NobodyWho Reduces TypeSafe's Jev to 25 Lines of Python

NobodyWho, a Copenhagen on-device inference startup, published a blog post on Sept. 22 that reproduces the core mechanic of TypeSafe AI's Jev in 25 lines of Python, using an unmodified, off-the-shelf Qwen3-0.6B model instead of the custom-trained system TypeSafe sells.

The post, published on NobodyWho's company blog and bylined to Duarte O. Carmo, loads Qwen3-0.6B-GGUF through the open-source llama-cpp-python library, reads the model's raw next-token scores for three labeled choices, and converts them to probabilities with a softmax function. Run on a sample phishing email, the 25-line script produced a clear ranking among the three options.

NobodyWho itself sells an open-source, on-device inference engine that runs large language models on phones and desktops rather than in the cloud, and says on its own site that it raised 2 million euros in December 2025 to build what it calls European AI infrastructure from Copenhagen. The post's closing pitch, urging readers to check out NobodyWho because it keeps data local rather than sending it to an API, doubles as an advertisement for the company's own product.

Choice	Probability
Legitimate	0.031
Spam	0.084
Phishing	0.885

What NobodyWho admits it skipped

The post is explicit about what its own script lacks. "We don't call it a System One decision model," it says. "We didn't call an API. We didn't create a bunch of synthetic data. We didn't train a model with Reinforcement Learning for Calibrated Decisions (RLCD) to calibrate the decisions and probabilities (even though they are not always correct). But yes. This is Jev." NobodyWho labels the post a parody in its own footer, while still presenting the code as functionally equivalent to Jev's classification behavior.

What TypeSafe says makes Jev different

TypeSafe opened early access to Jev on Sept. 15, charging $0.042 per million input tokens and nothing for output, according to the company's own announcement. Almeida, who wrote that he "helped build the methods that made language models useful at following instructions" during his time at OpenAI, credits Jev's behavior to RLCD, a post-training method TypeSafe says teaches calibrated confidence rather than raw next-token prediction. TypeSafe's launch materials do not disclose Jev's parameter count or the training data behind RLCD, gaps NobodyWho's post does not fill either.

The community response NobodyWho points to

NobodyWho's post closes by linking to three follow-on open-source projects it calls more complete than its own 25 lines: OpenJev, openjev-sglang, and an OpenJev build on Google's DiffusionGemma. One of those, AlexWortega's openjev, turns Alibaba's Qwen3.5 into an MIT-licensed cross-encoder that scores contradiction, entailment and neutral labels for reranking and guardrail tasks, hosted on Hugging Face under its own model card. It is the second widely shared rebuttal this month that undercuts an AI capability claim's framing without disputing the underlying result, following an audit that found gaps in GPT-6 Astra's Enigma-cipher solve.

What is still unresolved

Neither company's post addresses the other directly. NobodyWho's code does not attempt RLCD-style calibration training, so it cannot show whether Jev's confidence scores are more reliable than a stock model's raw logits, the specific claim TypeSafe's pricing and marketing rest on. TypeSafe has not published a benchmark comparing Jev's calibration against an untrained model like the one NobodyWho used.