What becomes worth automating
4 min read
Early LLM experiments applied models to work we already understood: summarizing a document, rewriting a paragraph, answering a question.
Products that would not have made sense without the model came later.
Jev may be at that earlier point for judgment. Decisions that happen too often to justify a larger model call may now be cheap and fast enough to sit inside the product loop.
My first experiment was boring. I used Jev to sort the podcasts and articles I had saved for later into high and low signal. It did the job.
What stands out in my use of Jev is how quickly it makes these small decisions. That makes me want to go beyond a single sorting pass. Something low signal for general reading might become useful when I'm investigating a particular company or workflow.
The queue made me look for decisions we usually postpone until the end. A bounded question could run on every item, every partial input, or every meaningful change in state.
A different product loop
TypeSafe introduced Jev in early access as a model for fast decisions inside software. A developer supplies the current state and questions. Jev returns choices, scores, or yes-or-no probabilities, without a generated explanation. A choice stays within the options the application supplies; that does not guarantee a correct judgment.
TypeSafe lists the input price at $0.042 per million tokens, with no output charge. That is a vendor price, not the cost of a useful decision. Preparing context, checking uncertain answers, and correcting mistakes still cost money.
Products often call a larger model after the user finishes a request or a workflow reaches a checkpoint. At Jev's listed price, a builder can consider asking bounded questions while the state is still changing. Judgment moves inside the task, where it can change what happens next.
Wait for the complete task
Ask as the state changes
The second loop can change course before the task is finished. The product still defines the questions and allowed answers.
A router inside an agent could choose which model gets the next task. During computer use, Jev could select the next approved action from the supplied state. A news queue could decide what deserves attention and what to leave aside as each item arrives.
Each follows the same loop. New state goes in, Jev answers a bounded question, and software chooses the next step before the task finishes.
Waiting for each choice adds up when a task needs many of them. A news queue may tolerate a delay that makes computer use frustrating. Fast answers would still need to select useful actions; price and speed alone do not establish that.
Try the decision before the task finishes
I'd keep the saved items fixed and give the queue a concrete research question: how does a company handle customer support? Then I'd narrow the question to what happens when the first reply fails to resolve the customer's request.
The queue would reconsider the same material as the question changes. An item could move into the reading list because it addresses follow-up work, while a general account of the company's support operation could move out. If none of the saved material answers the question, an empty list would be more useful than a confident recommendation.
I'd inspect those changes against the sources. Did a newly included article help answer the narrower question? Did an excluded item contain something I needed? I'd also look at what I read next, because producing plausible labels would say little if I still had to search and sort the queue myself.
The comparison would be with the original high-and-low-signal pass. Repeated decisions would need to bring useful material into view when I need it, with fewer irrelevant recommendations than the broad classification. I'd count useful material missed, time spent correcting the list, and the full cost of reconsidering it as the question changes.
Software would keep the sources available, handle uncertain results and let me change the criteria. If the list keeps changing without improving the reading, I'd narrow or stop the experiment. If it helps me pursue the changing question without repeatedly sorting everything myself, I'd have a reason to keep judgment running during the research.