Scrap the LLMs. Scoring 4.76% on the brand new ARC-3 using pure code, a 2012 AMD CPU, and zero AI tokens.[P]
A programmer has demonstrated that a simple Python script, running on a decade-old AMD CPU, can achieve a 4.76% score on the new ARC-AGI-3 benchmark. This feat highlights the inefficiency of current large language models, which struggle with the benchmark's dynamic, instruction-less environments and often achieve zero scores. The script utilizes basic computer vision techniques like centroid detection to solve spatial puzzles, outperforming many AI models despite its minimal resource requirements and lack of AI tokens. AI
![Scrap the LLMs. Scoring 4.76% on the brand new ARC-3 using pure code, a 2012 AMD CPU, and zero AI tokens.[P]](https://preview.redd.it/xv62h96r6d5h1.png?width=140&height=78&auto=webp&s=6e69473f7fe7a24721ec6540b49c587570bee6d3)
IMPACT Demonstrates that traditional programming can outperform current LLMs on specific benchmarks, highlighting LLM inefficiency.