On September 15 (US time), the top of Hacker News was taken over by a launch that sounds like fiction: TypeSafe AI, a San Francisco company that spent two years in stealth, announced it had built a model that "cannot hallucinate." It is named Jev. The same day, a $40 million seed round landed, led by DCVC.
The word "hallucination" has been discussed in AI circles for four years, and every vendor opens with the same line—we will hallucinate less and less. TypeSafe is the first to declare that hallucination simply does not exist.
That bold claim deserves serious attention. I went through their blog, public eval data, hidden FAQ answers, and over two hundred HN comments. Conclusion up front: there is something real inside, but it is wrapped in four layers of packaging.
What They Are Actually Selling
Kahneman split human thinking into two systems. System Two is slow, handling reasoning and writing; System One is fast, judging in 0.1 seconds whether there is a snake in the grass. TypeSafe's bet: most AI judgments in software are System One work.
Is this email a complaint? Is this transaction fraudulent? Should this agent trajectory be blocked? Today we ask large language models these questions, have them write a short essay, then scrape the answer out of the essay. TypeSafe cuts out the entire monologue. Input goes in, output is directly a predefined type—boolean probabilities, a distribution over options, or a score. Three primitives, named Noul, Choice, and Score. All questions are computed in parallel at once; adding a hundred new questions barely moves response time.
Engineers who have shipped production systems will smile knowingly. This is the glue of "call GPT + parsing code + a retry loop" packaged into a single API.
What about the architecture? Secret. The paper "might be written." The community's consensus characterization, endorsed by the founder, is "a generalized zero-shot classifier." Nothing shameful about that—doing classification well and cheaply is a big business.
The Numbers Start Nesting
The homepage slogan: 193.6x faster, 444.6x cheaper. Numbers precise to one decimal place deserve a look at their provenance.
They come from four business-process benchmarks built by the company itself. The comparison LLMs were wrapped in TypeSafe's own wrappers, which the company admits made the opponents slower and costlier. The "ground truth" for benchmark tasks was not human annotation but consensus between two models, GPT-6 Astra and Claude Fable 5.1. Their own tasks, their own wrappers, their own pipeline, with competitor consensus as truth—a fourfold overlap of identity.
More interesting is how the number travels. The same company's press release only dared say "up to 100x faster." The homepage's own comparison example shows 75x and 171x. Dividing the two closest items in the public eval table yields roughly 195x and 440x. The order of magnitude is credible. The decimal-point precision, nobody can reproduce.
This chart is their own, and it is the most honest thing on the page. Jev scored 67.8%, six points below the strongest, sol. On the invoice-processing task, it trailed by seventeen points. It wins on another dimension—accuracy in the same tier, at 1/40th of a second versus half a minute, and $0.0004 versus eight cents. Pareto-optimal: trading six points for cost-performance.
So What About "No Hallucination"
This is the best part of the whole piece. The blog has a passage explaining: Jev's hallucination rate in the chart is 0%, but that number "is not measured; output and schema matching are constructively guaranteed, so we can safely write 0%."
Translated: Jev guarantees it will never output anything outside the options—if you ask for a boolean, it can never answer with an essay. Mathematically true. But is a wrong yet valid answer a hallucination? Many on HN say yes. The model confidently says "this email is a routine inquiry" when it is actually a complaint—the output is perfectly legal, the judgment perfectly wrong.
The official docs twist the knife themselves: calibration holds only for the overall distribution of predictions, "individual responses are not guaranteed correct." Backed into a corner, the founder conceded the model "can absolutely be confidently wrong."
A model that cannot say a wrong-shaped word can still say wrong things. That is the precise description of the slogan.
The People Are Real, the Money Is Real
Having taken it apart, credit the foundation. Founder Diogo Almeida is one of the authors of the InstructGPT paper—the one that trained ChatGPT's behavior. The marketing line "co-invented RLHF" is self-flattery (the method predates the paper), but the authorship is real. The CTO came from Meta FAIR; the COO is a repeat founder. The $40M seed round, per Dealroom, ranks in the top 1% of AI seed rounds historically. The team works at the San Francisco waterfront, five days a week in office.
The model's name is not casual. Jev, as in the Jevons Paradox—the economist who found that every efficiency gain in steam engines led to more coal burned. A model selling "cheap" is named "cheaper means more use." Whoever named it knew exactly what they were doing.
What Is Worth Keeping
HN that day had plenty of abuse—"the most misleading marketing in months." Others posted Doom gameplay in praise—Jev reads the game state ten times a second to make decisions, at $7 per hour, though it reads textualized state rather than pixels, and the company itself admits "you could write a better bot without AI."
The critics and the fans are seeing the same thing. The real innovation is the interface: judgments as type-safe primitives, confidence as a first-class return value, twenty-plus cookbooks demonstrating reranking, guardrails, citation checking—real tasks. The eval site lays out all data, including their own ugly numbers. The founder answered forty-plus questions online, conceding everywhere. That transparency is rare in launch marketing.
And the suspicious part is the same thing. No public benchmark scores, with the FAQ literally saying "don't take public benchmarks seriously." Black-box architecture, no weights, no paper. A model proven only by its own evals is persuading the pipelines of ten thousand factories to trust its judgment.
My advice remains the old principle: buy its speed and price; leave "no hallucination" to the marketing department. Whether to trust its judgment—wait for third-party evals, or for that "might be written" paper.
System One is fast, too fast to explain itself. Humans spent a lifetime learning to attach review to intuition. The same rule for machines will probably not be an exception.
---
*Sources: typesafe.ai official blog and FAQ, evals.typesafe.ai public eval table, Dealroom coverage, Business Wire release, HN thread 49717558 comments (734 points), docs.typesafe.ai. Full fact-check records in factcheck_r46t.*