How I Built an AI Editor to Tame My News Feed — and Why Computer Vision Needs One Too

The Morning Scroll That Goes Nowhere

Every morning, I do the same thing: grab my phone and see what happened overnight. Dozens of unread messages in WeChat, a dozen new trending topics on Weibo, a long list of "important" push notifications from news apps, and then I dip into a few tech forums for the latest "just announced" or "major update" posts. After fifteen minutes of scrolling, I'm left with a strange feeling: I know a lot of stuff, but I can't name one thing that actually matters today.

This is the information paradox of our time. We used to worry about not having enough information. Now the problem is too much—so much that it swallows not just our time but our judgment about what deserves attention. So I ran an experiment: I built an AI agent to curate my daily news, and it worked better than I expected. But along the way, I realized the same principles apply to computer vision systems that have to make sense of a flood of visual data.

Step One: Stop Relying on Random Feeds

My first attempt was simple: set a morning task at 8 AM for the agent to gather five tech news items and send them to me. That worked for a few days. But then I noticed the same story appearing three times with different angles, and yesterday's news resurfacing as "latest" just because another outlet covered it. The agent wasn't lazy—it was just scraping the open web, which is a bag of mixed signals: official announcements, media reports, second-hand commentary, reposts, and clickbait all look like "new" information.

The fix? Curate the sources. Over the years, I've collected 161 RSS feeds. But quantity alone doesn't guarantee quality. You need a system. I sorted them into tiers by trustworthiness: primary sources like official company blogs (OpenAI, Google DeepMind, Anthropic), then serious outlets (Bloomberg, WSJ, Reuters), then aggregators and niche sites (The Verge, TechCrunch, MacRumors), and finally individual bloggers and KOLs for hands-on impressions.

Step Two: Organize Your Information Flow Like a Tree

With over a hundred feeds, management becomes a chore. I use Folo, an RSS reader that lets me organize feeds into folders—tech, gaming, culture, AI, cars. It feels like flipping through a magazine I've subscribed to myself. But the real breakthrough was Folo's CLI tool, which turned my curated feed list into a library my AI agent could query directly. Now the agent reads only from my pre-filtered sources, not the entire internet. Each item comes with a direct link, so the agent can't hallucinate URLs.

This is where the computer vision parallel hits home. A vision system fed raw, uncurated video streams will drown in noise. Just as I needed to build a ladder of trusted sources, a CV pipeline needs a hierarchy of data streams—some high-confidence, some exploratory—so that the model doesn't chase every pixel that moves.

Step Three: Teach the AI to Read Your Mood

Even with good sources, the agent initially pushed too much AI news. I told it directly: "I want more consumer electronics and hardware news, less AI model chatter." It wrote that preference into a memory file and adjusted its filters. The next morning, my briefing was full of laptop specs and phone release dates. That's the beauty of an agent: it doesn't know you by default, but it can remember what you dislike.

In computer vision, this is akin to fine-tuning a model on your own preferences. A generic object detector might flag every car, but your autonomous robot needs to focus on pedestrians and traffic lights. You have to teach it what matters in your context.

Step Four: Turn Your News into a Visual Paper

Chat bubbles are fine, but I wanted a better reading experience. So I asked the agent to generate an HTML page with a minimal card layout for my top five stories. Each card had a title, a summary with multi-source verification, a "why it matters" note, and buttons linking to original sources. It looked like a clean digital newspaper, and it felt like taking back control.

For a longer-running topic, like the foldable iPhone rumors, I asked the agent to build a live "dynamic encyclopedia" page. It had a status summary, a tree diagram of specs, a timeline of rumors, a keyword frequency chart counting independent sources (not reposts), and filterable cards with credibility labels: confirmed, multi-source, single rumor, unverifiable. The agent even avoided treating reposts as independent evidence. That's exactly the kind of rigor a computer vision system needs when fusing data from multiple sensors—each source must be weighted by its reliability, and duplicate information shouldn't be counted twice.

Why Computer Vision Needs the Same Discipline

You might wonder what a news curation hack has to do with computer vision. The answer is: everything. Both are about making sense of a noisy, redundant stream of data. In computer vision, you have cameras, lidar, radar, and each sensor provides a partial, sometimes conflicting view. A robust system doesn't just average them; it learns which source to trust in which context, and it fuses information to reduce uncertainty.

My news agent does the same: it cross-verifies details from multiple outlets, prefers primary sources, and ignores reposts. The result is a more accurate picture of what's happening. For a CV system, the equivalent is sensor fusion with confidence weighting. If one camera is occluded, the system relies more on the other. If a radar report conflicts with a camera, it checks the timestamp and source reliability.

In 2025, Merriam-Webster chose "slop" as its word of the year—AI-generated, low-quality content that floods our feeds. The same slop can poison a computer vision dataset if you scrape random images from the web without curation. Garbage in, garbage out. That's why building your own reliable information pipeline, whether for news or for training data, is the best defense against AI-driven noise.

The Takeaway: Build Your Own Dam

Outside information will always pour in like a flood. You can't swim faster, but you can build a dam. For my news, it's a curated set of RSS feeds and an agent that filters and verifies. For computer vision, it's a data pipeline that curates, labels, and validates sources before they reach the model. Both require deliberate choices about what to trust.

The end result? I spend less time scrolling and more time reading things that matter. And when I see a headline, I know whether it's worth my attention. That's the same clarity a well-built CV system achieves: it sees the world not as a blur of pixels, but as a structured scene with known objects and reliable cues.

So whether you're training a vision model or just trying to survive your news feed, remember: curation beats consumption. Filter early, verify often, and don't let the noise drown out the signal.