decision-making rather than slow, linguistic reasoning.
- The Present: The company scales its API, momentarily struggling to keep pace with an overwhelming influx of demand from enterprise developers.
The Mechanics of "Calibrated Decisions"
Unlike its conversational counterparts, Jev does not generate text. Instead, it outputs raw probabilities. In technical terms, the model acts as a highly optimized classifier that allows users to pre-define the potential outputs. Because the output space is constrained by the user, the model is physically incapable of "hallucinating" in the traditional sense.
Unpacking the Technical Advantages:
- Cost Efficiency: Because Jev does not generate long strings of conversational tokens, the computational footprint is significantly reduced. Input tokens are metered at a massive scale—by the billion rather than the million—making it economically viable for applications that require millions of decisions per hour.
- Speed and Latency: Without the need to decode complex linguistic sequences, the model’s inference time is cut down by orders of magnitude compared to standard LLMs.
- Reliability: The inclusion of "confidence scores" allows developers to treat AI outputs as data points. If a model returns a decision with a 50% probability, a developer can programmatically choose to ignore it. If it returns 95%, the system can execute the task automatically.
Industry Adoption: Evidence from the Field
The developer community has responded with immediate fervor. When TypeSafe launched its API, demand was so intense that the service briefly faltered—a testament to the pent-up need for "intelligence" that doesn’t require a chat window.
Pranit Sharma, a software engineer at Vercel, provided a compelling case study. His team had been using OpenAI’s Luna 5.6 to run a classifier for safety reviews on software commands. By migrating the workload to Jev, Vercel saw a performance jump that was 5 to 18 times faster, accompanied by an increase in classification accuracy.
Similarly, Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Google’s Gemini for the purpose of classifying business emails. While he noted that Gemini maintained a slight edge in raw linguistic nuance, Jev was 10 to 20 times more cost-effective. "It is the only one that hands back a real probability, which makes it ideal for automating workflows," Mudholkar noted.
Strategic Implications: Augmentation over Replacement
Almeida does not necessarily see Jev as the total successor to LLMs, but rather as an essential companion. In the current "agentic" era of AI, one agent is often tasked with checking the work of another. Using a massive, expensive LLM to monitor another LLM is a recipe for budget exhaustion.
Jev offers a leaner, more robust way to perform this oversight. By monitoring LLM "traces," Jev can act as a gatekeeper, identifying potential jailbreaks or logical failures in real-time. Armin Ronacher, CTO of Earendil, suggests that Jev could revolutionize model routing—the process of determining which AI model should handle a specific request. Because Jev is so inexpensive, it can act as a real-time dispatcher, sorting incoming requests and sending them to the appropriate, more specialized (and expensive) LLMs only when necessary.
The Philosophy of the "Jevons Paradox"
The name Jev is a deliberate nod to William Stanley Jevons, the 19th-century economist. The "Jevons Paradox" posits that as technological progress increases the efficiency with which a resource is used, the total consumption of that resource may increase rather than decrease. By making "intelligence" incredibly cheap and fast, Almeida believes we will see a proliferation of "smart software" that permeates the digital ecosystem.
"We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed," Almeida explains. He envisions a future that feels more like the early, decentralized internet than the current era of "mega-apps" and centralized AI platforms.
A Departure from the "Frontier Lab" Mentality
Perhaps most striking is Almeida’s rejection of the current "Frontier Lab" ethos. When asked about his vision, he distanced himself from the industry’s trend of "building God in a data center."
"The main product of Frontier Labs is fear or hype," Almeida said. "I would like our main product to be intelligence."
By focusing on synthetic data—specifically what he calls "reinforcement learning from calibrated decisions"—Almeida has built a lab that prioritizes statistical rigor over marketing bravado. His team creates their own data, betting that well-understood, proprietary datasets will outperform the "scraped-from-the-web" approach that has defined the last few years of AI development.
Conclusion: The Quiet Revolution
The rise of Jev signals a maturation point for the artificial intelligence sector. We are moving out of the "wow factor" phase, where the primary goal was to mimic human intelligence, and into the "utility" phase, where the goal is to integrate intelligence into the bedrock of software engineering.
While competitors will undoubtedly rush to replicate the success of TypeSafe’s model, Almeida’s early pivot has given his company a distinct advantage. By betting against the allure of the chatbot, he has uncovered a massive, underserved market: the need for reliable, fast, and cheap "System One" decision-making. As the dust settles on the AI gold rush, it is increasingly clear that the most valuable AI may not be the one that can win an argument, but the one that can perform a task with the silent, unwavering precision of a machine.
