benchmarks

10 articles, newest first

Creative Coding

Shorter Output, Same Cost: Declarative Animation Specs Didn't Beat Free-Form Code

The expectation is that constraining an LLM to a declarative animation format is cheaper, faster and more reliable than letting it write HTML and JavaScript. A controlled comparison across three Claude models and Qwen3-14B — 600 jobs, pre-registered analysis plans — found the shorter answers did not translate into less total output or more speed.

Oct 26, 7:25 AM · 5 min read

Research & Innovation

Evaluating an Image Model Without Fooling Yourself

Oct 19, 12:45 PM · 6 min read

Research & Innovation

Research Roundup: Judging Generated Worlds, Faking Fluid Physics, and Noise You Can Control in Time

Oct 12, 1:45 PM · 4 min read

AI & Creative Tools

Odyssey-3 Generates an Interactive World From a Sentence and Lets You Walk Around in It, Free

Oct 11, 8:10 AM · 4 min read

Research & Innovation

LLM-Generated Interfaces Pass the Brief and Repeat Each Other, With Creative Range at 0.581 Against 0.902 for Humans

Oct 10, 7:50 AM · 4 min read

Research & Innovation

Benchmark Tutorial: How to Measure a Model on Your Own Work Instead of Trusting a Table

Oct 9, 10:52 AM · 7 min read

Research & Innovation

A Benchmark That Asks Agents to Write the Code That Would Produce a Video

Oct 5, 9:30 AM · 4 min read

Emerging Interfaces

Generated Pointing Can Be More Accurate Than Human Pointing and Still Look Wrong

Oct 4, 10:35 PM · 4 min read

Research & Innovation

How You Measure Text-to-3D Matters More Than Which Generator You Use

Oct 3, 8:20 AM · 4 min read

AI & Creative Tools

Two 27B Decision Models Landed on the Same Day, on the Same Base, and Now Argue Over the Same Benchmark

Oct 3, 7:40 AM · 5 min read