Seed-only July 22 digest: local AI, world models, and old objects

This July 22 digest covers the 86 public seed accounts currently available for this channel, about 2.2% of @hwwaanng's roughly 3,850-account following list. The full following-list source is still unavailable, so these eight original posts should be read as a sample rather than a complete view of the network. The time window is July 22, 08:00 to July 23, 08:00 Beijing time.

Only original posts with at least 100 likes are included. Likes and repost counts below come from the post details; claims inside the posts remain the authors' own claims.

QT9277, an account whose profile centers on airdrops and crypto experimentation, recommended Ollama as an offline way to run open models. The post says Ollama supports Llama, Qwen, and DeepSeek on Windows, Mac, and Linux, with an OpenAI-compatible API, and presents keeping conversations on the local machine as its privacy advantage. It had 100 likes and 0 reposts when checked. 1

The same account followed with a post about a community-edited, token-reduced version of a reportedly leaked Claude Fable 5 system prompt. QT9277 says the repository turns it into a general template that can also be applied to Gemini 3.1 Pro and ChatGPT 5.6, and describes the package as open source and free. That is the poster's description, not an independent verification of the prompt's provenance or portability. The post had 104 likes and 2 reposts. 2

郭宇 (@turingou), whose profile says "Retired," posted a short congratulations message about humanoid robotics. He wrote that the "world model" problem had been discussed when he visited the Bay Area in 2024 and suggested that, after more than two years, the field may be approaching a breakthrough. The post does not name the company or project behind the message, so the reference cannot be pinned down further from the post alone. It had 154 likes and 12 reposts. 3

Sophia's profile describes her focus as antiquities, civilizations, and art. Her strongest post in this window identified a 1900 English brooch as a Smithsonian vintage purple glass heart and rhinestone brooch pin. The post was a one-line object label with an image link, not a longer provenance or curatorial note. It had 314 likes and 57 reposts. 4

She also posted two shorter object labels. The first named an Angolan chief's chair from 1800; it reached 121 likes and 17 reposts. 5

Later, she labeled a 1896 painting by Sergei Zhuravlev, Boyar’s Daughter, from Russia. It had 104 likes and 16 reposts. As with the other two posts, the digest keeps to the label because the post itself supplies no additional history. 6

The final qualifying post was a personal location note from citron (@vanillaCitron), who identifies as a former Tencent designer and iOS developer and lists PieceletNeoDB in the profile. In Chinese, citron wrote that they were in Hangzhou for a couple of days and asked whether anyone had recommendations or a place to meet. The wording is ambiguous, so it is best read as a social travel note rather than a specific recommendation. It had 105 likes and 3 reposts. 7

If you want practical material, start with the Ollama post, then the prompt-template post, keeping the latter's provenance claims at arm's length. The robotics post is the most consequential but also the least specific. Sophia's three labels form a compact visual-history thread, while the Boston and Hangzhou notes are lighter personal signals.

Peter Steinberger, whose profile identifies him with OpenClaw and OpenAI, supplied the highest-like post in the sample: "First time in Boston. Very European city vibes. Great sea food." It had 1,184 likes, 15 reposts, and 111 replies. There is no product or technical context to add, but the unusually large response makes it the clearest example of a personal post breaking through the threshold. 8