The generative artificial intelligence boom has fundamentally reshaped consumer technology, putting sophisticated models front and center in the daily lives of millions. Breakthroughs such as ChatGPT introduced conversational intelligence to the global mainstream, while developer-focused tools like Claude Code transformed individual engineering workflows. Yet, despite these remarkable leaps in capability, a lingering question has persisted across the tech industry: if artificial intelligence models have achieved superhuman levels of intelligence, why hasn’t software experienced a corresponding explosion in true, autonomous operation?
Addressing this exact gap, a new startup named TypeSafe has emerged from stealth to pursue the next fundamental paradigm shift in software architecture. Rather than focusing on conversational interfaces or human-assisted workflows, TypeSafe is introducing models designed to power software directly at runtime, executing countless programmatic decisions every second without any human intervention. The company’s debut model, named Jev, represents a calculated bet on shifting artificial intelligence from a slow, conversational assistant into the high-speed decision-making engine of modern applications.
The market response to TypeSafe’s vision has been swift and extraordinary. Within twenty-four hours of its official launch on September 15, Jev became the fastest-adopted model in the history of Vercel’s AI Gateway. Over the subsequent weeks, technology investors and early-stage venture funds have reported a massive surge in interest, as portfolio companies race to integrate the model into production environments for a wide array of new use cases. Industry observers note that this level of intense developer pull and organic adoption has not been witnessed across the ecosystem since the early days of GPT-3.
Decisions Instead of Strings
To understand the core innovation behind TypeSafe, one must examine how Jev fundamentally inverts the traditional mechanics of large language models. Standard language models operate by generating text sequentially, predicting and writing out answers token by token, word by word. This token-by-token generation makes traditional models inherently slow, deliberate, and expensive—characteristics that align them with what cognitive psychology describes as System 2 thinking: analytical, methodical, but unsuited for rapid, high-frequency operations.
Jev, by contrast, is engineered to act as artificial intelligence’s System 1: fast, cheap, intuitive, and built specifically to handle the thousands of small, routine judgment calls that software applications must navigate every single second. Instead of outputting sprawling strings of natural language text, Jev ingests comprehensive context alongside a structured set of queries, returning typed answers complete with confidence scores all at once. This architectural shift transforms how models interact with application logic, moving away from creative text generation and toward precise, high-throughput classification and decision-making.
The practical applications for this capability are vast and varied. Developers are utilizing Jev to determine which interface button an autonomous software agent should click next, grade evaluation metrics, route incoming requests, triage support tickets, and extract specific data fields from complex legal contracts. The model offers an out-of-the-box generality that functions with minimal friction, requiring no specialized labeled datasets or dedicated machine learning training teams to deliver immediate utility. Since its rollout, developers have pushed the boundaries of the technology in unexpected directions, deploying Jev for tasks ranging from playing the classic video game Doom and dynamically re-ranking personal news feeds to generating expansive Super Mario multiverses.
Jevons Paradox, by Design
The nomenclature chosen for TypeSafe’s flagship model is far from arbitrary. Jev takes its name from the economic concept known as Jevons paradox, which dictates that when technological advancements or resource efficiencies make a commodity dramatically cheaper to produce and consume, society ultimately uses vastly more of it rather than saving the surplus. TypeSafe is staking its foundational thesis on the prediction that the exact same economic dynamic will apply to artificial intelligence and machine intelligence.
Historically, when classification tasks or machine learning evaluations were expensive and computationally demanding, development teams were forced to ration their usage. Engineering organizations would evaluate only a small random sample of autonomous agent runs rather than auditing the entire operational pipeline. They would write brittle, hardcoded rules instead of invoking a flexible model, or they would bundle multiple inquiries into one broad, generalized question rather than executing ten specific queries.
When intelligence and classification become nearly free and instantaneous, those limiting tradeoffs instantly evaporate. This deflationary trajectory fundamentally alters what is economically viable to build. Many of the most persistent operational problems across modern industries are structured in a way that benefits directly from high-volume, low-cost decision-making—a reality underscored by the immediate explosion of volume and creative use cases discovered by early adopters. Industry analysts suggest that the most transformative applications built on top of infrastructure like Jev are likely still conceptual, waiting to be imagined and engineered.
Prod, Not God
The intellectual origins of TypeSafe trace back to the formative experiences of its co-founder and CEO, Diogo Almeida, who previously worked on InstructGPT and reinforcement learning from human feedback (RLHF) at OpenAI. During that period of intense research and model development, the industry focus centered heavily on training models to become exceptional conversational partners—capabilities optimized specifically for human interaction and human-in-the-loop oversight. However, Almeida recognized that achieving true, scaled automation required an entirely different technical blueprint. Real automation demanded artificial intelligence that prioritized operational reliability, high execution speed, and low cost above conversational charm.
This realization led Almeida to a conclusion he describes as the bitterest lesson of machine learning: selecting and defining the exact problem space to solve matters significantly more than raw computational scale. Born in the Philippines and formerly ranked as a top thirty North American League of Legends player, Almeida combines an iconoclastic, creative research background with sharp developer sensibilities and deep commercial instincts. Observers note that this unusual combination of unbridled ambition and pragmatic engineering realism provides the exact temperament required to navigate the transition into the next machine age.
While Jev has successfully captured significant industry attention right out of the gate, TypeSafe views this initial release merely as an opening salvo. The company’s long-term ambition extends far beyond a single model; TypeSafe aims to establish the foundational infrastructure layer for how software systems make runtime decisions. By offering a comprehensive platform of modular primitives that developers can compose and build upon, the company intends to provide the structural building blocks for modern intelligent software in much the same way traditional databases once powered the architecture of the web.
Backing this vision, venture capital firm a16z and other early partners have thrown their support behind the company. The alignment between TypeSafe’s leadership team—including co-founders Sasha Sheng and Erik Spock Gafni—and its investors highlights a shared conviction that the bottleneck in artificial intelligence adoption is no longer raw model intelligence, but rather the availability of fast, reliable infrastructure designed to put that intelligence directly into production.
Comments (0)