FAQ
A community-ranked intelligence board for people building with large language models. Think Hacker News, but focused on LLM tools, models, papers,…
What is LLM Atlas?
A community-ranked intelligence board for people building with large language models. Think Hacker News, but focused on LLM tools, models, papers, infrastructure, and the companies behind them. Anyone can submit, vote, comment, and curate dashboards. No account required.
How are items ranked?
The Top feed uses a time-decay ranking formula: Score = (votes + comments x 0.35 + saves x 0.75 + dashboardAdds x 1.2 - flags x 2) / (ageHours + 2) ^ 1.25 New items get a boost from the denominator being small. Sustained engagement (votes, comments, saves) keeps items visible over time. Items with no engagement decay naturally. The New feed is simply reverse-chronological. Category feeds (Models, Tools, Papers, etc.) filter by item type, then apply the same ranking.
How does voting work?
Votes are server-persisted. When you upvote, the count is sent to the server and stored in data/votes.json. All users see the same aggregate count. Your browser tracks which items you've voted on (localStorage) so the UI shows your votes. You can toggle a vote off by clicking the arrow again. There are no downvotes on items. If something doesn't belong, use the Report option in the ... menu.
How does the trust score work?
Every visitor has a trust score (0-100) tracked by IP address. You start at 50 (Normal tier). Your score goes up when you: pace your actions naturally, write substantive comments, submit content that isn't rejected. Your score goes down when you: submit duplicate content, post too-short or all-caps comments, burst-fire actions, trigger spam patterns. Tiers: - Trusted (70-100): generous limits, 15 votes/min, 40 comments/hr - Normal (40-69): standard limits, 10 votes/min, 20 comments/hr - Suspect (20-39): restricted, 4 votes/min, 10 comments/hr - Restricted (0-19): heavily throttled, 2 votes/min, 5 comments/hr Cooldowns escalate: 30s, 2min, 10min, 1hr. They reset after 30 minutes of no incidents.
What categories are available?
Top - Ranked by community score with time decay New - Latest submissions Models - Model launches, benchmarks, evals Tools - SDKs, platforms, infrastructure Papers - Research with linked repos and discussion Repos - GitHub-heavy feed Benchmarks - Evals, leaderboards, comparisons Show - Demos people built Dashboards - Community-curated collections Chats - AI chat interfaces and assistants News - Dedicated AI news and industry updates Companies - AI companies, funding, and startups Ask - Questions and discussion Categories are driven by data files, not hardcoded. New categories can be added by editing categories.json.
How do I submit an item?
Click 'submit' in the top bar. Enter a URL and select a type. The system checks the URL for duplicates and scrapes OpenGraph metadata (title, description, image) to pre-fill the form. You can edit everything before submitting. For text-only posts (Ask, Discussion), leave the URL blank and select the appropriate type. Keyboard shortcut: Cmd+Enter submits the form.
What happens when a submission is rejected?
If your URL already exists on the site, you'll see the existing item and can upvote it instead. You can also submit anyway if you believe your framing adds value. Rejected attempts are tracked and visible on your profile's Rejected tab. This is intentional — it shows both sides: what was accepted and what was caught by dedup.
What are dashboards?
Community-curated collections of items around a theme. Anyone can create a dashboard, and others can follow or fork it. To add items to a dashboard, use the ... menu on any item in the feed and select 'Add to dashboard'. Dashboards are how the community organises knowledge beyond simple ranked lists.
Can I delete my submissions and comments?
Yes. Use the ... menu on any item or comment you submitted, or use the delete button on your profile page. Deleting a comment removes all replies to it. Deletion is permanent. Deleting a submission removes it from all feeds, the URL index, and vote counts.
What do the ... menus do?
The three-dot menu on items and comments gives you: - Save / Unsave: bookmark for your Mine feed - Share link: copies a permalink (uses native share on mobile) - Add to dashboard: add the item to one of your dashboards - Report: flag for spam, harassment, misinformation, off-topic, self-promotion - Block user: hide all their content from your view - Delete: remove your own content (only visible on your items)
What is the Mine feed?
Toggle between Community and Mine in the header. The Mine feed shows only items you've interacted with: voted on, saved, added to dashboards, or in dashboards you follow. It's your personal filtered view of LLM Atlas.
How do comments work?
Comments are server-persisted and threaded (Reddit-style). You can reply at any depth, collapse threads, and use color-coded depth lines to follow conversations. Keyboard shortcuts: - c: focus comment box - Cmd+Enter: submit comment or reply - Escape: cancel reply Comments are quality-checked by the behaviour engine: minimum 3 characters, no all-caps, no character spam, no duplicates.
Where is my data stored?
Server-side: votes, comments, submissions, reports, and rejections are persisted on the server in JSON files. Client-side (localStorage): which items you voted on, your saves, dashboards, follows, blocked users, theme preference, and filter settings. No account or registration is required. No tracking cookies. No personal data collection.
What keyboard shortcuts are available?
Press ? anywhere to see the full shortcuts panel. Global: / or Cmd+K (search), Escape (blur/close), ? (shortcuts help) Feed: j/k (navigate items), Enter or o (open), u (upvote), s (save), arrows (pages) Comments: c (focus comment box), Cmd+Enter (submit), Escape (cancel reply) Submit: Cmd+Enter (submit form)
Is there a dark and light mode?
Yes. Click the sun/moon icon in the top bar to toggle. Your preference is saved in localStorage.
Are reposts allowed?
The dedup system prevents exact URL matches (normalised across protocols, tracking params, www, trailing slashes, and site-specific patterns like arxiv PDF vs abstract URLs). If a story has not had significant attention, submitting it with new context or commentary is acceptable. Don't game the dedup system with URL variants.