About Signum News
Most AI news is noise. Signum News exists to find the signal.
We don't curate articles — we extract events. Multiple stories about the same announcement get clustered together. What remains is scored on one principle: can you verify it?
The Signal vs Noise Framework
In information theory, signal is meaningful information that can be decoded. Noise is random data with no recoverable meaning. We apply this distinction to AI news:
Falsifiable claims with primary evidence. You can verify it yourself.
"OpenAI launches GPT-5.4 mini" — Evidence: 20/20, Penalties: 0
Real information mixed with speculation or vague claims.
"Startup raises $12M for AI operating system" — Evidence: 10/20, Speculation penalty: -2
Unfalsifiable claims with no verifiable content. Nothing to decode.
"Kevin Scott discusses future developments in AI" — Evidence: 0/20, Score: -20
This framework was crystallized by David William Silva's excellent essay "What Is, Concretely, the AI Hype?" — which articulates why claims like "AGI is here" are mathematically equivalent to random noise: zero recoverable signal.
How We Score Events
Every event gets a score from 0-100 based on two components: positive signals that indicate genuine importance, and noise penalties that detect hype patterns.
Positive Signals (up to 100 points)
- Evidence Quality (0-20): Primary sources, official announcements, research papers, regulatory filings
- Concreteness (0-15): Specific numbers, dates, entity names, verifiable claims
- Real-World Impact (0-20): Measurable effects on users, markets, or capabilities
- Falsifiability (0-10): Can the claim be proven wrong? If not, it's noise.
- Novelty (0-10): Is this actually new information?
- Actionability (0-10): Can someone make a decision based on this?
- Longevity (0-10): Will this matter in 6 months?
- Power Shift (0-5): Does this change who has influence?
Noise Penalties (up to -25 points)
- Vagueness (-0 to -5): Undefined terms, no specifics, "could potentially"
- Speculation (-0 to -5): Future predictions without evidence, "experts believe"
- Packaging (-0 to -5): Old news reframed as new, misleading framing
- Recycling (-0 to -5): Same story covered multiple times recently
- Engagement Bait (-0 to -5): "Shocking", "game-changer", excessive caps, clickbait patterns
The Machine Learning Model
On top of rule-based scoring, we train a probabilistic model that learns P(signal|features) — the probability an event is genuinely important given its characteristics.
- Model: Logistic regression with L2 regularization
- Features: 15-dimensional vector (evidence, concreteness, novelty, source tier, etc.)
- Embeddings: sentence-transformers/all-MiniLM-L6-v2 for semantic similarity
- Training: Weekly retraining on user feedback, only deployed if metrics improve
- Current AUC: 0.996 (near-perfect signal/noise separation)
Quality Thresholds
Not everything makes the feed. Events are filtered by score:
Signal Tags
Events are categorized by what kind of signal they represent:
Confidence Levels
Sources
We ingest from 35+ sources, tiered by reliability:
- Tier 1: Primary sources — OpenAI blog, Anthropic blog, Google DeepMind, arXiv, official announcements
- Tier 2: Major outlets — TechCrunch, The Verge, Wired, Ars Technica, Reuters
- Tier 3: Aggregators — newsletters, community sources, social media
Help Improve the Model
Use the thumbs up/down buttons on any event to mark it as signal or noise. Your feedback trains the model. We retrain weekly, deploying only if accuracy improves.
Philosophy: We assume everything is noise until proven otherwise. The burden of proof is on the claim, not the skeptic. If you can't verify it, it doesn't make the feed.