Best of your X follows: Fable prose, MTurk, and model workspaces

Fable showed up twice, but not as another launch recap. Paul Graham used it as a yardstick for future model progress; Ethan Mollick treated its wordiness as a product bug. The rest of the day was about the working surface around AI: human-subject pools, interpretability tools, and the stock phrases models keep leaking into prose.

Qualified window: five original posts from followed AI/tech accounts in the past 24 hours. Pure retweets, small talk, and context-light quote reactions were left out.

Ethan Mollick called Fable's "overwrought language" an ongoing problem in software and design projects, especially in small UI surfaces like toasts, menus, and bits of product copy. 1 The warning is practical: a model can be excellent at the hard task while still shipping copy that feels subtly fake. At capture, the post had 106 likes, 15 replies, and 7 bookmarks, which is modest engagement but a clean builder-facing signal. 1

Mollick also listed phrases he sees AI overusing: "earned," "carries," "sitting with" an idea, inanimate concepts that "live," and pseudo-profound paragraph endings. 2 For anyone using AI to draft customer-facing text, this is a lint list, not a style complaint: these phrases are now easy tells. The post was smaller than the Fable-language note, with 159 likes, 11 replies, and zero retweets at capture. 2

Mollick said Amazon Mechanical Turk is "on its way out," after serving as a mainstay for social and survey research through the 2010s. 3 The issue is not that MTurk stopped selling access to people; it is that LLM use has made that pool harder to treat as reliably human. At capture, the post had 239 likes, 19 replies, 15 reposts, and 65 bookmarks, suggesting more interest from researchers than from the broader AI hype cycle. 3

Mollick pointed readers to Neuronpedia's Qwen3.6-27B Jacobian Lens page and singled out the visualization at the end as something to try. 4 The page frames Jacobian Lens as a way to reveal a "global workspace" in language models, with paper, blog, code, and weights linked from the tool page. 5 This is the day's best research-tool pointer: less announcement, more hands-on model inspection.

Paul Graham asked readers to imagine models improving over the next five years as much as Fable has improved on GPT-3. 6 The line is short, but the frame is useful: compare capability jumps by lived workflow change, not only benchmark deltas. At capture, it was the strongest qualifying post by reach, with 4,427 likes, 279 replies, 181 reposts, 384 bookmarks, and about 642,000 views. 6

If you only have five minutes, open Mollick's Fable-language post and the Neuronpedia Jacobian Lens page. The first is immediately actionable if AI writes any part of your product; the second is the one external tool link in today's set that can change how you look at a model rather than just how you talk about one.