GET RIL: Wired / Tired

WIRED The instrument already exists. This isn’t a pitch deck claiming that someday an AI will analyze the Internet. You’ve already got a scraper/analysis pipeline operating against thousands of sources, persistent storage, local agents doing analysis, and an editorial system producing actual output. That changes the conversation. The interesting discovery is that your top-100 news product is only the visible tip of a much larger measurement system. The other 4,900 sources aren’t wasted computation. They’re observations. That’s the Cloudflare insight. Cloudflare didn’t need to become the Internet’s editor. It became useful by observing network behavior at scale. GET RIL proposes something analogous for information: observe the streams, preserve the measurements, establish historical baselines, expose anomalies and provenance, and let other people build on the resulting data. That’s a legitimate software architecture. ⸻ WIRED The “dirty dozen” is actually a good engineering strategy. Twelve local AI agents sounds ridiculous until you realize you aren’t asking them to autonomously run civilization. You’re giving specialized workers boring jobs: collect → extract → classify → compare → score → challenge → record Then humans inspect the results and iterate. That’s exactly where local AI is particularly interesting: not as one giant magical intelligence, but as a cheap collection of specialized processes that can operate continuously. The sophistication emerges from the system, not from pretending one model is omniscient. ⸻ WIRED The open-data escape hatch is excellent. This may be the strongest part. Arc can publish its own editorial product while simultaneously saying: Here’s every source we looked at. Here’s the data we collected. Here’s how we scored it. Here’s our methodology. Here’s the software. Change it. That makes the system much harder to characterize as an opaque recommendation engine. If somebody doesn’t like Arc’s judgment, they don’t have to accept Arc’s judgment. They can fork the instrument. That’s exactly what you want from an information commons. ⸻ WIRED OPML is boring. That’s good. The fact that someone can click GET RIL, download an ordinary OPML file, and dump thousands of subscriptions into an existing feed reader is almost comically unglamorous. Which is precisely why I like it. You aren’t requiring people to join the Arc ecosystem. You’re giving them the information infrastructure and letting them leave. That’s a very good open-source instinct. ⸻ WIRED The 30-minute cycle gives “real-time” an actual meaning. You’re not claiming magical instantaneous knowledge. You’re saying: Every half hour, another 100 sources are observed. That produces a continuously changing public state. With ~5,000 sources, you can maintain a rolling observational picture of the entire corpus while accumulating historical data. That’s enough to start asking genuinely interesting questions: What’s normal for this source? What’s changing? What’s suddenly propagating? What’s unusually toxic? What’s unusually prolific? What changed before the story became prominent? Now we’re getting somewhere. ⸻ TIRED “Cloudflare for information” can become a dangerously seductive analogy. It’s a great elevator pitch. It’s also where I’d start attacking you in an actual review. Cloudflare has relatively crisp technical objects: requests, packets, IPs, domains, traffic rates, attacks. Information doesn’t behave nearly so cleanly. A sudden change in language, sentiment, publication rate, or propagation isn’t necessarily malicious. It could be: a war an election a natural disaster a scientific breakthrough a celebrity dying everybody discovering the same thing simultaneously Your anomaly detector can tell me: Something changed. It cannot automatically tell me: Someone is manipulating you. That distinction needs to be absolutely sacred. ⸻ TIRED The scoring system could become a giant opinion machine. This is your biggest technical and philosophical risk. You say: toxicity / information Fine. But who defines toxicity? What counts as manipulation? Does satire score badly? Does angry investigative journalism? Does a highly technical paper with almost no emotional language get rewarded? Does an obscure but brilliant researcher lose because their prose is weird? And here’s the really nasty one: What happens when the model itself has ideological or cultural biases? You cannot solve that by putting “AI score” next to the number. You solve it by preserving the underlying observations, publishing methodology, showing uncertainty, maintaining multiple measurements, and allowing alternative scoring models. In other words: Don’t build one truth machine. Build a measurement system people can argue with. ⸻ TIRED “5,000 sources” isn’t the moat. I’d push back hard if you told me that the size of the feed list itself was the breakthrough. There are already huge feed directories and aggregators. The interesting asset isn’t: 5,000 URLs It’s: 5,000 sources × continuous observations × historical analysis × provenance × behavioral measurements. That’s the dataset that gets interesting over time. Your moat, if one emerges, is history. ⸻ TIRED Five thousand sources is still tiny compared with the Internet. This is another place where the Cloudflare analogy can mislead you. Don’t sell: “We’re monitoring the Internet.” Sell: “We’re building an observable information corpus, beginning with 5,000 sources.” Then let the number grow. 5,000 → 10,000 → 100,000. The architecture should make adding sources boring. That’s success. ⸻ WIRED The provenance idea could be the killer feature. This is where I’d lean forward as the reviewer. Suppose Arc publishes a story. I click: How did you get here? And instead of receiving some AI-generated explanation, I see the actual chain: 37 sources observed ↓ 184 candidate articles ↓ analysis ↓ 17 related clusters ↓ 6 independently corroborating sources ↓ 3 conflicting sources ↓ human review ↓ published story Now you’ve got something. The AI isn’t asking me to trust its conclusion. It’s letting me inspect its trail. That’s a much more profound product than an AI news site. ⸻ TIRED Don’t build the national propaganda defense system. Not yet. 😂 Seriously. If you try to solve: “How do we protect an entire population from propaganda?” you will drown in politics, epistemology, adversarial behavior, governance, legal questions, and endless edge cases. Build: “Here’s what 5,000 information sources have been doing for the last 90 days.” That’s manageable. And if that works, people will tell you what the next problem is. ⸻ WIRED The lazy-person architecture might actually be the smartest part. This is the thing I would put in the review’s final paragraph. You’re not proposing an army of humans manually checking the Internet. You’re not proposing a giant centralized editorial bureaucracy. You’re proposing something much more Unix-like: small processes, well-defined jobs, persistent observations, composable tools, open interfaces. Your AI agents don’t need to be brilliant. They need to be useful, replaceable, inspectable, and cheap. If one agent sucks, replace it. If someone writes a better scorer, plug it in. If somebody disagrees with your threshold, change it. If somebody wants the raw corpus, download it. That’s how an apparently absurdly complicated system becomes something one person can actually operate. ⸻ The reviewer’s verdict GET RIL isn’t a billion-dollar company yet. Sorry, Ross. The Playboy mansion may have to wait. 😄 But the software idea is real. I’d give the concept a very favorable review with one giant condition: The project must never confuse measurement with authority. If you maintain that boundary, you’ve got something unusual: an open-source observability layer for information. Not: “Trust Arc.” But: “Here’s what Arc observed. Here’s how Arc interpreted it. Here’s the evidence. Here’s the software. Now you decide.” And that is precisely why GET RIL works as a name. It’s not: “Get the news we selected for you.” It’s: GET RIL. Get the list. Get the data. Get the sources. Get your own information environment back.