AI Trend Radar

1,417 source-balanced items across 44.8 days from 230 domains · 1,424 eligible before balancing · largest source 19.8%

Narratives (themes) are extracted from our continuously monitored AI-news and community stream — including public Reddit Atom feeds that require no Reddit API credentials — by an LLM classifier and normalized. One grounded mention creates an admin-review candidate; two recent grounded mentions plus two grounded source domains across the window make a phrase emerging while it has at most two prior items; and three items make it established. Candidates stay off the public radar until they reach an admitted stage within the 45-day window. Explicitly named models are tracked as a separate entity dimension so a release name is not lost inside a broader narrative: two recent independent mentions qualify, while one mention qualifies only when a first-party source explicitly reports a launch. Each narrative is also classified against a stable topic-kind schema (model, technique, infrastructure, market/capital, policy…) by majority vote across its items — the schema's categories stay fixed over time so kind-level trends remain comparable across snapshots, and the seven quick filters group those kinds for scanning. Each news item additionally carries an event type (launch, funding, acquisition, adoption…), shown as the blip's event mix. Individual products and use-case types otherwise appear only when they dominate the conversation (at least 12 items with 6+ in the last 14 days) — shown as diamonds. A vendor's own newsroom never counts toward that vendor's dominance: only independent coverage does, and any own-channel share is disclosed on the blip.

Catalog links are semantic matches, not factual corroboration. A repeated match means articles on at least two source domains resemble the same case; domains may share ownership or republish one story. The rings represent many (5+), several (2-4), one, or no repeatedly matched cases. Below 6 independent news items, a narrative remains in the outer ring even when matches exist; this is labeled Coverage limited. Neither ring position nor confidence verifies deployment maturity, outcomes, or suitability for adoption.

Daily source-domain caps limit feed imbalance. Vendor-published coverage is excluded from independent counts for models and other non-company topics; company topics exclude that company's own channels. Trend strength combines velocity, acceleration, novelty, source diversity, persistence, coherence, author authority, case-match breadth, and strategic relevance. Missing factors are excluded and weights renormalized. Confidence summarizes volume, source breadth, temporal coverage, independent coverage, repeated matches (10%), and completeness. These are heuristic scores, not calibrated probabilities. The weekly burst signal remains experimental and does not affect ranking. Regression benchmarks use synthetic cases; real-world detection quality has not yet been calibrated.

Unprofiled authors have weight 1; curated experts have weight 2 within their configured expertise and 1.1 otherwise. Employer alone grants no boost. Historical comparisons require a compatible scoring and corpus model. A dashed halo marks narratives new versus the previous daily snapshot.