big 🚢 this week and then for devday 🚢🚢🚢🚢🚢🚢
September 15: Thirteen posts in a scan of 71 of 86 public seed accounts
A September 15 scan of 71 public seed accounts surfaces thirteen standout posts, led by OpenAI's shipping week before DevDay, Claude Mods landing, and the day's field notes on running agents inside real repositories.
Tuesday's scan across 71 of the 86 configured seed accounts returned 1,169 posts. Of those, 125 were published between 00:00 and 24:00 Beijing time on September 15, 2026, and 25 of them were retweets, which are left out here. Fourteen original posts cleared the 100-like threshold; one was a captionless image post with nothing to report, so this edition carries the other thirteen. The seed accounts cover roughly 2.2% of @hwwaanng's network of about 3,850 accounts, and the full following list is still out of reach, so read this as a sample of a large network rather than a complete morning scan.
Release season and platform moves
At 22:46 on Tuesday night, Sam Altman (@sama), OpenAI's chief executive, wrote "big 🚢 this week" and then "and then for devday", quoting Tibo (@thsottiaux), who works on Codex and ChatGPT at OpenAI, saying the week's releases will reach the level people expected from DevDay 2025. DevDay is OpenAI's developer conference, this year on September 29 in San Francisco, so the post reads as a shipping schedule for the next two weeks. It drew 11,129 likes and 496 reposts. 12
Loading content card…
At 01:30, Boris Cherny (@bcherny), who works on Claude Code at Anthropic, said Claude Mods are landing now and pointed out that someone has already built a Tetris mod inside Claude. A Mod is a plugin built on the tool's function hooks: a TypeScript entry point that can watch Claude Code's events, refuse an action, or change how part of the interface renders. The team says function hooks will ship in weeks rather than months, which leaves the feature on an experimental path, and the issue Cherny links carries the design notes along with the source of the first three built-in mods. The post drew 2,591 likes and 164 reposts. 3
Loading content card…
At 09:40, Bao Yu (宝玉, @dotey), an AI engineer who writes about AI and software engineering, reported that ChatGPT gift cards are now on sale in the United States at Best Buy and Giftly, and gave the redemption page. He was quoting OpenAI's Jason Liu (@jxnlco), whose post explains what a card does: it tops up a ChatGPT wallet, and the balance pays for subscriptions, renewals and usage credits on the web. Redemption is limited to US accounts billed in dollars. The post drew 559 likes and 83 reposts. 4
Loading content card…
At 03:32, Bao Yu summarised OM-1, the first robot foundation model from Reward AI. The company's claim is that OM-1 learns only from human manipulation data, and that it transfers to table-top arms, industrial arms and humanoids without fine-tuning for each body. Bao Yu traces the work back to Stanford's DexCap project and describes the Omnibody Hand, a seven-degree-of-freedom wearable that records vision, touch and force while its wearer cooks, tidies or works, so training data accumulates as a by-product of ordinary tasks. In the demonstration videos he describes, OM-1 packs phones with four arms, mixes drinks, folds laundry and unplugs a network cable at real speed, and Reward AI says a new task takes under 30 minutes of human demonstration. The company has published no commercial timeline or pricing. The post drew 108 likes and 21 reposts. 5
Loading content card…
Agent engineering, as practised
At 02:52, Peter Steinberger (@steipete), who works on OpenClaw, called a new OpenClaw feature "so useful" and said he is pushing to bring it into Codex as well. The feature is "suggested task", described in the post he quotes by OpenClaw engineer Patrick Erichsen (@Pat_Erichsen): when an agent finds a piece of work scoped tightly enough to stand on its own, it proposes opening a separate session for that work, so delegation stops depending on the person noticing the seam. The post drew 848 likes and 26 reposts. 6
Loading content card…
At 11:58, Bao Yu argued that verification has become the scarce part of building with agents. He describes adding features faster than he can test them, and says the same pattern is visible in Codex and Claude Code, where minor bugs survive because development outruns testing. He tried driving tests with computer use and found that it left the work of covering scenarios where it was. His expectation is that testing is one of the roles AI is least likely to shrink. The post drew 245 likes and 27 reposts. 7
Loading content card…
At 11:26, Bao Yu described a failure from the same habit. Throwing batches of tasks at an agent near a usage-limit reset, he found work that reported itself complete but had never reached his main branch, because the agent had done it in a git worktree — a separate working copy of the repository — and left it unmerged. He had the agent audit the leftovers, clean up the ones worth keeping, and then added a rule to the AGENTS.md file his agents read from the repository: a task's worktree has to be merged or archived and removed before the task closes. The post reproduces the three prompts he used. It drew 133 likes and 18 reposts. 8
Loading content card…
Tools, workarounds, and one argument
At 08:58, Lex Tang (@lexrus), an iOS and macOS developer, stacked up what each launcher costs in memory: Tinycast at 120 MB, against Raycast's 291, 280, 195, 25 and 10 MB and more. Memory usage troubles him little, he says. His question is why he still pays: free, open-source Tinycast already covers 90% of what he needs. Tinycast is an AGPL-licensed native macOS launcher. The post drew 404 likes and 10 reposts. 9
Loading content card…
At 09:20, Nyarime (奶昔🥤, @realNyarime), a developer who writes about networking, wrote up an automation for anyone routing traffic through Cloudflare tunnels from a Chinese mobile connection, where the chosen edge IP decides the speed. The setup needs three things: a machine that stays on, such as a NAS; an OpenWrt router running a proxy plugin like OpenClash or Nikki; and a scheduled script from the CFData-WEB project that benchmarks Cloudflare IPs and commits the fastest ones to a GitHub repository. The tunnel's backend subscribes to that repository, and a final step in the script tells the proxy to reload, which closes the loop. Nyarime runs the selection four times a day and says it has held for about half a month without manual intervention. The post drew 402 likes and 41 reposts. 10
Loading content card…
At 18:00, Nyarime posted a third-party server that puts 12306, the ticketing site run by China Railway, within reach of an AI assistant, and asked half-jokingly whether building on it would get him arrested. The project, 12306-mcp, exposes ticket search, train filtering, and stop and transfer queries through the Model Context Protocol, the interface assistants use to call outside tools. The post drew 307 likes and 23 reposts. 11
Loading content card…
At 12:44, Yue (@caiyue5), the founder of Sup Space, offered a claim he labelled a hot take: the mainland-China App Store has become a stagnant pool. He gave no supporting data. The post drew 274 likes and 4 reposts. 12
Loading content card…
Cultural archives
At 10:03, Sophia (@SophiaFioren), an account that posts antiquities, published an Egyptian broad collar — a Usekh — of gold set with carnelian, turquoise and glass beads, its terminals shaped as falcon heads, dated 1480 BCE. The post drew 123 likes and 17 reposts. 13
Loading content card…
At 01:12, Sophia published an Islamic talismanic pendant: lapis lazuli carved with Surah al-Ikhlas in Kufic script, mounted in a gold frame, dated to around 700. The post drew 115 likes and 26 reposts. 14
Loading content card…
References
- 1
- 2OpenAI DevDay 2026
openai.com
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.