The video explores the Viral Decision Model called Jev, introduced by Diogo Almeida, co-inventor of ChatGPT and CEO of Type Safe AI. Jev is a novel AI model designed for decision-making rather than text generation, boasting claims of being 20 to 200 times faster and 40 to 400 times cheaper than traditional large language models (LLMs). Unlike LLMs, Jev does not generate free-form text but instead answers specific, typed questions with calibrated probabilities, making it highly efficient for classification and decision tasks. The model operates by evaluating multiple options in parallel from a single read of the input state, returning probabilities for each choice, which enables rapid and cost-effective decision-making.
The creator of the video developed a mini reproduction of Jev, dubbed “mini Jev,” using a smaller pre-trained language model (Qwen 3 with 0.6 billion parameters) that was modified to mimic Jev’s decision-making process. This mini Jev was trained specifically to play a custom Sonic-style game by predicting actions that minimize damage. Unlike traditional chat models that generate text token-by-token, mini Jev processes the input once and scores predefined options without generating text, making it faster and more efficient for this task. The training involved running thousands of game simulations to calibrate the model’s probability estimates for different actions, enabling it to make informed decisions during gameplay.
The video compares the performance of the mini Jev with the actual Jev model on the custom game. The real Jev, when used zero-shot without game rules, performed poorly, often making suboptimal moves. However, when provided with explicit game rules, it improved by avoiding damage but became overly cautious, frequently choosing to wait rather than take risks. In contrast, the mini Jev, trained specifically on the game, outperformed both versions of Jev by completing the course faster and achieving higher scores, demonstrating the advantage of task-specific training despite its smaller scale and simpler architecture.
Beyond the game, the video also tests Jev and mini Jev on a Wiki racing task, where the goal is to navigate from one Wikipedia page to another using only linked articles. The real Jev excelled at this task, efficiently finding short paths between pages with quick decision times and low costs, showcasing its strength in structured decision-making scenarios. Conversely, the mini Jev struggled with this broader and more complex task due to its limited training and capacity, often getting lost in irrelevant links. Attempts to fine-tune mini Jev on a small dataset of Wikipedia links improved performance slightly but were insufficient to match the real Jev’s capabilities.
In conclusion, the video highlights Jev’s potential as a fast, cheap, and specialized decision-making model that excels in tasks suited to its design, such as classification and structured navigation. The mini Jev reproduction serves as a proof of concept, illustrating the core principles behind Jev but also its limitations when applied to more complex or generalized tasks. The creator suggests that while Jev is not a replacement for LLMs, it offers exciting possibilities for specific applications, especially where speed and cost are critical. The video ends with an invitation for viewers to share their thoughts and a hint at possible future explorations of Jev’s capabilities.
Useful Links
- TypeSafe AI — Introducing System One Models & Jev — Directly explains the main subject Jev and its design principles and claims.