AI ResearchMachine Learning2 min reading time

Qwen3.8 27B addition in words

Simon Willison's Weblog
Read the full article
The Qwen3.8-27B-Q4_K_M.gguf model was tested on adding positive integers and expressing sums in English words. Without reasoning, it achieved 23.57% numeric accuracy, dropping sharply for larger numbers, but with reasoning enabled, it correctly answered 167 of 169 cases.

More in AI Research

AI Research1 min read

Different eval awareness may track different graders

LessWrong
AI Research1 min read

Tracing Subliminal Learning Through Pretraining

LessWrong
AI Research13 min read

What Is Data Poisoning? How Attackers Corrupt AI Models and Datasets

Unite.AI