Jev: The Decision Model That Filters Your Second Brain

0:00Cole Medin has been running Jev — TypeSafe's "decision model" — inside his AI second brain "for real" for a while, and this video distills the four use cases that actually stuck. The framing is deliberately high-level: not a tutorial, but patterns to steal. The through-line is that his second brain makes thousands of decisions a day, and most of them are filtering — deciding what external information deserves his attention — which is exactly what Jev is built for.

What Jev is

1:57Jev is not an LLM, and Medin is honest about the marketing: "They try to present Jev as this brand-new type of model. It's really not — there have been things like it before." What it does: it doesn't generate text. Its sole job is to make decisions, fast and cheap. You give it a state plus a list of multiple-choice questions, and it answers all of them in parallel, each with a confidence score. TypeSafe's framing is that it's a "System One" model (the fast, intuitive thinking in Kahneman's taxonomy), contrasted with the "System Two" deliberative reasoning of LLMs. The claimed numbers: 20–200× faster than an LLM depending on which one, and 40–1,000× cheaper. The practical pitch: a lot of places where you're currently using an LLM just to make a yes/no call can be swapped for Jev and get faster, cheaper, and sometimes more reliable.

The economics

0:58Medin routes Jev through OpenRouter's Decisions endpoint, and his dashboard makes the value concrete: in a single week of testing he made 38,000 calls to Jev, pushed 55.3 million tokens in, and the total bill was $21. For thousands of decisions a day that's the whole argument in one number — decision-model pricing is a different order of magnitude from LLM pricing.

Use case 1: prompt-injection guard

4:00Every piece of external information entering a second brain is a potential prompt-injection vector, so Medin used to run a separate read-only LLM session to screen each incoming email before the main agent ever saw it. Now Jev does it. The prompt is simple: append the email to a hard-coded state and ask a battery of yes/no questions — does it contain instructions to send or forward data, edit memory or settings, do boundary-escapes, or relay orders while impersonating the owner? 7:21The payoff of parallel questions: Jev doesn't just answer "is this an injection," it answers "what kind," returning a confidence score per question. A threshold — anything above ~0.7 — triggers quarantine. Running every check for every email took under a second and "a fraction of a fraction of a penny." Medin says he's been "super impressed" with Jev's injection detection, though it's only as good as the state and questions you craft.

Use case 2: "worth replying?"

4:55The second half of inbox management: not every email deserves the tokens to draft a reply. Jev labels each incoming email one of three ways — reply, skip, or injection — using a hard-coded list of things to definitely skip ("urgent demo sponsored video for $100" is his example) plus the email itself. 2:56The ones marked "reply" go to an LLM — he's experimenting with open-weight models like GLM 5.3 Flash rather than burning his Claude subscription — to draft the reply in his own tone. Notably, he never lets the system send the reply: drafting is automated, sending is not.

Use case 3: news filtering

8:08Same idea applied to the fire hose: RSS feeds and YouTube channels feed into his second brain, and Jev decides what's relevant. But here's the honest limitation that makes the video worth watching: 8:41Jev can't fully replace the LLM. It's great at the first cut — "is this just noise?" — but the decision "is this actually a good fit for my audience?" needs a lot of context about what he's currently covering and what he cares about. He found that even a small LLM (Claude Haiku, now GLM 5.3 Flash) beats Jev on that last, high-context relevance pass. So Jev handles the cheap bulk filtering and hands the borderline cases to a small LLM for the final say. 10:16He even ran a comparison on 14 news sources: Jev's composite relevance score (weighted across its parallel questions) does the triage, then defers to Haiku when it isn't sure the source fits his audience.

Use case 4: duplicate detection

11:33His second brain maintains a permanent knowledge base of entities he cares about (AI coding tools, MCP servers, companies like Anthropic and OpenAI). When new information arrives, he needs to know: is there something to add or update, or has he already processed this? That duplicate check used to be semantic-similarity matching, which works well — but he found Jev "more confident in a good way, and more accurate overall," and cheaper. 12:14The nuance: he biases toward re-processing (erring on the side of not missing something), so Jev isn't perfect, but it still stops a lot of wasted LLM calls.

The pattern

12:48Across all four, the same architecture emerges: Jev as the fast, cheap, parallel first-pass filter — the LLM as the high-context last mile. Decision models don't replace LLMs; they shrink the surface area where you need one. For a second brain that's mostly deciding what deserves attention, that's where most of the value lives.

Claims checked

Sources & further reading