[KongchangAI]
· 3 min read· 1,787 words

Behind the Jensen Huang Resignation Rumor: How the DSH Open-Source Plugin Uses AI to Rebuild the Intelligence Flow

Behind the Jensen Huang Resignation Rumor: How the DSH Open-Source Plugin Uses AI to Rebuild the Intelligence Flow

A Bilibili creator tests a DSH sidebar plugin that filters 295 news items into 5 and prepares context for follow-up questions.

Based on a Bilibili creator's hands-on share about a DSH plugin, this article argues that the real problem of information overload is not too little news but that no one judges for you. The plugin exists as a sidebar next to the chat box, turning the information flow into queryable context and addressing the triple cost of time, judgment, and context. The tested funnel is 295 items in, 104 scored, 8 above threshold, 5 shown. Its core design removes fame bias by excluding source names from the scoring prompt. It is fully open source and explicitly avoids RSS/ATOM support, a front-end config UI, and historical backfill.

Title

Behind the Jensen Huang Resignation Rumor: How the DSH Open-Source Plugin Uses AI to Rebuild the Intelligence Flow

Meta Description

Starting from the rumor of Jensen Huang resigning from NVIDIA, this piece analyzes the DSH open-source plugin tested by a Bilibili creator: a filtering funnel that takes in 295 news items and outputs 5, a 5-axis scoring system, a design that removes fame bias, and how it turns the information flow into a context you can query.

Article Content

295 news items came in, and he only read 5. This numerical contrast comes from a Bilibili creator's hands-on share about a DSH plugin, and what triggered this share was an analysis he scrolled past in the morning titled "Jensen Huang Resigns as NVIDIA CEO."

The truth of this news is unknown, and the comment section has long been split into two camps. The creator's stance is clear: he does not vouch for the truth of this news, but only discusses what stands between it and your question. The answer is—one click, and a sidebar that can complete the filtering and question-preparation work for you.

The Real Problem of Information Overload: It's Not That There's Too Little News, It's That No One Judges for You

The creator breaks down the cost of traditional news scrolling into three layers: time cost, judgment cost, and context cost. When you see an interesting headline, you first open three tabs and flip through two reports—20 minutes pass, and the nature of your follow-up question is long gone; the same event has ten different framings, and you have to act as your own editor to decide which to trust; news lives in the browser while the conversation tool lives on the desktop, and everything in between relies on copy-paste—by the time you finish pasting, you've already forgotten what you wanted to ask.

There's so much news that no one judges for you

This explains why RSS readers aren't enough—subscribe to 200 sources and all you get is a longer unread list; aggregation tools solve "don't miss anything," while this plugin solves "which five items should I read, and what can I ask after reading them." It doesn't give you unread red dots; it only tells you which few items to read today.

What the DSH Plugin Is: Turning the Information Flow into a Context You Can Query

In one sentence: a sidebar installed in DSH that turns an information flow you could only scroll through into a context you can query.

It doesn't build a webpage or standalone software, and the reason is product-driven—after reading a webpage you still have to move the link back to the chat window, and standalone software competes with the main chat window for screen space; only a sidebar right next to the chat box truly works. When the position is right, the action flows: it's right there after you finish reading, so you'll naturally ask a question.

The structure has two layers: the upper "Today's Must-Read" gives judgment conclusions, and the lower "Raw Stream" expands as evidence. Conclusions must be traceable to evidence, yet evidence without conclusions is too tedious to read—this two-layer design is exactly meant to satisfy both reading states at once.

The Tested Funnel: 295 In, 5 Out

The plugin is backed by a set of tested data: 295 signals entered the database that day, short pieces and low-quality items were filtered out first, and only 104 were actually sent to the model for scoring; among them, 8 reached the 80-point threshold and qualified for Today's Must-Read, but the homepage hard-caps at 5. This "5" is deliberate—any more and it becomes another endless list.

Behind this is a set of tested data

Filtering relies on 5-axis quantitative scoring, evaluating across 5 dimensions such as impact, technical depth, and timeliness, with 6 public sources connected by default, covering blogs, communities, and daily digests. Scoring sets 3 thresholds: 80 points enters Today's Must-Read, 85 points earns an S-level badge, and low-quality signals below 60 are filtered directly from the raw stream.

Taking that Jensen Huang analysis as an example, the card lays out four things: a high score of 100, an S-level badge, five-element reporting, and "why it matters"—the in-database summary analysis states that he controls tens of billions in chips and hundreds of billions in leasing guarantees; once finances tighten, compute credit and capital ties will contract, directly shaking startup financing and the open-source ecosystem landscape. The creator only added his own follow-up: "How will the stock price decline after Jensen Huang resigns affect the industry chain? If there's no exact data, just say there isn't any, and only analyze the transmission path." The answer on the right didn't fabricate data; instead it broke out three lines: changes in cloud providers' capital expenditure, accelerated substitution by in-house ACC, and tightening compute leasing guarantees affecting startup financing.

5-axis quantitative scoring refers to a mechanism that breaks news quality into five independent dimensions and scores each separately. Common dimensions include impact (the scope of the event and the number of stakeholders), technical depth (whether it involves verifiable technical details rather than generalities), timeliness (relevance to the current time window), source reliability (whether the original source is traceable), and information density (the amount of effective information per unit of length). Each dimension is scored independently, then weighted and aggregated into a total score out of 100, which is then layered by threshold. The value of this design lies in breaking the vague judgment of "is this news important" into explainable sub-questions, avoiding a single dimension (such as popularity) dominating the ranking, and letting users understand why a piece of news scored high or low instead of facing a black-box score.

Removing Fame Bias: The Most Core Design Decision

When scoring, the source name is not written into the prompt given to the model—the model has no idea which big-company blog this news came from, and all weight is settled post-hoc in code. Fame cannot move the score. This is the most noteworthy design of the entire plugin: it counters the most hidden source of bias in the information flow, namely "who said it" overpowering "what was said."

Source names as a source of prompt contamination is an easily overlooked source of bias in LLM scoring scenarios. Large models have been extensively exposed to the names of well-known institutions during pretraining; when a prompt contains "a certain big-company blog" or "a certain top journal," the model unconsciously raises its prior judgment of content quality even if the body text itself is mediocre. This bias is isomorphic to the human editor's "halo effect." Stripping the source name from the prompt and instead applying post-hoc weighting to the final score at the code layer is equivalent to decoupling two variables: "content quality" and "source weight." The former is left to the model's judgment, while the latter is explicitly controlled by the developer—preserving the source as a reference factor while preventing it from stealing the show.

What It Explicitly Doesn't Do: Three Things That Draw the Boundaries

  • No RSS/ATOM feed support—web interfaces are unstable, and connecting them means a long-term maintenance black hole
  • No front-end source configuration UI—just edit a config file, no piling up 10 input boxes
  • No historical backfill—only today and yesterday are considered

Web interfaces are unstable

In terms of rhythm, it manually syncs 5 seconds after the host starts, then syncs every 60 minutes, with at most 120 model calls per day, and it only fetches without scoring. The plugin is fully open source, and the code can be viewed and modified.

RSS/ATOM feeds are a standardized content distribution protocol that has been in use since the early days of the internet: websites expose updates in XML format at a fixed URL, and readers pull them at intervals. It seems open but is actually fragile: many websites have stopped maintaining feeds, or only output summaries rather than full text, and interface fields and rate limits vary widely—integrators need to write separate adaptation and fault-tolerance logic for each source. For a small plugin, this means ongoing maintenance cost rather than one-time development cost, which is exactly why the article calls it a "long-term maintenance black hole." By contrast, the 6 public sources the plugin connects by default are selected interfaces with relatively stable structures, trading limited coverage for sustainability.

Three Steps to Get Started and One Habit Worth Building

Step one: install the plugin with a command; after installation, press CMD+Q to fully quit the desktop app and restart—just refreshing the window won't work. Step two: configure the model in the host; you can still view the raw stream without configuring it, just without scores and summaries. Step three: click the sidebar icon in the top right corner to open it directly.

Back to that Jensen Huang analysis from the beginning. Whether it's true or false, and where it goes next, the creator himself doesn't know, nor does he expect a plugin to judge the truth for him. But the plugin prepared the judgment and context in advance, so he didn't get stuck on "where to start thinking" and within 30 seconds asked the question that was truly valuable.

This may be a more pragmatic use of AI tools: not drawing conclusions for you, but lowering the threshold for asking questions enough. Next time you scroll past a breaking news story, your first reaction might as well become—how will this affect me three months from now?

One-line Summary

A Bilibili creator tests a DSH sidebar plugin that uses 5-axis scoring to filter 295 news items down to 5 and prepare the context for follow-up questions.

Paragraph Summary

Based on a Bilibili creator's hands-on share about a DSH plugin, this article discusses that the real problem of information overload is not too little news but that no one judges for you. The plugin exists as a sidebar right next to the chat box, transforming the information flow into a context you can directly query and solving the triple cost of time, judgment, and context. The tested funnel is 295 items entering the database, 104 entering model scoring, 8 reaching the 80-point threshold, and at most 5 displayed on the homepage. The core design is removing fame bias: source names are not written into the scoring prompt, and weights are settled post-hoc in code. The plugin explicitly draws three boundaries—no RSS/ATOM support, no front-end source configuration UI, and no historical backfill—and is fully open source. Using the Jensen Huang resignation rumor as an example, the article shows how the plugin lets people ask the truly valuable question within 30 seconds rather than judging truth for them.

#DSH plugin#information overload#AI news filtering#5-axis scoring#open-source plugin#Jensen Huang#context
Share:

Related articles