OpenJev (openjev.com)

Article summary

The article presents OpenJev, a live experiment that allows users to compare the performance of a local decision model in their browser, using either direct readout or generation of option probabilities. The model can be run on different sizes and devices, with options to switch between models like MiniCPM5 2B and Qwen3 0.6B. The experiment aims to measure the difference in performance between the two approaches. The article also mentions that the model's weights come from Hugging Face and remain in the browser cache, with no inputs leaving the page.

Main themes

• OpenJev experiment
• decision model performance
• LLM-generated websites
• encoder-only models
• novelty and uniqueness
• user experience and design
• model biases and limitations
• replicability and comparison with other models

What commenters say

• The approach used in OpenJev is not novel and has been done before with other models.
• The naming of OpenJev is misleading and could be legally problematic.
• The speed and performance of OpenJev are impressive, but may not be groundbreaking for all use cases.
• The model's ability to classify input text into categories with confidence scores is a key feature.
• The use of encoder-only models is not new, but OpenJev's implementation may have fewer restrictions or be smarter.
• The design of LLM-generated websites, including OpenJev, can be off-putting and prioritize function over user experience.
• The experiment's results may not always align with expected outcomes, highlighting potential limitations or biases in the model.
• The performance of OpenJev can be replicated with other models and bespoke inference engines, potentially undermining its uniqueness.