Someone Rebuilt TypeSafe AI's Jev in 25 Lines of Local Python
Follow-up to: 2026-09-21, TypeSafe AI's Jev Skips Text Entirely and Just Answers Structured Questions for $0.042 per Million Tokens
Three days ago I wrote about Jev, TypeSafe AI's product for answering structured, multiple-choice-style questions without generating free text, priced at $0.042 per million tokens. This week, a developer published a post titled "Jev in 25 Lines of Python," an explicitly parodic reimplementation of the same core mechanism, running locally against an open model, no API key and no per-token bill required.
What the 25 lines actually do
The post targets NobodyWho, an open-source local-model project, and it's upfront about being a parody rather than a production-ready competitor. But the mechanism it reproduces is the real one: give a model a prompt with a fixed set of choices, and instead of generating free text, read off the probability the model assigns to each choice. That's classification via logits, a technique that predates any of the current wave of AI product wrappers by years. The post skips TypeSafe's RLCD training process and its synthetic data generation pipeline entirely, and it runs against a model sitting on your own machine instead of a hosted API.
The author's stated case is privacy and cost: nothing leaves your machine, and you're not paying a per-token fee for something that's fundamentally reading a probability distribution the model was already computing internally. They point to NobodyWho as the fuller open-source path, while noting more complete community implementations already exist under the OpenJev name.
Whether this actually threatens Jev's business
I want to be careful here, because "someone wrote a parody blog post" is not the same claim as "a product's moat has collapsed." TypeSafe AI's pricing, hosting, reliability guarantees, RLCD-tuned model quality, and support are all real value that 25 lines of local Python doesn't replicate, even if the core classification mechanism is genuinely the same idea. A hosted API with SLAs and a tuned model is a different product from a weekend script, even when they're doing structurally the same thing under the hood.
But the honest version of the take is this: the mechanism is not the moat. If Jev's entire value proposition rested on "we're the only ones who know how to turn a prompt with fixed choices into a probability output," this post is a direct rebuttal, published in under 30 lines, two days after their pricing made headlines here. What actually holds up TypeSafe's business is everything around the mechanism: the training investment in RLCD, the synthetic data generation, the hosted infrastructure, and whatever accuracy gains that tuning produces over a raw base model doing the same trick. Whether that gap is worth $0.042 per million tokens versus free and local is a real question, and it's one you can only answer by comparing actual accuracy on your own use case, not by reading either blog post.
The pattern worth noticing
This is not the first time a "sophisticated" AI product has turned out to be a thin wrapper around a technique that's been public for a while, and it won't be the last. The interesting part isn't that it happened to Jev specifically. It's how fast the reveal came: three days from "here's a paid product with a specific price point" to "here's the same trick, free, in 25 lines." That compression is new, and it's a direct result of how easy it now is to stand up a local model and test a hypothesis about what a paid API is actually doing under the hood.
What I'd actually do
If you're building on Jev or a similar structured-output API today, don't panic-migrate off it based on a parody post, but do run the actual comparison: take your real prompts and choice sets, run them against a local open model using the same logit-reading trick, and measure accuracy against what you're currently paying for. If the gap is small for your specific use case, you've found a real cost reduction. If the gap is large, that's your answer for why the hosted product's price is justified, and you now have the receipts to know it rather than assuming either way.
The honest counter-take: I have not run this comparison myself against production data, and neither has the parody post's author, as far as the source indicates. Treat "the mechanism is the same" as a starting hypothesis to test against your own accuracy requirements, not as proof that the paid product is overpriced.
Author
Lukas
@lukcombinator