How to use Jev to automate your business (Step-by-step w/ Treg)
In a Nutshell
Jeff is a new model that outputs probability scores for predefined options rather than generating text, delivering 5-7x faster and 5x cheaper performance than standard LLMs while enabling near-100% accuracy business automation. It excels at high-volume decision tasks like real-time fraud detection, browser automation, SEO link mapping, and lead qualification by pairing with traditional LLMs for text generation. The key innovation is using confidence scores to build autonomous business logic that only escalates uncertain cases to humans, making previously uneconomical AI workflows viable.
These notes were generated by AI and may contain inaccuracies.
A new model called Jeff has been appearing on Twitter. The core problem is that while the latest models excel at advanced tasks like coding and mathematics, they still struggle with basic business workflows and automations. Large models can handle Olympic-level problems but companies rarely trust them to handle customer inquiries involving billings autonomously. Two main reasons explain this limitation: overconfidence and cost/speed constraints.
Agents frequently state "you're absolutely right" when they are actually wrong. Models make judgment calls without calibrating their actual confidence level. Most business workflows require almost 100% accuracy, which is why humans are still preferred. Additionally, the cost and speed need to make sense because many business workflows involve massive scale and volume.
Large models go through reinforced learning from human feedback, where they are fine-tuned based on human feedback in chat assistants. Jeff uses reinforced learning for calibrated decisions instead. This model is designed for reliable, high-quality decision making rather than being creative or assisting humans. It makes decisions given a set of options but cannot make up options itself, cannot output text, cannot write code, and cannot reason step by step.
Unlike normal large language models that output results token by token, Jeff does not predict text but outputs probability of a list of given answers in an extremely fast and efficient approach. While this might feel limited at a high level, it is significantly faster and cheaper than even the cheapest large models available.
Jeff is consistently 5 to 7 times faster and 5 times cheaper compared to the cheapest intelligent models like GPT 5.6 Luna. One limitation is that Jeff has a context window of 32K. If responses exceed this limit, users must implement workarounds like map reduce. When Jeff outputs responses, it provides the probability distribution of what it thinks is the right answer.
The confidence score enables building business logic guards around every answer, allowing models to operate fully autonomously. This opens use cases that weren't possible before. One example is playing games, where real-time fast decision making is required. Because games have limited sets of inputs and options, and Jeff processes so quickly, it can play games fully autonomously.
A demo showed a system that can live monitor any call and give real-time classification for fraud, issues, or frustrations. This demonstrates real-time intelligence capabilities. Jeff also excels at browser and computer use because the hardest part for agents using browsers is deciding which UI elements to click or type, as DOM structures are often messy.
Large language models either make mistakes or consume context windows rapidly when using browsers. Browser use has already integrated Jeff into their stack and achieved much faster browser and computer use. The process involves sending Jeff the task to achieve, browser interaction history, and a list of DOM elements. Jeff predicts which type of action to perform next (click or type) and which UI element to interact with, outputting probability scores for each answer.
Jeff pairs with a small large model to handle actual text generation for interactions like typing. This demonstrates how Jeff plus large language models can complete tasks end-to-end. Jeff is designed for business automation, especially for tasks requiring high accuracy or processing massive amounts of data that weren't economically viable with expensive large models.
For internal link mapping in SEO and AEO, models need to crawl through hundreds of pages and accurately locate links between pages. With normal models like Claude 3.5, this can take hours. Jeff scanned more than 500 pages and completed mapping in less than 50 seconds. Creative uses include real-time customization of copy based on visitor personas.
Calling Jeff is similar to using structured output from large language models. Users pass a prompt (called a state) and a set of questions. For support ticket handling, questions could include which department should handle a request. Instructions guide how Jeff should make decisions, and a list of options is provided with further descriptions for each option.
Jeff supports three types of questions: multiple choice (as demonstrated in support ticket examples), true and false questions, and score questions. True and false questions can prevent jailbreaks or prompt injection in AI chat platforms by asking if user requests are jailbreaks and providing probability scores.
Score questions can evaluate whether Twitter posts are organic or paid advertisements based on engagement statistics. Users send post details as state and ask Jeff to provide an authenticity score with different criteria ranging from "clearly manipulated" to "clearly organic." Each option added increases the index by one, and Jeff outputs probability for each option to calculate a final score.
The confidence score allows building sophisticated business logic decision trees. For prompt injection cases, logic can include: if hazard score exceeds 70%, automatically block the message; if jailbreak score is between 35% and 70%, request human review; otherwise, send as normal.
For more accurate and aligned responses, detailed descriptions work better than simple ones. Within criteria options, users can include examples similar to few-shot prompting. In customer support scenarios, simple instructions might route double charge issues to billing, but additional instructions focusing on root cause resolution might route to technical teams instead.
Two products experience heavy bot attacks: Super Design (vibe design platform) and Track (open router for data and tooling connecting to thousands of API endpoints). Handling fraud is difficult because attackers frequently change domains, making programmatic checks ineffective. Every 15 to 30 minutes, signups and product usage data are fetched and classified for fraud likelihood, with automatic banning above certain thresholds.
Track provides access to more than 3,000 different data points including people and company enrichment, buying signals, and other services. Instead of charging subscription fees, Track charges per usage with 0% markup. For the same people enrichment task, Track achieves similar or better accuracy at 85% cheaper cost - enriching about 290 people costs $1.49 versus $10.30 with Clay.
Every website signup goes through email verification and person enrichment via Track, then grouped information including product usage data feeds into Jeff for classification (fraud, upsell value, or affiliate/influencer partners). The extremely low cost of both Jeff and Track enables scanning hundreds of signups for around $1 per day, triggering different outreach, in-app support, or fraud banning.
Track fetches popular LinkedIn posts relevant to specific verticals like enrichment, then retrieves people who interact or comment on those posts as potential buyers. This data feeds into Jeff for qualification and scoring, classifying leads into different personas and roles. For analyzing 20 posts and nearly 400 leads, costs are minimal. Top leads proceed through proper enrichment pipelines for contact information.
Other potential use cases include identifying companies hiring on LinkedIn or making job posts, then matching with candidate databases, and job trend analysis. Track provides endpoints for collecting these signals. Getting Claude or Codex to write scripts using Track to fetch buying signals is straightforward.
A pipeline grabs viral posts from the past 24 hours, uses Jeff to filter product launch-related content, and classifies which posts appear organic versus paid or boosted traffic. Track fetches trending posts with statistics and comments, then Jeff classifies organic-looking posts versus paid content while filtering noise. This provides a daily feed of genuine trending content.
A free tool at track.j allows pasting any Twitter post to trigger the pipeline, fetching post information, author profiles, and replies, then using Jeff to classify organic versus paid content with full reasoning and analysis.
Keep AI Jason in your library
Save the videos and channels worth coming back to, and find them again in one place.
