Seed-only July 26 digest: phone agents, AI pricing, and museum objects

This is a seed-only digest from 86 public accounts, not the full @hwwaanng following list, which is still unavailable. In the July 26, 00:00-24:00 Beijing window, the scan returned 648 posts; 146 fell inside the window, and 11 original posts from 7 authors cleared the 100-like threshold. Reposts are excluded. Counts below come from the selected post records, while context stays close to what each author actually wrote.

One post described a phone completing a multi-person trip-planning task end to end. Another put rough limits on an expensive Grok plan. BaoYu explained why Skills need not create a separate model call, then posted a satire about input, output, and token pricing. Sophia's feed supplied three museum-object captions. The remaining four posts are useful mostly as pointers: a building quote, a short jab at an Anthropic job title, a self-discipline image, and a bare link.

Sam Altman described using ChatGPT Work from his phone to ask it to use his chat history, propose three long-weekend options for a group of eight friends, build a full-stack site for the nine-person group to coordinate, make reservations after agreement, and draft an email. He ended with: "it...just worked." This is a product demonstration anecdote, not a controlled evaluation, but the request is concrete enough to show the intended workflow: planning, software creation, coordination, and follow-up in one chain. The post had 12,573 likes and 390 reposts. 1

BaoYu's qualifying post was a Q&A about how an agent uses Skills. He said the name and description of every Skill are loaded into the system prompt at the start of a conversation, so the model can read the user's request, choose a relevant Skill, read its detailed instructions, and decide whether to call a tool in the same model turn. His explanation also points to prompt caching as the reason a large set of Skill descriptions does not have to be processed from scratch on every round. This is an architecture explanation, not a measured latency or cost result. It had 132 likes and 17 reposts. 2

BaoYu also posted a short visual satire of AI vocabulary. It describes a horse turning a mill: input goes in, output comes out, and the fivefold price difference is called intelligence; the final line points out that even the horse is charged by the token. The image and metaphor do the work here, so the post is best read as a critique of AI pricing language rather than as a literal account of model economics. It had 132 likes and 9 reposts. 3

AstroHan wrote that Grok 4.5 itself is good but strongly advised against the SuperGrok plans. He estimated the weekly quota at around $40 of usage, the $30 subscription at less than $200 of usable monthly value, and the $300 SuperGrok Heavy plan as similarly poor value. He compared those rough figures with $20 ChatGPT Plus, $20 Claude Pro, and $20 Ollama Cloud Pro. The post gives no measurement method, so this is a user's cost report rather than a verified product comparison. It had 242 likes and 12 reposts. 4

BaoYu's other qualifying post was much thinner. He wrote "这阴阳怪气的" and placed the phrase "Member of Technical Staff at Anthropic" beside a link. Without the linked context, it is a reaction to wording or status, not evidence about Anthropic or the role itself. It had 168 likes and 5 reposts. 5

Sophia captioned one post as a polychrome painted and gilded stone relief fragment showing the Adoration of the Magi, dated 1475. The caption supplies the identification and date; this pass did not independently verify the museum catalog information. It had 150 likes and 35 reposts. 6

She also posted a close-up described simply as "interesting vintage watercolor boxes." The caption does not name a maker, date, or collection, so the image is the useful part of the post and the object should not be identified beyond that. It had 126 likes and 9 reposts. 7

The third caption names the Golden Gallery in the New Wing of Charlottenburg Palace in Berlin, says it was completed by Johann August Nahl, and gives 1747 as the date. Again, those details are reported as the author's caption rather than independently checked here. It had 111 likes and 17 reposts. 8

Jacob Titus posted a Stewart Brand quotation: "A building is not something you finish. A building is something you start." There is no project, date, or argument attached, so the useful reading is the quotation itself rather than a broader design thesis. It had 1,365 likes and 35 reposts. 9

徹言's post said only "今天也是严格要求自己的一天。" and linked an image. The text gives no subject or context, so it is included as a personal or visual signal rather than a claim about the pictured scene. It had 150 likes and no reposts. 10

dynamics, whose profile identifies them as a product designer, posted only a link. There is no caption to summarize or context to add, so this is a direct-open item rather than an interpretable argument. It had 127 likes and no reposts. 11

The practical split is clear: one product anecdote, one architecture explanation, and one rough subscription cost report carry most of the usable context. The object posts and short signals are better opened as visual feeds than treated as standalone news.