Feed
See what deserves attention.
Browse recent, relevant stories with a short summary and topic tags. Filter or search, then open the original article or its HN discussion. No model runs when you load the page.
The HN hive mind, made queryable.
Eieye narrows Hacker News to the AI stories people engage with, preserves the articles and conversations around them, and makes that evidence useful through a Feed, a research Chat, and four weekly briefings.
01
From submission to saved story
- 01
Wait for a signal
Stories are considered after two days. Only submissions with at least 20 points or 10 comments enter the pipeline.
- 02
Collect the context
The app keeps the HN story, its complete comment tree, and the readable text from the linked article.
- 03
Find the AI stories
AI must be central to the story itself or a sustained theme across several independent parts of its HN discussion. Isolated mentions do not count. Accepted stories receive topics and a short Feed summary.
- 04
Make it researchable
Stories, article passages, discussion branches, and comments are indexed for keyword and meaning-based search, while the original source text stays separate. A story's first successful index time also fixes its weekly publication window.
02
Three ways into the conversation
Reports
Follow the week in four dimensions.
Read Friday briefings on models, agentic use cases, agent harnesses and runtimes, and enterprise AI usage. Each edition links source evidence beside relevant claims and remains in the archive.
Chat
Research the conversation.
Ask a question in plain language. The agent searches across HN stories, reads promising articles and their discussions, then synthesises an answer with checked links beside its claims.
03
How a weekly report is made
- Freeze the weekly inventory
- Search within the captured stories
- Read useful articles and discussions
- Brief the key stories
- Resolve source links
04
How Chat builds an answer
- Map the question
- Find and rank story leads
- Read the articles
- Inspect the discussions
- Answer with checked citations
Chat distinguishes article claims from HN opinion, includes meaningful disagreement, and says when the available material is too thin. It does not force every question into the same answer shape.
05
Design decisions
The model writes the weekly briefing
Each report searches only within the week's captured stories, reads the useful articles and discussions, and briefs the key stories with both sources together. A quiet week stays short; there is no story quota or per-story checklist.
Engagement is the first filter
Points and comments are HN's strongest signal that a story deserves attention. They decide what is processed and strongly influence research ranking.
Comments are first-class evidence
The discussion is not an attachment to the article. Comment ancestry and replies are preserved so opinions keep their conversational meaning.
Discovery is not evidence
Generated summaries and search records can lead the agent to a story, but every substantive claim must come from article text or comments it inspected for that answer.
Coverage before convenience
For broad questions, Chat works across the relevant stories and question facets. When it reads an article with an available HN discussion, it must inspect that discussion too.
One bounded research agent
Chat uses one agent with three controlled research tools, up to 20 research calls, and a bounded evidence budget. It cannot browse the web or query the database directly.
Citations are checked
The agent cites source identifiers, not URLs it invents. The server accepts only sources returned during the current research and turns those citations into links.
One bad item does not stop the day
After limited retries, a failed story is recorded so it can be retried without blocking the rest of the daily update. Failures that would make progress unsafe still stop the run.
Daily is current enough
This is a considered view of the HN conversation, not a breaking-news wire. The saved stories update once each day.