NED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS ITNED SHIPS IT
A free guide from the Ned Ships It video
What Is Jev? The AI That Only Decides, Plus Laya
Big models write. Decision models decide.
Commented JEV? This box is yours.
From
Ned, @nedshipsit
To
You
In the box
3 free tools
Checked
October 7, 2026
Why sorting with a big model is so expensive
A big model like Claude or ChatGPT writes its answer one word at a time. Every word takes time and every word costs money. That's perfect for writing a reply, and wasteful for sorting one, because sorting isn't writing. It's a decision: is this email an order, a complaint or a question?
Jev, from TypeSafe, is built only for decisions. You send the text and your options, and it scores every option in one pass and picks. It never writes a single word. That's why it answers in 70-500 ms and costs $0.042 per million input tokens, with output free, per TypeSafe's docs and pricing.
The math from the video: 200,000 items at about 500 tokens each is 100 million tokens. At $0.042 per million, that's about $4.20.
Kevin used a novelist to sort the mail. The novelist wrote a lovely essay about every envelope. Then sent the bill.
New to this? Start hereWhat you need before you try Jev or Laya
For Jev: a free TypeSafe account and an API key from the TypeSafe console. TypeSafe opened to everyone on September 27, 2026.
For Laya: Python on your computer. It runs on CPU, Apple Silicon or a GPU.
Your questions written down. Jev only answers the question you ask, literally. Write the options and one line describing each before you start.
Where things go: lines starting with npm, uv, pip, laya or claude are terminal commands. Plain-English prompts go in the Claude Code chat.
Paying a writing model to make choices. Jev takes a state (your text) and typed questions: a Choice (pick one of up to 255 options), a Noul (yes/no, as a probability) or a Score (2-10 levels). All questions in one call run together against the same text. The endpoint is POST https://api.typesafe.ai/v1/systemone with "model": "jev-latest".
The easiest way is to let Claude Code learn it. TypeSafe publishes an official skill:
Paste in Terminal
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
Prefer the SDK yourself? npm install @typesafe-ai/sdk or uv add typesafe-sdk.
Try it
Paste this in the Claude Code chat.
You, to Claude
Use the TypeSafe skill to build a small script that sorts customer emails with Jev.
Ask one Choice question: "What kind of email is this?" with the options order, complaint, question and other, each with a one-line description.
Also ask a Noul question: "Does the customer ask for a refund?"
Read my TypeSafe API key from an environment variable called TYPESAFE_API_KEY. Never print it.
Run it on these 3 test emails and show me each answer with its confidence:
1. "Can I get 3 of the blue ones by Friday?"
2. "My flowers arrived dead. I want a refund."
3. "Do you deliver on Sundays?"
Heads up Jev reads text only (no images, audio or video), is weak at math, counting and dates, and reads your question literally. More
Add an other option whenever the list might not cover everything. The request limit is 64k tokens. English is its main language; other languages work but not equally well.
Sending every email to someone else's server, and paying per token. Laya is an open-source decision model from Convai Innovations (Apache-2.0, 31.5k stars on GitHub). It returns the same kinds of typed answers, runs on your own machine, and its server copies Jev's API shape.
Then run it as a local Jev-compatible server and point your app at it:
You, to Claude
pip install "laya[serve]"
laya-serve
It listens on port 8000 and accepts the same POST /v1/systemone request as Jev. In your code, change https://api.typesafe.ai/v1/systemone to http://localhost:8000/v1/systemone.
Heads up Laya's own benchmarks are honest about the gaps. More
It gets much less accurate with more than about 20 options, Score questions are its weakest type, and the base checkpoints aren't a zero-shot decision engine: use the fine-tuned typed-decisions checkpoint and test on your own data. The server has no password unless you set LAYA_API_KEY, so don't expose it to the internet as is.
Same idea, your laptop, no bill. Kevin cried a little.
Not trusting AI to act alone. Every Choice and Score answer comes with a confidence number. The pattern: when the model is very sure, your software acts on its own. When it isn't, a human gets it.
Extend my email sorter:
- If the Choice confidence is 0.95 or higher, label the email automatically.
- If it's lower, put it in a "Needs a human" list instead.
Log every decision with the email id, the answer and the confidence, so I can check how often the auto-labels were right.
Run it on my last 200 emails and show me: how many were auto-labeled, how many went to a human, and 10 random auto-labeled examples for me to spot-check.
Try it
Start strict (0.95) and spot-check. If the auto-labels are right almost every time, lower the bar a little. If not, improve the option descriptions before you touch the threshold.
Heads up Confidence is what Jev is trained to report, but your data decides how much to trust it. More
Keep a human on anything that moves money or talks to a customer until the log shows it's reliable.
If it still breaks
Answers look random. Rewrite each option's one-line description so they don't overlap, and add an other option.
Everything goes to "Needs a human". Your threshold is too strict for your data, or your options are too similar. Fix the options first.
Laya gets a big list wrong. Split it into two steps: first a few broad groups, then the detail inside the group you picked.
Dates or amounts come out wrong. That's a known weak spot. Pull numbers and dates out with normal code, and use Jev or Laya only for the judgment.
401 errors from Jev. Check that TYPESAFE_API_KEY is set in the same terminal and that the key is active in the TypeSafe console.
Questions people ask
What is Jev?
Jev is TypeSafe's first System One model, launched September 15, 2026. You send it some text (the state) and typed questions, and it returns answers your code can use directly: one choice from a list, a yes/no probability, or a score, with probabilities and a confidence number. It never generates text.
How is Jev different from ChatGPT or Claude?
Chat models write their answer one token at a time, which is great for writing and slow and pricey for picking. Jev scores every option in a single pass instead. TypeSafe lists 70-500 ms end to end and $0.042 per million input tokens, with output free. Use a big model to write and Jev to decide.
Is Laya really free?
Laya's code and weights are Apache-2.0, so you can run it yourself for free; you pay only for your own computer or server. It runs on CPU, Apple Silicon or a GPU, and laya-serve speaks the same POST /v1/systemone shape as Jev.
Is Jev accurate?
It depends on your task and how well you write the options. TypeSafe's own speed and price comparisons come from its own workflows, and TypeSafe says those gains are on the higher end of real world results. Independent reviewers ask the same question: test it on a few hundred of your own examples before you trust it.
What can't Jev and Laya do?
They read text only (no images, audio or video), they're weak at math, counting and dates, and they read your question literally. Laya also gets less accurate with more than about 20 options, according to its own benchmarks.
Where Ned does his homework
The tools are free. Knowing what to ask them for is the real skill. Professor Glitch teaches it from zero, in order, inside the community: 8 courses and 200+ lessons. These two are where Ned started.
