{
  "id": 11414463,
  "title": "Claude-shaped science",
  "url": "https://urgent.news/2026/10/02/claude-shaped-science",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-02T10:07:15.000Z",
  "source": {
    "name": "Lobsters",
    "slug": "lobsters",
    "url": "https://www.anthropic.com/research/claude-shaped-science"
  },
  "original_language": "en",
  "account": "Claude-shaped science refers to a new approach to utilizing AI-accelerated science by leveraging the capabilities of large language models (LLMs) like Claude. In this approach, Claude is allowed to find \"Claude-shaped\" problems, which are the most suitable for the current generation of LLM tools. This led to the creation of BootLoops, a toolkit for exact calculations in quantitative science.\n\nBootLoops functions as a harness for the LLM, similar to Claude Code or Claude Science, which are harnesses for Claude and GPT, respectively. BootLoops has proven to be especially well-suited for a class of quantitative problems in science and has been open-sourced for use with any model.\n\nAs Claude worked with BootLoops, it started noticing patterns where many fields have problems that can be solved using techniques from mathematics, physics, or computer science. These problems are referred to as \"Claude-shaped\" problems. By collaborating with domain experts, BootLoops was able to make substantive advances in various research areas, including ecology, population genetics, geology, biology, economics, and linguistics.\n\nClaude's capabilities are impressive, allowing it to understand and process vast amounts of information, code, and analyze scientific papers at a rapid pace. However, working like a human scientist is not what LLMs are best at. There is often a disconnect between the impressive capabilities of LLMs and the needs of academic scientists. The core conflict lies in the fact that while these models are brilliant, they are not scientists and require significant hand-holding to produce scientifically valuable results.\n\nTo bridge this gap, the author proposes treating Claude as an AI collaborator that excels in specific areas, such as coding and quantitative problem-solving. By identifying problems that suit Claude's strengths, researchers can harness its potential in a more efficient and effective manner. BootLoops represents one step towards resolving the impedance mismatch between human and AI scientists, enabling the successful application of LLMs in various research fields.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Hacker News",
        "title": "Claude-Shaped Science",
        "url": "https://urgent.news/2026/10/02/claude-shaped-science",
        "published": "2026-10-02T13:38:03.000Z"
      }
    ]
  },
  "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."
}