LLMgram is a curation layer. It doesn't produce news — it ingests public AI content from roughly 100+ primary configured sources (with an optional expansion pack), enriches weak excerpts with the dedicated LLMgram URL Reader, runs structured analysis, and ranks by source trust, evidence, recency, and usefulness. Everything on this page explains exactly how that happens, because if you can't see how a score is made you shouldn't trust it.
Data sources
Every public surface on llmgram pulls from a named upstream. Ranking is not “we picked what we like” — items arrive from feeds, APIs, curated registries, or scheduled scrapers. A few surfaces (Hardware Red Room today) are still curated seed graphs and are labeled as such.
| Surface | Upstream | Method | Cadence |
|---|---|---|---|
| AI Signal | ~105 primary RSS/scrape/X sources in fetch-rss.py (+ optional expansion pack) | Fetch + LLMgram URL Reader enrichment + local/Grok structured analysis | Ops cron (freshness target ≤6 h) |
| Use Cases | Hand-curated evidence registry | Human review against a public repo, documentation, demo, or reproducible workflow | On verified editorial update |
| Weekly | AI Signal snapshot | Editorial Top 10 digest from scored articles | Built with AI Signal critical sync |
| Git Signal | Curated GitHub repo list / API snapshot | Repo metadata + README + analysis | Ops sync script |
| AI Papers | ai-papers-data.json research graph | Scored/grouped paper index for the AI Papers surface | On dataset sync |
| Skills Signal | External skills snapshot → skills-signal-data.json | Sync + publish | Ops sync script |
| Compute Map | Epoch AI — AI Data Centers dataset | Compacted static JSON for the globe UI | On refresh script run |
| Hardware Red Room | Curated seed graph (AI Signal candidates planned) | Conservative source-backed seed MVP | On curated update |
| LLM Architectures | Sebastian Raschka LLM Architecture Gallery | Gallery scrape/builder + attribution | Suggested every 6 h |
| Company Radars | Labs registry + public profiles / evidence queue | Radar pipeline + roster verification | Rotation / varies by lab |
| AI Mercato | Labs registry + roster-backed affiliations | Verified people/movements graph | On rebuild |
| Wiki | AI Signal-derived entity graph | Seed/enrich from signal (full sync) | Scheduled full sync |
| Academy | Tutorial-like items filtered from AI Signal | Regex prefilter + optional Grok verification | With Academy builder / signal |
What "structured analysis" means
AI Signal uses a local analysis model by default, with Grok as fallback/legacy path when explicitly configured. Before analysis, weak or title-only inputs are enriched through llmgram-url-reader, the LLMgram-owned fork/adapter for URL extraction. Celeste URL Reader can inform the implementation, but the production contract is LLMgram's own POST /summarize service.
- Category — one of ~20 buckets (model release, paper, framework, infra, agent, safety, etc.). Lets users filter.
- Themes — up to 5 free-form keywords capturing what the item is about. Drives the search index.
- Audience — researcher / practitioner / both. Not every paper matters to every reader.
- Editorial analysis — signal, summary, context, critique, key takeaway, impact score, novelty, and warnings where extraction is partial.
How the signal score is prompted
The analysis prompt asks for a score based on a weighted mix of novelty (is this new information, or the Nth write-up of the same thing?), importance (does this change how practitioners work?), rigor (is there evidence, or is it hype?), and freshness (when was this published?). The weekly digest adds an extra editorial filter: unverified X leaks and speculative frontier-model rumors can remain searchable in AI Signal, but they are not allowed to dominate the canonical weekly issue.
0.85 + Rare. Genuinely important — models launches, breakthrough papers, major framework releases. These are the items you'd regret missing.
0.50 – 0.84 Useful context. Good reads, but skippable if you're pressed for time.
Below 0.50 Noise. Kept in the corpus for search, hidden from default views.
Limitations you should know
Grok is a language model. That means the scoring is opinionated and inconsistent across runs. We mitigate but don't eliminate this:
- The prompt is fixed and versioned. Changes are rare and noted in the changelog.
- Scores are rounded to 2 decimals to discourage false precision.
- Every item links to its raw source. If Grok miscategorized something, you can see the original in one click.
- Sampling bias exists. English-first sources dominate the feed. Labs outside the US/EU/China are underrepresented.
Refresh cadence
Different surfaces refresh at different rates. Exact crontab lines live on the ops host; the contracts below match the scripts and health checks in this repo.
- AI Signal / Weekly — cron-driven fetch + critical sync. Health treats datasets older than ~6 h as stale.
- Git Signal / Skills Signal / AI Papers — published when their sync/dataset jobs run; not a fixed public 2-hour promise.
- Compute Map / Hardware Red Room — refreshed when their build/seed scripts run (Hardware remains a curated seed MVP).
- Company Radars / AI Mercato / Wiki — rotation, rebuild, or scheduled full sync depending on the surface.
- LLM Architectures — gallery builder; suggested ops cadence every 6 hours.
- Academy — rebuilt from AI Signal tutorial-like items (optional Grok verify), not a static interview PDF.
Use Cases evidence gate
Use Cases turns selected capabilities into evergreen workflows. AI Signal and other sources may surface candidates, but no model can publish a case automatically. A public case requires a concrete outcome, explicit inputs and steps, at least one reachable evidence link, and a dated editorial verification.
- Evidence beats attention. A repository, official documentation, working demo, or reproducible process can support publication; likes and reposts cannot.
- Claims stay narrow. The card describes only what its linked evidence supports. Product announcements do not become performance claims.
- Verification expires. Links and claims are reviewed again after roughly 180 days; stale cases are rechecked or retired rather than silently refreshed by a model.
- AI Signal remains chronological. Use Cases is a separate curated registry, so rebuilding the news feed cannot erase or mutate the published workflows.
What we don't do
- We don't tell you what to think. The score is an opinion. Always click through to form your own.
- We don't re-write articles. Summaries are Grok-generated from the original; we don't paraphrase or re-publish content.
- We don't accept paid placement. Ranking is signal-based, never sponsored. If that ever changes, it'll be labeled clearly.
- We don't sell data or track you beyond basic analytics. No fingerprinting, no cross-site tracking.
Provenance & source code
LLMgram is built and operated at llmgram.app as a public lab notebook. Pipelines, HTML, and data are in private repos for now. Every content item links back to its original source — nothing is published without attribution.
If you find a miscategorized item, a stale score, or a bug in the scoring, ping @iamsupersocks on X/Twitter or @llmgram.
Changelog
- 2026-07 — v1.2 adds the Use Cases evidence gate: human-reviewed workflows, durable provenance, and stale-case rechecks.
- 2026-07 — v1.1 trust audit. Surface inventory, source counts, cadences, and Academy attribution aligned to live pipelines; retired Hermes Live / Claude Code Live rows removed.
- 2026-04 — v1.0 methodology page published. Signal score definition locked.