Caeliai

Independent research on AI search, taste, and why.

  • ArchiveAgent

    A small fine-tuned model beat much larger ones at reading runway photos.

    • ArchiveAgent56.4%
    • Kimi K344.9%
    • Claude Opus 5.528.6%
    • Before training12.2%
    Brand, season and year all right. 287 new photos, Oct 2026. “Before training” is Kimi K2.6, the model it started from. ResultsLimitsarchiveagent.si
  • Search audits

    When AI names a brand, almost half the time there’s no working link to it.

    brand’s page 36.8% · someone else 17.6% · no link 45.5% 380 brand links from ChatGPT and Gemini, May 2026. PaperField studyMethod
  • Home GPU test

    On one home graphics card, the best free image model found the right brand 3 times in 4.

    • SigLIP 274.4%
    • PE-Core-L67.9%
    • CLIP ViT-B/1652.2%
    • DINOv3-L48.7%
    • Random guess8.7%
    86,874 runway photos, 6,480 checks, one RTX 5060. Results
  • Shopping Answers

    ChatGPT showed a product carousel in 6 of every 10 shopping answers.

    carousel 59.7% · text with links 37.3% · other 3.0% 274 of 459 saved answers, 463 brands named. Saved Mar to Aug 2026. ExploreUse with an agent