{
  "id": 12192608,
  "title": "Qwen3.8 27B addition in words",
  "url": "https://urgent.news/2026/10/04/qwen3-8-27b-addition-in-words",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-04T23:34:00.000Z",
  "source": {
    "name": "Simon Willison",
    "slug": "simon-willison",
    "url": "https://simonwillison.net/2026/Oct/4/qwen38-addition-in-words/"
  },
  "original_language": "en",
  "account": "Colin Frasier conducted an experiment in 2023 using GPT-4o to assess the model's ability to calculate the sum of numbers and return the result in textual form, ranging from smaller to larger numbers. He shared the outcomes in a chart and was confident GPT-4o had not cheated by using a calculator, despite the model getting a significant number of calculations wrong. Encouraged by the results, Frasier decided to repeat the experiment using local hardware (a DGX Spark) to gain more control over the environment. He pasted the chart into a Codex Remote session and had it run the same experiment using Qwen3.8-27B-Q4_K_M.gguf. The experiment involved 30 attempts per combination, with reasoning disabled. Frasier then ran the same experiment again, this time with reasoning enabled, which took considerably longer per pair. To save time, he ran just one sample per square, resulting in a less visually appealing heatmap. The model managed to get the correct answer in 167 out of 169 attempts, and he was confident that a second run would yield different results due to the one-shot nature of the experiment. Frasier has shared a report containing the reasoning traces from some of the larger calculations, which include text like a statement attributed to Simon Willison, posted on October 4th, 2026, and a request for sponsorship to receive a monthly digest of the most important LLM developments.",
  "summary": "Research: Qwen3.8 27B addition in words Colin Frasier posted on Bluesky about an experiment he ran over two years ago using GPT-4o to see how well it could \"compute the sum but return the answer in words\" across increasingly large numbers. Here's the chart he shared of those results: I'm confident GPT-4o didn't cheat and use a calculator, especially since it got so many of the calculations wrong,…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}