Stringers
SAOS-3624/dsh-see-now · SAOS-3624

"Five minutes for twenty cents": a Bilibili UP host got DeepSeek to watch the screen cheaply. 9 days to 530 stars

What it is A real-time vision plugin for DeepSeek Harness: it picks frames by how much the screen changes, stitches several frames into one image, sizes it to the model's billing grid, and costs almost nothing on a static screen.

Why it took off The author runs a Bilibili series on "ultra-cheap DSH plugins." This episode hooked viewers with "a few minutes for a few cents," which directly addressed the fear that always-on screen watching burns too many tokens. The video has 350k views, and most of the stars trace back to it.

BHow you could follow this Don't build another DSH screen watcher. Borrow its open-sourced frame-selection and stitching method and point it at something that "watches all the time" (phone, games, cameras), or port it to another host and language. Ship a demo video with concrete prices the same day you create the repo.

Grade B: a clear play, low cost, and the method is open and reusable, but this exact niche is taken, the star ceiling is probably a few hundred, and whether it takes off depends entirely on whether your video gets pushed. Chatter in the niche
Data (updated daily) · Repo created 2026-11-01 · 714 stars · 7 days +106
GitHub · stars
020040060080011-0311-0711-1111-1611-20714
Data 2026-11-03 → 2026-11-20 (18 days, 18 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

SAOS-3624/dsh-see-now is a real-time vision plugin for DeepSeek Harness (DSH). It keeps the AI watching your screen, but only sends the frames where the picture changed, stitched into one image, so a static screen costs almost nothing. The author is a Bilibili UP host with 5,500 followers. Ⅰ posted a demo video 22 minutes after creating the repo, and by 11-10 it had 350k views. Within 9 days the repo grew to about 530 stars, nearly all of them from that video. Below: what it does, where the stars came from, why a similar project from 10 days earlier has only 18 stars, and whether you could do this yourself.

Repo created at 9:30 PM, video posted at 9:54 PM

2026 Nov 1, Beijing time 21:32: on GitHub, SAOS-3624, an account created in Oct 26, created the dsh-see-now repo and published the first release one minute later. 21:54, Bilibili UP host "苍穹晓雷" posted a 3-minute 31-second video titled "Ultra-cheap real-time vision DSH plugin!" The first line of the description is the repo link, followed by a paragraph:

"I actually built it! Works even better than I expected! In five days I built full vision from scratch. Plenty of parameters still aren't tuned, but I was too excited and posted it before the audio was even synced!!! I'm so excited, everyone!!!"

The video sits under the "DeepSeek Big Fish" topic (大肥鱼 is Bilibili's nickname for DeepSeek) and was later added to Bilibili's "Trending" feed. It has 11-10:352k views, 20k likes, 29.5k favorites, and 6,405 comments (from the Bilibili video API). Favorites outnumber likes, which suggests many people saved it to install later (inferred).

The comment section shows an unusual pattern: a batch of accounts named "Someone's Big Fish" are replying on behalf of their owners. One comment reads "@新萌迷途 Power poles, the CN Tower, sea turtles at the aquarium, all named in five minutes for twenty cents," and another "Little Big Fish" replies: "The best part of twenty cents is that you dare to leave it running. At the price where you'd turn it off after one look, I can't afford it." Judging by the names and tone, these are viewers' own DSH instances posting after watching the video (inferred). A large share of the people who watched it already run a "Big Fish" on their own computer.

What it does: not "can it watch," but "can you afford to let it"

The obvious way to have an AI watch your screen is to send it a screenshot on a timer. The README opens by listing the problems with that: a screenshot every 2 seconds means a minute of 30 frames that are mostly identical, and "tokens burn through the floor"; a screenshot every 30 seconds means you lose everything that happened in between, and the AI can only say "looks like you're using a computer."

See-now's approach is to decide first whether a frame is worth looking at, then decide whether to send it:

  • Score each frame for change: from 0 to 255, compute color differences on a 64×36 grid of small cells. Only frames where the picture suddenly changed, or where motion has built up enough, are kept. When the screen is still, almost no images go out. Every 20 seconds it sends one fallback frame to tell the AI "nothing changed."
  • Stitch several frames into one image: frames 4 through 8 that survive the filter go into a single image. An image has a fixed price, so packing in 4 frames or 8 costs the same. That makes "keeping an extra frame is nearly free, missing one is expensive."
  • Size the image to the model's billing grid: DeepSeek-V4 charges for images per 42-pixel tile (per the README, not independently verified). The author changed the canvas width from 800 to 798, which comes out to exactly 19 tiles, saving 11 to 24 tokens per image. The default setting makes a 798×462 image worth 222 tokens. The audio track is also placed in leftover pixels that are already paid for, so it adds no cost.

The README also works through the longer-term math: why costs rise the longer you watch, and roughly how many images later the host's "image omission" mechanism breaks the cache. The price the comments repeat most comes from these numbers: "Watching for 6.4 minutes costs eight cents" and "Five minutes for twenty cents, cheaper than my electricity bill."

Beyond that, it has a floating window (the AI's replies are displayed and read aloud), a /seenow command, and system-audio transcription, and it captures on Windows, macOS, and Linux. Install with one line: dsh plugin add dsh-see-now. It shipped 11 versions in 8 days, and more than a dozen outside developers opened PRs. The Linux capture and process recording were contributed by others.

Where the stars came from: one video, a 1.5 per-thousand conversion

DatestarGained that dayfork
11-03160—8
11-04252+9219
11-05334+8229
11-06383+4937
11-07417+3441
11-08461+4443
11-09493+3246
11-10527+3453

(GitHub daily star counts; the breakdown for 11-01 and 11-02 wasn't available. Later, the GitHub page showed 11-10 stars and 534 forks. 54)

350k views brought in about 530 stars, a conversion rate of roughly 0.15%. The author's previous project, dsh-video-storyboard (created 10-26, which lets DSH watch video files cheaply), had two videos totaling 486k views and 495 stars, roughly 0.1%. The two repos land at the same scale, which suggests the stars largely come from the videos. On X, the only posts from the past 7 days were two small-account posts with 0 likes each, and no English-language media has covered it.

Fewer people are starring it, but the number of users hasn't dropped. npm downloads peaked at 11-03 on day 3,593, then settled at about 2,200 to 2,500 a day, totaling 18,775 over 30 days (npm download stats). How much of that is double-counted from upgrades and mirrors is unclear.

One more thing: this is a series. The creator's videos are all titled "Super Cheap X". The 10-26 episode on the video-watching plugin got 332k views, the sequel 10-28 got 154k, and the 11-01 episode on the "watch together" plugin got 65k, same night it went see-now. The audience and recommendation tags built up by the previous episode carried straight over to the next one.

The same idea, but the one that came 10 days earlier has only 18 stars

"Let the Big Fish watch your screen with you" wasn't see-now's idea first. On 10-22, pairun325/dsh-whale-sense ("Whale Watch") went live: it took a screenshot every 2.5 seconds and only spoke when the picture changed, with a cooldown and an hourly cap. It also posted a Bilibili video, which got 84k views and 4,253 likes. So far it has 18 stars.

The top comment on its video is "That must burn a lot of tokens, right?"

The gap between the two projects comes down to that one line. whale-sense proved people want "watching together," and it also exposed the worry. see-now turned the worry into the selling point: the title says "ultra-cheap," the demo quotes the price outright, and the README has a full section walking through the math. Both were Bilibili videos. One stalled at 80k views and 18 stars; the other reached 350k views and 530 stars.

The timing is also right. DSH has been hot in the Chinese developer community for months: the official repo has 277k stars, and the "DSH: everything is a plugin" topic on Bilibili has 23.7M views. Hosts can now bill image inputs per tile, so "always watching the screen" is practical for the first time. What was missing was someone pushing the cost down.

Who else is in the space

ProjectScaleCompared with alternatives
liustack/modlens4,879 starsDSH's earliest vision plugin: paste one image and convert it to structured text. Not continuous watching.
ysr666/dsh-vision-router1,253 stars, 65k npm downloads in 30 daysGives text-only models an image channel; per-request image viewing
Anionex/dsh-vision-toolkit936 starsPasting images for recognition, screenshot-to-UI restoration; also single images
About 40 official ones, such as dsh-computer-useAbout 59k npm downloads per month from official packagesScreenshots are for operating the computer for you, not for watching along
pairun325/dsh-whale-sense18 starsContinuous watching, but cost is unresolved
Doubao desktop, Gemini Live, ChatGPT screen sharingBig-company featuresFree and works out of the box, but you can't use your own DSH or key

(GitHub and npm, as of 11-10.)

Inside DSH, "have the model look at one image" is already crowded. "Watch continuously and cheaply" has only see-now as a serious option so far. Downstream projects are already building on it: nacivageer/dsh-see-now-more (11-09) added a chat-style floating window, and someone in the issues said the frame-selection criteria could be copied straight over to control a phone via adb and "work right away." No direct copycat with the same features exists yet.

If it were you

B: a clear play, low cost, and replicable, but this exact niche is taken. The ceiling is probably a few hundred stars, and whether it takes off depends on whether your video gets pushed. Two things are worth learning from it: in an agent ecosystem that's booming, find a feature people want but fear is expensive, then use math to make it something people dare to leave running; and ship a demo video with concrete prices the same day you create the repo.

Who it suits: developers who can write screen or device capture and are patient enough to do the billing math on how models charge; people willing to record and explain on camera, with a Bilibili or Douyin account (a few thousand followers is enough; the author only has 5,500).

Who it doesn't suit: people who only want to publish a repo and skip the video (whale-sense got only 18 stars even with a video; without one it's harder still); people aiming for thousands of stars; people planning yet another "DSH screen watching" plugin.

How to do it: three ways to follow, ranked from best to weakest:

  1. Swap in a different "always watching" target: phone screen mirroring (someone has already confirmed the criteria transfer), game companions, cameras watching the house or pets, live streams or online classes. The change-gating plus stitching method is Apache-2.0 open source and can be borrowed directly.
  2. Swap the host or the language: almost no one in the English community talks about it. Build the same thing for other agents and models that bill per image. Each provider's billing grid differs, so redoing the math becomes your own README.
  3. Make it a series: follow the author's approach of running several episodes on one money-saving idea (watching videos, watching together, watching the screen), with consistent titles, and have each episode funnel viewers to the next.

How to get users: post the video the day you create the repo, with a title using words like "ultra-cheap" and one concrete price, put the repo link in the first line of the description, and attach trending topics from the ecosystem. Write the README with real numbers, and the people who understand the math will come in and open PRs. Keep the package name free of prefixes, make install a single line, then submit to the plugin aggregator sites. Run the English version through X and Reddit.

Timing: see-now's daily star growth has already dropped from 92 to 30, so the freshness of DSH plugins probably has a few months left (inferred). Ideally ship before Q1 2027.

Signs you've misread this: ① your video passes 100k views but the repo gets fewer than 100 stars in a week, meaning video-to-stars converts worse than estimated here; ② DSH officially releases its own token-efficient real-time vision, which takes this niche away; ③ conversely, if someone takes the same play to a thousand stars, the ceiling is higher than estimated here, and you should move faster.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report
Competitors
NameWhereLaunchedScaleHow it differs from this one
pairun325/dsh-whale-sense (Whale Watch)Open source2026-10-2218★; Bilibili video 84k viewsScreenshot every 2.5 seconds, speaks only when the picture changes; no stitching or billing-grid design; the top comment worries about token cost
liustack/modlensOpen source2026-02-224,879★Pastes a single image and converts it to structured JSON; static, does not watch the screen continuously
ysr666/dsh-vision-routerOpen source2026-08-131,253★; npm 64,589 downloads in 30 daysGives text-only models a free vision chain; per-request image viewing, no streaming
Anionex/dsh-vision-toolkitOpen source2026-08-13936★Pasting images for recognition, multi-image Q&A, screenshot-to-UI restoration; static
DSH computer-control types (official dsh-computer-use, dsh-computer-use-win, etc.)Open sourceFrom ¥2026-08Official npm ~59k downloads/mo; win build 554★Screenshots are for acting on screen for you, not for watching along. Saving tokens isn't the core point.
SAOS-3624/dsh-video-storyboard (same author)Open source2026-10-26495★; npm 28,492 downloads in 30 daysSame 42px alignment + multi-frame-per-image approach, but for video files; see-now moves it to live screens
oxbshw/watch-skillOpen source2026-07-05563★Turns video/audio/screen into a searchable timeline. General-purpose agent, doesn't pitch token cost.
Doubao desktop app shared screenBig-company features2025 (exact date unverified)Bilibili review video 276k viewsFree and works out of the box, but not inside DSH and can't plug into your own agent
Gemini Live / ChatGPT voice screen sharingBig-company featuresGemini 2025; ChatGPT ~2024-12 (unverified)HugePrototypes of live screen watching. see-now is the cheap version for DSH users or people bringing their own key.
Where users come from
  • Verified the author's own Bilibili video: BV1c3xb6UEWE, posted 2026-11-01 13:54 UTC (22 min after the repo was created), 351,570 views / 20,184 likes / 29,475 favorites / 6,399 comments. The first line of the description is the repo link.
  • Verified Same author's Bilibili "super cheap DSH plugin" series cross-promoting: storyboard two episodes 332k + 154k views, watch-together 65k, see-now update video 38k/7,582, all titled "super cheap"
  • Verified Platform event: DSH ecosystem boom + Bilibili topic: Bilibili topic "DSH is everything a plugin" 23.7M views; videos tagged "DeepSeek fat fish"
  • Verified DSH plugin aggregators/marketplaces + unscoped npm package name: indexed on dshopen.org, findharness.com, dshpluginhub.ai, etc.; install with dsh plugin add dsh-see-now
  • Inferred Chinese AI daily newsletters/communities, awesome lists: recall card says the tweet had about 260★ at the time and it was added to an awesome list; the original post wasn't found this round.
  • Verified X: only two small-account posts with 0 likes in the last 7 days, almost no contribution

SAOS-3624/dsh-see-now: why it could run

Channel: GitHub (open source, Apache-2.0). Data as of 2026-11-10. Sources: our github repo tables (repo_day / repo_profile), GitHub API, npm registry / downloads API, Bilibili public endpoints (read via Jina), Jina / Exa search, live X search (x live).

Bottom line

The repo's stars came from the author's own Bilibili video, not from GitHub or X. The author goes by "苍穹晓雷" on Bilibili, had about 5.5k followers then, and was building a "super cheap DSH plugin" series. 22 minutes after the repo was created he posted "Super cheap near-real-time vision DSH plugin!" That video now has 352k views, 20k likes, 29k favorites, and 6,399 comments, and the first line of the description is the GitHub link. The previous episode in the series, the same author's dsh-video-storyboard, had 332k views, so the audience for "cheaply letting DeepSeek (the big fat fish) watch video" was already built.

The product's pitch is that it turns "AI watching your screen in real time" from "tokens burned through the roof" into "a few cents for several minutes of watching." It does three things: filter frames by how much the picture changed, tile several frames into one image, and size the canvas to the host model's 42px billing grid. People already wanted "watch-along" plugins, but the earlier similar one (dsh-whale-sense, which screenshots every 2.5 seconds) got the first comment "that must burn a lot of tokens." see-now addresses that exact worry and makes it a video hook with concrete prices (comments: "eight cents for 6.4 minutes of watching", "five minutes for twenty cents, cheaper than my electricity").

1. What it is

  • A real-time vision plugin for DeepSeek Harness (DSH), also usable as a standalone Node library. It captures the screen, a webcam, or a video file, gives each frame a "change score" from 0–255 (HSV-weighted difference on a 64×36 grid), and keeps only frames where the scene cut or changed enough. The kept 4–8 frames are tiled into one image and handed to the host model. Static scenes almost never produce an image; a fallback frame goes out every 20 seconds.
  • The cost math is in the README: based on the host DeepSeek-V4 billing of one 42px cell per 14px patch, the default setting 798×462 is one image at 222 tokens. Width is aligned to 42 (800→798), saving 11–24 tokens per image. The bottom audio-track bar sits in leftover pixels that get billed anyway, so it costs nothing extra. The README even works out when the host's "image omission" mechanism breaks the prefix cache (about 120 images, a half hour to an hour).
  • Extras: a floating window (the AI speaks via say and it's read aloud), the /seenow slash command, system-audio transcription (WASAPI process loopback / PulseAudio monitor), and capture on Windows, macOS, and Linux (xdg-desktop-portal).
  • Author: the GitHub account SAOS-3624,2026-10-26 was only just registered, with 25 followers and 4 public repos, and no bio (GitHub API). The Bilibili UP host "苍穹晓雷" (mid 109659565) has 5,555 followers, 88 uploads, and 94,701 total likes. The two match: the video description links the GitHub repo and says the full vision plugin was built from scratch in five days.
  • Timeline: repo created 11-01 13:32 UTC, v0.0.2,15:33 released 13:33, published to npm. As of 11-08 there have been 11 releases (up to v0.2.1), and 18 versions on npm. The repo has 100+ commits total, a dozen or so of them from external contributors' PRs.

2. Growth data

Datestars (our github database)fork
11-031608
11-04252 (+92, the largest in the observation window)19
11-0533429
11-0638337
11-0741741
11-0846143
11-0949346
11-10527 (GitHub API later the same day: 534)53
  • Daily star gains: 92 → 82 → 49 → 34 → 44 → 32 → 34, cooling off.
  • npm downloads (api.npmjs.org): 11-01 552,11-02 2,258, 11-03 3,593 (peak), then settled at 2,200–2,500 per day. 30-day total: 18,775 downloads. Stars are falling but downloads aren't, so the people installing it are steadier than the people clicking star.
  • The same author's two other repos (our github tables): dsh-video-storyboard, created 10-26, now 495★, npm 28,492 downloads over 30 days; dsh-watch-together, created 10-31, 36★, 6,898 npm downloads.
  • Quick conversion math: 350k views → about 530 stars, roughly 0.15%. The storyboard's two videos total 486k views → 495 stars, roughly 0.1%. Both repos land in the same range, which suggests the stars come mostly from the videos.

3. Where users come from (growth tactics)

ChannelEvidenceConfirmed
The author's own Bilibili video (main channel)BV1c3xb6UEWE "Super cheap near-real-time vision DSH plugin!", posted 2026-11-01 13:54 UTC, 22 minutes after the repo was created. 351,570 views / 20,184 likes / 29,475 favorites / 6,199 coins / 6,300 shares / 6,399 comments. The first line of the description is github.com/SAOS-3624/dsh-see-now. Tagged "DeepSeek fat fish".Yes
The same series cross-promoting before and after10-26 "I made a super cheap video-watching plugin for dsh!" 332k views (storyboard); 10-28 "A plugin for dsh video watching even cheaper than super cheap!" 154k; 11-01 00:09 "I made another super cheap, fast dsh watch-together plugin!" 65k (watch-together); 11-04 one dsh-see-now v0.1.0 video at about 38k views (from the Bilibili topic page summary; bvid not retrieved); 11-07 "See-now new version released and open-sourced" 7,582. All series titles use "super cheap" and link to each other in Bilibili's related recommendations.Yes
Timing: DSH topic is hot on BilibiliBilibili topic "DSH is everything a plugin" 23.7M views, 120k discussions; many plugin videos under the "DeepSeek fat fish" topic (Bilibili topic page).Yes
DSH plugin aggregators / marketplacesIndexed by dshopen.org, findharness.com, dshpluginhub.ai, googleharness.com, dsh-xray, viralagent.si, theagenticleaderboard.com (Exa / Jina search). The npm package name has no scope: dsh plugin add dsh-see-nowYes (how much traffic it brought is unknown)
XLive X search for "dsh-see-now" and "SAOS-3624" over the last 7 days turned up only two small-account posts with 0 likes (@danielsspo, 6 views; @anoymous_elangm, 18 views). x-radar local corpus: 0 entries.Yes (X contributed almost nothing)
Chinese AI daily newsletters / communities, awesome listsThe recall card says "about 260★ when the tweet was posted" and "added to awesome-dsh-plugins." The original post wasn't found this round.No (unverified)
Paid acquisition / star farmingNot seen. Favorites, coins, and comments on Bilibili are all high, the fork ratio is about 10%, and there are many external PRs. Doesn't look farmed.—

4. Competitors

ProductWhereLaunchedScaleHow it differs from see-now
pairun325/dsh-whale-sense (whale watch-along)Open source + Bilibili2026-10-2218★; Bilibili video 84k views, 4,253 likesClosest in direction, and 10 days earlier: it screenshots every 2.5 seconds and speaks only when the picture changes, with a cooldown and an hourly cap. But it has no tiling and wasn't designed around per-token pricing. The top comment was "that must burn a lot of tokens". That's the exact problem see-now solves.
liustack/modlensOpen source2026-02-224,879★"DSH's first vision plugin": the user pastes an image and it's turned into structured JSON (OCR + layout + semantics). It handles a single static image, not continuous screen watching.
ysr666/dsh-vision-routerOpen source / npm2026-08-131,253★; npm 30 downloads over 64,589 daysAdds a free vision chain to text-only models for pixel-level Q&A / localization. Looks at images on request, no streaming.
Anionex/dsh-vision-toolkitOpen source2026-08-13936★Paste-to-recognize, multi-image Q&A, screenshot-to-UI. Static.
Elohia/dsh-plugin-mm-vision, oil-oil/dsh-vision, linenxi-ctrl/dsh-vision, and othersOpen sourceFrom ¥2026-08A few to a few dozen ★All are "image → text" bridges. Small volume.
DSH computer-control plugins (official @deepseek-ai/dsh-computer-use, Yu-tao-Li/dsh-computer-use-win, and about 40 others)Open source / npmFrom ¥2026-08About 59k npm downloads per month from official packagesScreenshots there are for acting on screen for you, which is "operating on your behalf." see-now is "watching along / observing," and saving tokens is the core.
Same author as dsh-video-storyboard / dsh-watch-togetherOpen source10-26 / 10-31495★ / 36★Same 42px alignment and multi-frame-in-one-image approach. storyboard handles video files; see-now brings it to live screens
oxbshw/watch-skillOpen source2026-07-05563★Turns video, audio and screen activity into a searchable timeline for any agent. Not DSH-specific, and it doesn't talk about token cost
Doubao desktop 'screen share'Big-company features2025 (exact date not verified)Bilibili review video 276k viewsScreen share + voice call, free and works out of the box. But it's not inside DSH, and you can't plug in your own agent
Gemini Live screen share / ChatGPT voice screen shareBig-company featuresGemini: 2025 year(s) (full Android rollout during I/O); ChatGPT: about 2024-12 (unverified)HugeLive screen viewing is what they pioneered. see-now is a 'budget version' for DeepSeek users who open-source it or bring their own key
pluely (open-source Cluely alternative), cueOpen source2025-08 / 2026-072,773★ / 1,514★Screen-watching assistants for meetings / interviews. Different use case

Takeaway: On DSH, "letting the model see images" is already crowded, but "continuous screen watching that's cheap" has only one early mover, whale-sense, and it doesn't solve the cost problem. Big tech has had real-time screen viewing for a while; DSH users just can't access it.

5. Why it took off

  1. The author brings their own distribution: a Bilibili UP creator's 'ultra-cheap DSH plugin' series. The previous storyboard episode got 330k views, so the audience and algorithm tags were already built. see-now was posted the day the repo was created (22 minutes later), 350k views → 530 stars.
  2. The hook is a concrete price: the title says 'ultra cheap', and the demo says '6.4 minutes for eight cents'. The comments repeat the number and play with the 'big fish has eyes now' meme. People in the scene already worry that 'live screen watching burns tokens' (from whale-sense video comments), and this hook hits that exact concern.
  3. Timing: DSH is breaking out (official repo 277k★, awesome-dsh-plugin 24k★, the Bilibili 'DSH is everything a plugin' topic 23.7M views). The host model already bills image input per patch (per the README's billing math), so 'watching the screen' became feasible for the first time, and all that was missing was someone bringing the cost down.
  4. The engineering credibility is what keeps people: the README has hardcore token accounting, 8 days and 11 releases, support for all three platforms, and a dozen-plus external contributors sending PRs (Linux capture, process-loopback audio, the /seenow command). One person in issue #41 said its criteria could be lifted straight onto adb phones and "work right away." npm downloads run steadily at 2,000+ per day.
  5. Low install barrier: the npm package name has no scope, one line dsh plugin add dsh-see-now, and it works after a restart. It's auto-indexed by a cluster of DSH plugin aggregator sites.

6. Differentiation

  • Compared with similar tools, the difference isn't 'can it see', it's how it saves cost. It uses change gating to decide whether to send a frame at all, tiles multiple frames so one image's fixed cost is spread across 4–8 frames, and sizes the canvas to the model's billing grid. Others take timed screenshots and recognize single images; this one turns 'live screen watching' into cost engineering.
  • The differentiation is one reason it took off, because the video's hook is 'cheap'. But differentiation alone isn't enough: the method is public, and the license is Apache-2.0. The real amplifier is the author's own Bilibili account and the serialized videos. The same code from an author without a Bilibili following would most likely get only a few dozen stars, like whale-sense did.

7. Visibility and copycats

  • Visibility: already widely seen in the Bilibili DSH circle (350k views). But almost nobody mentions it on X, there's no English media coverage, and no revenue. It's 'spreading in a small circle', and that circle isn't small, and it's Chinese-language.
  • Followers / downstream: nacivageer/dsh-see-now-more (created 11-09, 0★, added an IM-style floating window and mid-turn interjection around the same sampling core, and said hi in issue #64); iacore ported the criterion to adb phone control (private, issue #41); QiuYeDx/dsh-bili-eyes uses the same author's storyboard frames. There's no direct copy with the same name and function yet.

8. What others can learn

  • 'Build the repo + post the video the same day': Bilibili is the main distribution channel for Chinese AI tools. Demos need concrete prices or concrete result numbers, and a serialized title ('ultra-cheap X') sends viewers from one episode to the next. The first line of the video description carries the repo link.
  • Find the 'everyone wants it but fears the cost' need in a breakout ecosystem, and solve it with cost engineering: align sizes to the model's billing grid, tile multiple frames into one image, gate on change. This token-saving approach carries over to any scenario where an agent keeps watching something (phones, surveillance, games).
  • Writing hardcore billing derivations into the README attracts people who know the field to submit PRs, which builds a maintenance community.

9. Still unclear

  • The Chinese AI daily tweet mentioned in the recall card (about 260★ at the time) and the date it got into awesome lists could not be found this round.
  • We don't have the 11-01 to 11-02 star curve (the database starts at 11-03), so how much it gained on day one is unclear. The GitHub stargazers endpoint returned 404 this time, so we couldn't get a star-time breakdown.
  • It's unclear how many npm downloads come from CI, upgrades or mirrors. The plugin's real active usage is unknown.
  • 'DeepSeek-V4 natively bills image patches' comes only from the README's description. The host model's vision capability and pricing weren't verified separately.
  • Release dates for Doubao and ChatGPT screen sharing were not verified one by one.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.