For years, the narrative surrounding Artificial Intelligence has been defined by the pursuit of human-like conversation. From the early breakthroughs of GPT-3 to the current era of multimodal agents, the industry has fixated on making machines sound more like us. But for Diogo Almeida, one of the original architects of OpenAI’s reinforcement learning from human feedback (RLHF)—the very mechanism that made ChatGPT’s conversational fluidity possible—this fixation has been a profound misdirection.
"We have lightning in a bottle, and yet it is not useful," Almeida told TechCrunch in a recent interview. After leaving OpenAI two years ago, Almeida founded TypeSafe AI with a singular, disruptive thesis: the industry’s obsession with human language as the primary interface for AI is actually the greatest barrier to meaningful machine automation.
This week, TypeSafe AI took a bold step toward proving that thesis by launching "Jev," a transformer-based model that refuses to speak. By stripping away the conversational layer, Jev promises to reshape how developers integrate intelligence into software, signaling a pivot from "chatty" AI to "decisive" AI.
The Core Problem: Why Computers Don’t Need to Talk
The fundamental disconnect, according to Almeida, lies in the medium of exchange. Large Language Models (LLMs) are optimized for human communication—nuanced, probabilistic, and prone to the inherent ambiguities of syntax and semantics. However, computer systems operate on binary logic and structured protocols. When developers force an LLM to interact with a machine backend, they are essentially asking a poet to perform complex accounting.
"We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language," Almeida explains.
Jev represents a radical departure from the LLM paradigm. Instead of outputting natural language tokens, Jev outputs calibrated probabilities—what the company calls "calibrated decisions." By eschewing the need to generate fluent text, the model becomes significantly faster and cheaper to run. Because the potential outputs are defined by the user in advance, the model is physically incapable of "hallucinating" in the traditional sense. It doesn’t guess at a sentence; it calculates the likelihood of a specific, predefined action.
Chronology: From OpenAI to a "System One" Future
To understand the emergence of TypeSafe, one must look at the evolution of the current AI boom.
- 2020–2022: Almeida serves as a key researcher at OpenAI, contributing to the development of RLHF. During this time, he witnesses firsthand the limitations of training models on human preferences. He observes that while RLHF makes models likable and coherent, it often obscures the underlying reliability required for robust engineering.
- 2022: Almeida departs OpenAI, concluding that the path toward Artificial General Intelligence (AGI) via conversational agents is being hampered by the "language bottleneck." He begins incubating TypeSafe AI.
- 2023–2025: TypeSafe focuses on internal research, pioneering techniques in synthetic data generation. Unlike the "frontier labs" that rely on scraping the entire internet, TypeSafe invests heavily in "reinforcement learning from calibrated decisions."
- September 2026: TypeSafe officially unveils Jev. Demand is so immediate that the startup experiences temporary service outages, marking a rare moment in the AI industry where technical demand outstrips the supply of a non-LLM specialized tool.
Supporting Data: Efficiency at Scale
The performance metrics associated with Jev have already caught the attention of major engineering outfits. In early testing, developers are finding that Jev is not just a replacement for LLMs, but a superior tool for specific functional tasks.
Pranit Sharma, a software engineer at Vercel, noted that his team replaced an OpenAI "Luna 5.6" model with Jev for safety-classification tasks. The results were stark: Jev was 5 to 18 times faster than its conversational predecessor, with improved accuracy.
Similarly, Nikhil Mudholkar, CTO of Bryo AI, conducted a comparative test between Jev and Google’s Gemini for classifying incoming business emails. While Gemini offered a slightly higher raw accuracy in some contexts, it proved to be 10 to 20 times more expensive. More importantly, Mudholkar highlighted the "confidence score" feature. "It is the only one that hands back a real probability, which makes it ideal for automating workflows," he noted.
For developers, the ability to receive a numerical confidence level—rather than a vague, text-based assessment—is the difference between a brittle script and a production-ready automated system.
Implications: The Death of the "Mega App"
The name "Jev" is a deliberate homage to the 19th-century economist William Stanley Jevons. The "Jevons Paradox" posits that as technology increases the efficiency with which a resource is used, the total consumption of that resource increases rather than decreases. Almeida believes this applies perfectly to intelligence. As the cost of "smart" decision-making drops toward zero, we won’t see fewer AI models—we will see them embedded into everything.
"We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed," Almeida said. "Much more like the early internet than the mega-apps that people are trying to build right now."
The "System One" Architecture
The industry is currently debating how Jev works under the hood. While Almeida remains tight-lipped about the exact architecture, experts suspect the model leverages open-weight LLMs as a base, which are then refined through specialized synthetic data.
Almeida prefers the term "System One model," a nod to the psychological framework popularized by Daniel Kahneman, referring to fast, intuitive, automatic thinking—as opposed to the slow, deliberate "System Two" reasoning often associated with the high-latency "Chain-of-Thought" prompting currently popular in the LLM space.
Preventing the "Hallucination" Trap
The utility of Jev extends beyond replacing LLMs; it functions as a "guardrail" for them. By deploying Jev as a monitor, companies can track LLM agent behavior in real-time. If an agent begins to veer off-script or exhibit signs of a jailbreak, Jev can calculate the probability of that deviation instantly.
Armin Ronacher, CTO of Earendil, notes that this shifts the burden of responsibility in a productive way. "At the end of the day, it delegates the hallucination problem a little bit to the user," Ronacher explains. "The user has to say, ‘Okay, if this only comes back with 50% probability, maybe this is a coin toss.’ But if it’s 95%, sure, I can do something with it."
Official Responses and Industry Outlook
The reaction from the broader developer community has been one of belated realization. As Ronacher suggests, the industry’s reliance on "subsidized" LLMs—where large labs provide cheap, powerful conversational models—has allowed many developers to ignore the need for architectural creativity. Now that the limitations of conversational agents are hitting a ceiling, the value of specialized, cost-effective models like Jev is becoming undeniable.
When asked if TypeSafe intends to become a "frontier lab"—a term often associated with the high-stakes, capital-intensive race toward god-like AGI—Almeida is dismissive. "The main product of Frontier Labs is fear or hype," he says. "I would like our main product to be intelligence."
For now, TypeSafe AI is focused on expanding Jev into new modalities. By proving that synthetic data and calibrated decision-making can outperform the massive, brute-force language models of the current era, the company has effectively declared that the future of AI isn’t in what the machine says, but in what it can reliably do.
As the AI bubble continues to inflate with expensive, text-generating agents, Jev stands as a quiet, efficient reminder that sometimes, the most intelligent thing a machine can do is stay silent and simply get the job done.
