Executive Overview
The generative AI boom, for all its dazzling poetry, image generation, and conversational fluidity, suffers from a foundational identity crisis. According to Diogo Almeida—an early OpenAI researcher instrumental in inventing reinforcement learning from human feedback (RLHF), the breakthrough training technique that catalyzed the modern AI era—the industry has fallen into a trap. We have managed to bottle lightning, but we are using it exclusively to write prose.
Two years after quietly walking away from OpenAI, Almeida has resurfaced with a radical thesis and a disruptive new product. Through his startup, TypeSafe AI, Almeida has launched Jev, a transformer-based model that explicitly rejects large language models (LLMs). Instead of generating text, stories, or conversational filler, Jev outputs probabilities, or what TypeSafe terms "calibrated decisions."
By stripping away the complex overhead of human language generation, Jev bypasses the industry’s most persistent headaches: hallucination, prohibitive compute costs, and sluggish inference speeds. Early developer metrics indicate that Jev can execute classification and safety-filtering tasks between 5 and 20 times faster than top-tier LLMs, while costing a fraction of the price.
As the developer community rushes to adopt this alternative paradigm, TypeSafe AI is sparking a necessary industry-wide reckoning. The question is no longer just how cleverly an AI can converse, but whether human language was ever the right interface for automation in the first place.
Detailed Chronology: From OpenAI’s Breakthroughs to TypeSafe AI
The Disillusionment at OpenAI
To understand the genesis of TypeSafe AI, one must return to the formative days of the generative AI explosion. Diogo Almeida was in the trenches at OpenAI, helping construct the foundational architectures of ChatGPT. More importantly, he helped invent RLHF—the process by which human preference data shapes model outputs, steering neural networks away from toxic or nonsensical language and toward coherent, helpful dialogue.
Yet, as the technology scaled and the world marveled at ChatGPT’s conversational depth, Almeida experienced a profound sense of disappointment. The models were brilliant at mimicking human rhetoric, but when it came to practical software automation, they were cumbersome, unpredictable, and expensive.
"We have lightning in a bottle, and yet it is not useful," Almeida reflects. "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."
Realizing that the industry was careening down a dead-end street by forcing probabilistic intelligence through the bottleneck of human text, Almeida made the bold decision to leave OpenAI two years ago. He set out to build something fundamentally different—a company dedicated to operationalizing intelligence without the linguistic baggage.
The Birth of TypeSafe AI and the Launch of Jev
Operating largely in stealth, Almeida founded TypeSafe AI to tackle the automation bottleneck head-on. This week, the startup pulled back the curtain to release its flagship transformer-based model, named Jev in honor of 19th-century economist William Stanley Jevons and his famous paradox regarding resource efficiency.
Jev represents a dramatic departure from standard LLM architecture. It does not output words, tokens of prose, or chat logs. Instead, it computes and returns strict, quantifiable probabilities—calibrated decisions tailored specifically for backend software systems.
Because Jev avoids the massive computational overhead required to generate syntax and grammar, its operational footprint is drastically lighter. Output tokens are entirely free, and input tokens are metered on a vastly different scale—by the billion rather than the million. Most importantly, because the possible outputs are predefined by the user prior to execution, the model cannot hallucinate.
The market response was immediate and overwhelming. Demand for Jev’s API upon launch was so fierce that TypeSafe AI briefly lost its capacity to serve incoming traffic—a testament to the pent-up frustration developers feel with bloated, over-engineered LLMs.
Supporting Context & Metrics: How Jev Stacks Up
The industry’s embrace of Jev is not merely philosophical; it is driven by hard performance metrics in production environments. Software engineers grappling with the latency and financial drag of traditional LLMs are finding that Jev offers a leaner, more deterministic alternative.
Real-World Developer Benchmarks
- Vercel’s Safety Classifier: Pranit Sharma, a software engineer at agentic infrastructure firm Vercel, tested Jev against OpenAI’s ChatGPT Luna 5.6 in running a classification system to review commands for safety. When Vercel swapped out OpenAI’s model for Jev, the performance leap was staggering: execution speeds were 5 to 18 times faster, paired with enhanced accuracy.
- Bryo AI’s Email Automation: Nikhil Mudholkar, CTO of Bryo AI, benchmarked Jev against Google’s Gemini for business email classification. While Gemini offered a marginal edge in raw semantic accuracy, it proved to be 10 to 20 times more expensive to run. More critically, Mudholkar highlighted Jev’s confidence scoring system: "It is the only one that hands back a real probability, which makes it ideal for automating workflows!"
Augmenting and Guard-dogging LLMs
Rather than viewing Jev purely as a replacement for language models, industry observers note its unique capacity to augment existing LLM workflows. Running complex multi-agent LLM systems to monitor other agents quickly introduces unsustainable compute costs. Deploying Jev as an oversight mechanism—tracking LLM agent traces and actively preempting jailbreaks—presents a cost-effective alternative.
Armin Ronacher, CTO of Earendil (which builds the open-source model harness Pi), explains how developers interact with Jev’s outputs:
"At the end of the day, it delegates the hallucination problem a little bit to the user. 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 also points to model routing as a prime use case. Accurately predicting whether a specific computational workload demands an expensive frontier model or a lightweight alternative is notoriously difficult. Jev’s high speed and low cost make real-time intelligent routing economically viable for the first time.
Synthetic Data and Architecture
While Almeida remains tight-lipped about the exact structural blueprint of Jev—leading outside observers to suspect it leverages an underlying open-weight LLM—he emphasizes that the company classifies Jev as a "System One model." Prioritizing rapid intuition over recursive reasoning, Jev is tailored specifically for targeted classification tasks.
Crucially, TypeSafe AI broke from conventional industry pipelines by training Jev exclusively on synthetic data using a proprietary technique Almeida calls "reinforcement learning from calibrated decisions."
"We made an early bet that we will be making all of our data, and that has been one of the best bets I’ve ever made in my life—better than our launch, in my opinion, better than RLHF," Almeida shares. "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."
Official Statements & Industry Perspectives
The emergence of TypeSafe AI has forced a critical dialogue among software architects and AI veterans regarding the actual utility of current frontier labs.
Armin Ronacher notes that the industry’s heavy reliance on language models stems largely from how heavily subsidized and ubiquitous LLMs have been. Because developers have had cheap access to text-generating behemoths, the incentive for deep architectural creativity has been muted until now.
When asked whether TypeSafe AI considers itself a "frontier lab" in the vein of OpenAI or Anthropic, Almeida draws a sharp boundary line:
"The main product of frontier labs is fear or hype. I would like our main product to be intelligence… [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."
Instead, Almeida’s vision aligns with the Jevons paradox: as the economic cost of intelligence trends toward zero, its deployment footprint must expand exponentially.
Future Outlook: The Rise of Distributed Intelligence
TypeSafe AI’s early success points toward a paradigm shift in how artificial intelligence will integrate into the digital landscape. Rather than consolidating compute into massive, centralized "mega-apps" hosted by a handful of monopolistic cloud providers, Almeida envisions a decentralized, emergent future.
"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 predicts.
As competitors inevitably enter the calibrated-decision space—recognizing the clear market appetite for fast, un-hallucinatable, probabilistic models—the era of defaulting to an LLM for every computational problem may be coming to a close.
TypeSafe AI has proven that stepping away from human language is not a step backward in capability; rather, it is the key to unlocking the true, utilitarian potential of machine intelligence. By teaching computers to speak their own native language of probabilities, Almeida and his team have moved past the parlor tricks of chat and laid the groundwork for the next generation of industrial software automation.
