Jev Explained: Demos and Use Cases

Jev, a new AI model from Types AI, co-invented by Dio Almata (known for ChatGPT and RHF), is remarkably fast and cheap. It is 20 to 200 times faster and 40 to 400 times cheaper than traditional Large Language Models (LLMs). Crucially, Jev is not an LLM; it's a classifier designed for decision engines, fitting into what's known as "system one" thinking—fast, automatic, and immediate, as opposed to LLMs which are "system two" (slow, effortful). To use Jev, you provide it with data (known as "state"), such as a customer email or code, and fixed-option questions. Jev then responds with a probability for each option. This allows for pairing Jev's output with deterministic code that branches based on its classifications, leading to specific actions. Jev supports three types of questions: 1. **Nules (Yes/No questions):** For example, "Is this a refund request?" 2. **Choice (Pick one from options):** For example, "Which team should respond to this email?" 3. **Score (Rate on a scale):** For example, "How frustrated is the user?" The integration with Jev is straightforward, similar to other AI APIs using an SDK. You install Typesafe AI, provide your state (e.g., an email), and specify questions using `choice`, `null`, or `score` imported from the SDK. The output is a strongly typed object, enabling immediate code branching based on Jev's classifications. This approach differs significantly from current agent development, which often relies on large, error-prone LLM prompts. Traditional agents struggle with consistent structured output, often requiring `try-catch` blocks due to LLM hallucinations or truncated responses. Jev, however, consistently provides specific, fast answers, making it ideal for deterministic code. Jev acts as the initial classification layer in agent workflows. An incoming message is quickly classified by Jev, and then your code takes action. For instance, an order status request can trigger an immediate tool call without an LLM. Complex product questions might then be routed to an LLM, while angry or unclear messages could be triaged to a human. Malicious messages can be instantly blocked. Jev enables writing more deterministic code, reducing reliance on LLMs' unpredictable outputs. Examples of Jev's applications include: * **Browser automation:** A user booked flights in 7 seconds by hooking Jev to browser snapshots to fill in information, a task typically slow with LLM-based agents. * **Data classification:** Classifying over a thousand AI research papers cost 8 cents and took 256 milliseconds per classification, highlighting Jev's speed and cost-effectiveness for tasks previously done by LLMs. * **Resume screening:** Jev can classify if a resume is a good fit for a job posting, proving 10 times cheaper than smaller LLMs for this task. * **Email management:** Jev can classify personal emails for priority, spam, or reply necessity in real-time. * **Real-time text analysis:** Jev can analyze text as it's being written, extracting tone, conviction, and urgency, useful for word processors or document creation. It can also detect AI-written content based on deterministic rules. * **Browser extensions:** An X (formerly Twitter) extension uses Jev to hide rage bait, crypto, or political posts in real-time. A similar YouTube extension could filter recommendations. * **Debate analysis:** Jev can fact-check spoken words in real-time during a debate, providing a "BS meter." * **Gaming AI:** Jev can play Tetris by deciding moves (press down, rotate) in real-time via API calls. * **Self-driving cars:** Jev can make real-time decisions in a driving simulator (accelerate, stop, turn) based on inputs like stop signs, traffic, and pedestrians. * **Code analysis:** Jev can perform code reviews, scoring modified files for security issues or bad coding practices based on defined criteria. Practical applications at Syntax and Sentry demonstrate Jev's utility: * **Podcast summary verification:** Syntax uses LLMs to summarize podcast transcripts. Jev can then verify claims made by the LLM, ensuring truthfulness and preventing hallucinations. * **Model routing:** Sentry's Slackbot, Junior, uses a "turn router" to decide which model to use. This can be replaced by Jev asking two questions: "How much reasoning is required?" and "Which profile does it fit?" to classify the message before an LLM is engaged. A chatbot demo built without any LLM, using only tool calls with Jev, showcases its capabilities. The chatbot can answer questions like "What's the weather in Denver tomorrow?" by classifying the user's intent, identifying the necessary tool (e.g., weather API), extracting parameters (city, time), and making deterministic tool calls. It handles conversational history for follow-up questions and can convert units or search Wikipedia/web, quoting specific sentences from retrieved articles. This approach grounds responses in factual tool calls, reducing hallucinations common with LLMs. Another demo hooked Jev to Home Assistant, allowing fast, real-time control of smart home devices (e.g., turning lights on/off, changing brightness/color) with plain language commands. Jev processes these commands and makes API calls within 300 milliseconds. Jev's speed stems from its ability to return all necessary information (tool to call, arguments) in a single response, eliminating multiple prompts. Jev is accessible via a waitlist or the Vercel AI gateway.