Humans and Machines Make Decisions Using Two Systems

In his bestseller “Thinking, Fast and Slow” (2011), Daniel Kahneman, a psychologist and pioneer of behavioral economics, distinguished between System 1 and System 2 in human thinking. System 1 is fast, automatic, and intuitive, but prone to errors; System 2 is slow, resource-intensive, and capable of calculating, verifying, and reasoning. And these two systems actually mirror the two-stage approach of AI: Decision Models – the AI’s “gut feeling” – make a quick preliminary decision, while a large LLM is only activated when uncertainty arises and a complex analysis of a problem is required.

Calibrated Decisions count

Decision models are often not entirely new architectures: Many are built on top of existing models, such as Quyet, which is based on Gemma, or Clef, which is based on Qwen. Jev from TypeSafe, which sparked the current hype, is, by contrast, a closed model. It is trained using Reinforcement Learning for Calibrated Decisions (RLCD). The system does not reward the most agreeable answer; instead, the training system provides the model with positive feedback for a correctly calibrated probability. The output is not continuous text, as is common with Claude, ChatGPT, and similar systems, but rather a probability, for example: “Spam with 0.96” or “Yes with 0.82.” If a model is uncertain, it indicates this with a value close to 0.5.

Three question types cover most use cases:

• Choice: Selection from several given options

• Scoring: Rating on a scale

• Boolean: Yes/no or true/false

Advantages of Decision Models Over LLMs

Decision models are extremely fast and cost-effective. In our test, Jev answered a single question in about 250 milliseconds. The cost for this is about 0.003 U.S. cents. Decision models can be used to classify large volumes of data in batches. Another advantage: the models do not invent new options. If the possible answers are “Yes, No, Maybe,” they choose exclusively from those – which reduces hallucinations and makes the systems more controllable for production environments.

Practical Use Cases for Decision Models

• Voice Agent: A decision model routes requests such as “Show me the weather” directly to the appropriate tool, at a median of about 266 milliseconds.

• Generative UI: Based on a prompt, Jev determines which UI components are needed and in what order they should appear; a language model then fills in the text and details. In our test, the selection and sequencing took just over 0.7 seconds.

• Browser Filter: An extension evaluates YouTube videos or Hacker News posts based on personal interests and filters out suspected clickbait.

• Coding Agent Routing: Decision models can determine the level of effort required for a prompt – such as low, medium, or high reasoning effort. Switching between different large models may result in the loss of the prompt cache.

• FFmpeg Helper: Instead of guessing through a confusing array of parameters, the model decides which operation is necessary, whether it is destructive, and whether another tool is a better fit.

• Game Autopilot: In our shooter game Sprint Doom, Jev made decisions regarding movement, weapons, targeting, and actions – at a median of 288 milliseconds and for about $0.33 per hour.

However, the model can also fail

The model only knows what you show it, and it even falls for clickbait. “CLAUDE CODE IS DEAD” matched the reader's interests with a score of 0.93; it wasn’t until a second question about clickbait that the model flagged the headline.

Conclusion: Listen to Your Gut and Your Head When It Comes to AI

Decision models are not a universal substitute for LLMs. Even powerful models can get math problems wrong: In a disguised math trap, GPT Luna chose the wrong answer, while Jev and Clef got it right. Such models tend to recognize patterns rather than perform thorough calculations. The key takeaway is this: For many everyday AI problems, a small, fast decision model is sufficient instead of a large LLM – especially for filtering, routing, classification, and tool selection.

Good AI systems combine a quick “gut” with a powerful “mind.” And a rule for who takes over and when.