How Does Jev Classify Text Without Writing?

TL;DR

Jev reads complex text and selects only from predefined answers, making it a general-purpose classifier rather than a writing model. It fits workflows that need interpretation followed by a category, score, or approval decision, and its constrained output can provide major speed and cost advantages over using a text-generating LLM for every judgment.

Transcript

Jeb is like an LLM that can only talk in multiple choice. It can read whatever you hand it, but it can't write even one word back. So, you supply all the possible answers in advance and it's going to only pick one. And I know what that sounds like to you. It sounds useless. Why would anybody want a model that cannot write in the age when we have LL... Read More

Key Insights

• Jev is a general-purpose classifier that reads supplied material and selects only from answers defined in advance. It cannot generate even one word outside those choices, so its limitation is also the feature that makes its output predictable and easy for programs to consume.
• A Jev-shaped problem contains complicated text as input and a small set of possible outputs, such as a category, score, approval, rejection, or escalation. The task still requires interpretation and judgment, even though the final answer is structurally simple.
• Deterministic rules are appropriate when a condition can be stated precisely, such as flagging an invoice after 30 overdue days. They are less suitable for interpreting whether an email signals customer dissatisfaction, contractual risk, or a serious business opportunity.
• Semideterministic software combines probabilistic language judgment with predefined outcomes and actions. The model determines which choice best fits the input, while developers or agents control the available choices and specify what the system does after each possible result.
• Traditional machine learning classifiers require examples, labeled data, training or fine-tuning, evaluation, and maintenance as the underlying problem changes. That investment can make sense at large scale, but it is difficult to justify for many small judgments scattered throughout software.
• General-purpose LLMs already classify messages, choose tools, judge relevance, screen documents, and check other models' work. Jev targets these same decision points with an architecture focused on structured choices rather than paying for unrestricted text generation that the workflow does not need.
• Agent safety is a potential Jev use case because the classifier can review complicated proposed actions, text, or code and choose whether an agent may proceed, must stop, or should ask a human. Its speed and cost could support checks across more of an agent run.
• Jev should be evaluated against the system it might replace, including an LLM, an existing machine learning classifier, or individual human decisions. The strongest candidates are workflows with complex language input and constrained output, but the presenter emphasizes testing rather than assuming universal superiority.
• More videos from this channel:
  • What Happens When AI Picks the Problem to Automate?
  • How to choose single vs multi-agent AI architectures?
  • What AI Super Week signals for enterprise strategy
• More from this channel: AI News & Strategy Daily | Nate B Jones

Questions & Answers

Q: What is Jev and how does it classify text?

Jev is presented as a general-purpose AI classifier that can read complicated supplied text but cannot write an unrestricted response. Users or agents define every permitted answer in advance, and Jev selects among those choices while returning associated scores and probabilities to the program. Several questions can also be evaluated together, allowing one input to support multiple structured decisions.

Q: What kinds of problems are best suited to Jev?

Jev is best suited to problems with complicated text as input and a small, clearly defined set of outputs. Suitable outputs include a category, a score, a yes-or-no decision, permission to proceed, or a request for human review. The central requirement is that understanding the input demands interpretation, while the useful result does not require an open-ended written response.

Q: How is Jev different from deterministic software rules?

Deterministic rules produce the same outcome whenever a precisely stated condition is met. For example, ordinary code can flag an invoice when it is 30 days overdue or flag an order whose total exceeds a limit. Jev addresses less precise questions, such as whether an email suggests a customer may leave, where language must be interpreted before selecting a predefined outcome.

Q: What does semideterministic software mean?

Semideterministic software is the presenter's term for connecting probabilistic language judgment to a controlled set of choices and actions. Jev still uses a transformer architecture to interpret complex input, so its judgment is not a handwritten rule. However, developers or agents constrain the possible outputs in advance and determine exactly what should happen after each selected result.

Q: How can Jev support safer AI agents?

Jev can act as a decision checkpoint before an agent executes a proposed action. It can inspect complicated agent text and code, then choose among outcomes such as permission to proceed, rejection, or escalation to a human. Because the classifier is described as fast and cheap, these checks could be placed across a wider portion of an agent run than more expensive reviews.

Q: How does Jev compare with traditional machine learning classifiers?

Traditional machine learning classifiers generally require teams to collect examples, label data, train or fine-tune a model, evaluate performance, and maintain the system as conditions change. Each new classification category may require additional specialized work. Jev is described as general purpose because it can address different classification problems from written descriptions and predefined choices without training a new custom classifier each time.

Q: What real software workflows could use Jev?

The cited workflows include classifying inbox messages, routing support tickets to the appropriate team, identifying customer risk, detecting serious business opportunities, selecting tools, judging relevance, screening documents, checking another model's work, and reviewing proposed agent commands. Each example requires understanding nuanced material but ultimately produces a limited decision that another part of the software can act upon.

Q: How should developers evaluate whether to use Jev?

Developers should first locate workflows where complicated text enters the system but only a simple choice comes out. They should then test Jev against the actual alternative, such as an LLM call, an existing machine learning classifier, or individual human decisions. The comparison should examine whether Jev performs the required judgment while delivering the claimed advantages in speed and cost.

Summary & Key Takeaways

• Jev addresses tasks where software must interpret complicated language but only return a simple decision. Developers define the permitted outcomes and their downstream actions in advance. The model then evaluates the input and selects among categories, scores, or yes-or-no options instead of generating unrestricted text, creating what the presenter calls semideterministic software.

• Traditional deterministic rules work when conditions can be written precisely, while custom machine learning classifiers require labeled data, specialized expertise, evaluation, and ongoing maintenance. General-purpose LLMs reduced those barriers, but developers often use costly text-generation systems for tasks that only require classification. Jev is designed specifically to fill that architectural gap.

• Potential applications include routing support tickets, classifying inbox messages, detecting customer churn signals, identifying genuine business opportunities, selecting tools, screening documents, checking model output, and reviewing proposed agent commands. The presenter recommends testing Jev wherever complex text produces a limited choice, then comparing its quality, speed, and cost against existing alternatives.