Jev AI: Fast, Cheap, Type‑Safe Classifier Redefining LLMs
Last week, a new AI model named Jev was released by Typesafe AI, a company founded by Diego Almeida, an ex-OpenAI researcher. Jev represents a significant shift in AI, particularly because it addresses a common flaw in large language models (LLMs): their verbosity and high cost. Unlike traditional LLMs that generate extensive text, Jev is designed to be fast, cheap, and precise, focusing on specific, structured outputs.
What is Jev?
Jev is a novel type of classifier that operates without language generation. It cannot engage in conversation, write code, or compose essays. Its core innovation lies in removing language from the large language model paradigm, resulting in a system that is:
- 200 times faster than traditional LLMs.
- 400 times cheaper, with free output tokens.
- Zero hallucinations, meaning it does not invent information.
This makes Jev particularly suitable for applications requiring quick, instinctual decisions.
How Jev Works: A Type-Safe Approach
Typesafe AI, the company behind Jev, provides a clue to its functionality. Similar to a type-safe programming language like TypeScript, Jev processes questions and context (unstructured text) but demands a strongly typed question that must return one of three specific shapes:
- Choice: Selecting from predefined options.
- Score: Assigning a numerical value.
- New: A simple yes or no (binary) response.
This schema matching is guaranteed, making type errors mathematically impossible.
Jev is categorized as a "system one" model, a concept derived from Daniel Kahneman's "Thinking, Fast and Slow." System one models are fast and rely on gut instinct, contrasting with "system two" models (like GPT-6 or Claude Fable) that are slow and deliberate, consuming many tokens for reasoning.
Practical Applications and Benefits
For app developers, Jev offers significant advantages:
- Cost-effectiveness: It is approximately 440 times cheaper than major brand LLMs.
- Speed: Its rapid processing allows for real-time applications.
Examples of its current and potential uses include:
- Content Moderation: Instantly banning non-compliant accounts on platforms (e.g., identifying non-horses on a horse-dating app).
- NPC Behavior in Video Games: Implementing dynamic character responses.
- Real-time AI Calculators: Performing calculations instantly.
Calibrated Confidence
While Jev's output is type-safe, it's not always deterministic or guaranteed to be correct. Like other LLMs, sending the same question and context can yield different results. To address this, Jev provides a "calibrated confidence number" with each response. This confidence is achieved through Reinforcement Learning for Calibrated Decisions (RLCD). For instance, a 60% confidence means the model is correct 60% of the time. This contrasts with traditional LLMs, which are often trained to please human evaluators and may express high confidence even when incorrect.
The Mystery of Jev's Architecture and Skepticism
The exact architecture of Jev remains proprietary, with the CEO indicating that a paper might be released in the future. This secrecy has led to some skepticism and comparisons to existing technologies.
- Zero-Shot Classifiers: Some critics argue that Jev is not fundamentally different from zero-shot classifiers that have existed for over a decade, with pioneers like Jiny Yang developing similar techniques.
- Open-Source Alternatives: A developer claims to have built an open-source equivalent of Jev, called OpenJev, a year ago. OpenJev reproduces Jev's interface by reading option probabilities from a frozen Llama 4B model in a single forward pass. It requires no new training and can run on a 3090 GPU, with a web GPU demo available for browser use.
MX: A Solution for Video Infrastructure
The video also highlights MX, a sponsor offering a highly customizable API for integrating video features into applications. MX simplifies video hosting and streaming, providing features like:
- Automatic Transcripts, Storyboards, Thumbnails, and Clips: Generated upon video upload.
- Structured Data: Information about video content.
- MX Robots: AI-powered workflows for tasks like audio translation, content moderation, and more, without requiring users to host models or maintain pipelines.
- Automated Workflows: Directives allow users to define workflows once, which then run on every new upload, with payment only for executed jobs.
Companies like Perplexity and Patreon utilize MX. A free plan is available, including 10 videos and 100,000 delivery minutes per month, with an additional $50 credit offered through a specific link.
Takeaways
- Jev is a new AI model from Typesafe AI that functions as a classifier without generating language, delivering only structured outputs such as choices, scores, or binary answers.
- By eliminating text generation, Jev claims to be up to 200 times faster and 400‑plus times cheaper than conventional large language models, with free output tokens and no hallucinations.
- The model enforces a type‑safe schema, guaranteeing that every response matches one of three predefined shapes, making type errors mathematically impossible.
- Jev provides a calibrated confidence score for each answer, derived from Reinforcement Learning for Calibrated Decisions, so users can gauge the likelihood of correctness.
- Although its architecture is proprietary, critics note similarities to zero‑shot classifiers, while an open‑source project called OpenJev replicates its interface using a frozen Llama 4B model.
Frequently Asked Questions
How does Jev's type‑safe schema prevent hallucinations and ensure correct output shapes?
Jev removes language generation and forces every query to request one of three predefined response types—Choice, Score, or New—so the model can only produce data that fits the schema, eliminating the possibility of fabricating irrelevant text and thus preventing hallucinations.
What does the calibrated confidence number indicate in Jev's responses?
The calibrated confidence number, produced via Reinforcement Learning for Calibrated Decisions, represents the empirical probability that the given answer is correct; for example, a 60 % confidence means the model is correct about sixty percent of the time, offering a more truthful measure than the often over‑confident scores of conventional LLMs.
Who is Fireship on YouTube?
Fireship is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
that must return one of three specific shapes: * **Choice:** Selecting from predefined options. * **Score:** Assigning
numerical value. * New: A simple yes or no (binary) response.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.