SlopBench ranks 18 models on AI 'slop'; Kimi cleanest, Mistral worst

Model choice measurably changes how 'AI-sounding' your writing reads — while MCP reaches Bloomberg's data and Anthropic interviews 80k people about AI.

Nowline SEP 30 1:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • SlopBench: a leaderboard for the least sloppy writing model

    A new arXiv benchmark scored 18 models across 112 writing tasks (~20k outputs): Kimi K2.6 came out least sloppy (21.1), Mistral Large the most (40.6). Handy for picking a model for user-facing copy — but the authors warn the ranking swings under reweighting and AI-slop detectors are unreliable, so read it as directional, not gospel.

  • MCP reaches Bloomberg's data — the protocol keeps eating enterprise

    Bloomberg shipped an Enterprise MCP server (Sept 29) that exposes Data License Plus to agents: discovery across 100M+ securities and 50k+ fields, with semantic context attached, not just raw values. It follows Rubrik's governed MCP for security data — MCP is fast becoming the default way agents reach premium data, so standardizing your agent stack on it keeps getting safer (Bloomberg's is gated to its data customers).

  • Anthropic's Interviewer quizzed ~80k people on what they want from AI

    Anthropic opened its Interviewer — an LLM that runs structured, open-ended interviews at scale — and published findings from ~80,000 participants; reportedly the top concern wasn't job loss. The pattern (conversational interviews over static forms) is worth stealing for your own user research, and the writeup is a free read on what people actually want built.