Introducing 'Decision Models (Jev)' that output only the necessary answers for immediate action without long explanations, making them dozens of times faster than existing LLMs, along with 'Flide', a technology for operating them efficiently.
Imagine you are in a situation where you need to classify spam emails in real-time on your smartphone or capture the fleeting buy/sell timing in stock data. What if you asked an AI, “Explain to me in three paragraphs why this data is spam or why this is a sell signal”? If the few seconds it takes to wait for the answer pass by, you might miss the important timing.
Recently, in the artificial intelligence industry, a new type of model that makes ‘decisions’ right away instead of explaining in long text is gaining attention. These are ‘Decision Models’ like ‘Jev’.
Why It Matters
The Large Language Models (LLMs) like ChatGPT that we commonly use are very smart, but they take quite a while to generate answers. This is because they go through the process of considering sentence structure and adding polite explanations. However, in actual data processing fields, immediate signals such as “This is spam (Yes/No)” or “Importance is 85” are much more valuable than detailed explanations.
‘Jev-like’ decision models target this practical domain. Instead of a human reading and judging a long report written by an existing LLM, it is a method where the AI directly outputs a score and a category, and the software receives it and immediately takes action. This process can be handled more than ten times faster than existing LLMs, making it efficient Source: Introducing System One Models and Jev.
The Explainer
Let’s use an easy analogy. If an existing LLM is a ‘friendly chef’ who explains recipes in great detail, a decision model is a ‘high-speed sorter’ that classifies ingredients as soon as it sees them and puts them in a basket, saying “This is for stir-fry!”.
- System One Approach: Just as the human brain makes immediate decisions unconsciously, these models return immediate categories or scores without a complex logic generation process Source: Scaling JEV-like Decision Models with SGLang.
- Efficiency: Decision models do not waste computational resources because they do not write long texts. Therefore, they can be driven sufficiently in local environments like your computer or smartphone without relying on cloud servers Source: Jev-like decision models are flooding the AI market….
In fact, according to recent research, models running locally have shown response speeds nearly 20 times faster than general cloud-based models Source: Локальная модель Laya на 421 млн параметров превзошла….
Where We Stand
Currently, decision models including Jev are flooding the market Source: Jev-like decision models are flooding the AI market…. These models do not just get the right answer, but also calculate how confident they are in the decision they made (Confidence Level) Source: Introducing the Artificial Analysis Decision Index v2.
For example, when an AI determines whether something is spam, if the confidence level is low, it passes it to a human for confirmation, and if the confidence level is high, it works by blocking it immediately. Libraries like ‘Flide’, which was recently unveiled, boost the processing speed of these models by up to 9.5 times, breaking down the wall of real-time judgment Source: flide· PyPI.
What’s Next
In the future, not all AI will try to have chatbot conversations like humans. Instead, there will be more ‘invisible AIs’ that make optimal decisions and classify information within 0.01 seconds inside apps, even if they are not visible to the user’s eyes.
These models are lighter, faster, and run directly on my device, so there is less risk of personal information leakage. We are entering an era where AI goes beyond the age of transcribing our conversations to an era where AI makes judgments lightning fast on our behalf.
AI’s Perspective (MindTickleBytes’ AI Reporter’s Perspective)
If language models proved their ‘ability to speak’, decision models are proving the practicality of their ‘ability to judge’. In a digital world where speed is competitiveness, quick decisions rather than slow explanations will become the new standard for artificial intelligence.
References
- flide· PyPI
-
[Introducing the Artificial Analysis Decision Index v2 Artificial Analysis](https://artificialanalysis.ai/articles/decision-index-v2) -
[Jev-like decision models are flooding the AI market… Tom’s Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/jev-like-decision-models-are-flooding-the-ai-market-after-researcher-claims-typesafe-copied-their-research-microsoft-openai-among-companies-releasing-lightweight-system-one-ai-models) - bestjev: 896 Open Source Jev Projects & Tools
- Scaling JEV-like Decision Models with SGLang - LMSYS Org
-
[LitJev: Benchmarks & research with Jev Calibrated Decisions](https://calibrateddecisions.com/projects/litjev) - Introducing System One Models and Jev - TypeSafe AI Blog
- Scaling JEV-like Decision Models with SGLang - LMSYS Org
- Локальная модель Laya на 421 млн параметров превзошла…
- They write very long reports
- They immediately return a Category or Score
- They always require human verification
- Because they only run on the latest computers
- Because they can process as quickly and efficiently as human intuition
- Because they are the first AI developed
- Because it adjusts the color of AI
- Because it makes decision models run much faster
- Because it makes AI answers longer