We Built a Live AI News Terminal — Labs, Papers, GitHub & HN, All in One Place

Most AI news lives in 14 different places at once. Anthropic's blog, OpenAI's news feed, Google DeepMind, Meta AI, Mistral, xAI. arXiv cs.LG and cs.CL drop dozens of papers a day. Trending repos on GitHub, the front page on HuggingFace, AI threads bubbling on Hacker News. And the writers everyone reads — Simon Willison, Latent Space, Interconnects.

If you want to keep up, you're either tab-juggling or you miss things. So we put them all in one place: probbrain.com/news.

What it is

A live, dense, terminal-style page that pulls from every source above, dedupes by URL, and renders the headline + source + how-fresh in a single scrollable feed. Every row links straight to the original article. Filter by category — lab, paper, code, writing, community — or hit a source pill at the top to filter by publisher. The page auto-refreshes every 5 minutes; the fetcher runs every 15 minutes via a local timer.

Sources, by category

• 🧠 Labs — Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, xAI
• 📄 Papers — arXiv cs.LG, arXiv cs.CL (one-line LLM summaries on every paper)
• 💻 Code — GitHub trending (AI-keyword filtered), HuggingFace top trending models
• 🔥 Community — Hacker News front page (score > 100, AI keywords)
• ✍️ Writing — Simon Willison, Latent Space, Nathan Lambert's Interconnects

Under the hood

RSS where it exists, sitemap parsing for the labs that don't publish a feed, the official HN firebase API, and HTML scrapes only as a last resort (one fetch per run, polite). Every paper and trending item gets a one-sentence neutral summary via the local Claude Haiku CLI — cheap, fast, cached so we don't re-summarize anything we've already seen. Items are stamped in UTC ISO at ingest, displayed in your local time on render. Every item gets a stable 10-character SHA-1 anchor so a Telegram link can deep-link straight to the row on the website.

The whole thing is a few hundred lines of plain Node, one small dependency (rss-parser), and a systemd user timer. The static page lives on Vercel and reads a sibling JSON file the fetcher rewrites every 15 minutes.

The Telegram channel

Every new item also posts to @ProbBrain on Telegram. Headline plus a clickable link back to the page (deep-linked to that exact row). Posts are throttled to one per second, oldest first, no link previews — clean and fast. We silently seeded the backlog so the channel only fires on stuff that's genuinely new from launch onwards.

Design choices we're proud of

• Live source ticker — each pill shows that source's 24h count and softly pulses when something fresh lands. Click a pill to filter the feed to it.
• Lazy render — only 15 rows in the DOM on first paint; an IntersectionObserver pulls 15 more as you scroll. Page loads fast even with hundreds of items.
• Honest timestamps — ingest UTC, display local. No "1 hour ago" on something published last week because some feed didn't include a date.

Try it

👉 probbrain.com/news

👉 @ProbBrain on Telegram for the live feed

Feedback welcome. We're building ProbBrain in public — the dashboard, arbitrage scanner, Kelly calculator, methodology, and now the news terminal all live at probbrain.com, with source on GitHub.