AI Signals

An AI signal is the smallest possible observation of change in the AI industry: a single event, dated, sourced to a primary URL, and tagged with the entities it involves. A model release is a signal. A benchmark score is a signal. A regulatory document, a pricing-page diff, a senior hire, a paper, a 10-Q disclosure — each is one signal.

1. LAYER 01

   Signals

   One event, one record. Atomic, dated, sourced. The raw material.

2. LAYER 02

   Trends

   Clusters of signals with stable density and positive velocity. The pattern.

3. LAYER 03

   Intelligence

   Briefings, predictions, and decisions made on signal-backed trends. The output.

Each signal Steek captures is assigned exactly one type. Type determines weight and how the signal contributes to a candidate trend's velocity score. The full taxonomy is documented in The AI Signals Taxonomy.

• Release

  A model, product, or feature shipped to the public.

  e.g. Anthropic releases Claude 3.7 Sonnet.

• Benchmark

  A measured score on a recognized evaluation.

  e.g. GPT-5 reaches 85.6% on SWE-bench Verified.

• Research

  A paper, preprint, or technical report from a credible source.

  e.g. New paper on inference-time scaling published on arXiv.

• Pricing

  A change in published API or subscription pricing.

  e.g. Provider X cuts input tokens 60%.

• Enterprise

  A procurement, partnership, or earnings disclosure indicating production use.

  e.g. Fortune 500 names AI line item in 10-Q.

• Hiring

  A senior hire or org-design change at a frontier lab or buyer.

  e.g. New head of AI agents named at major SaaS vendor.

• Policy

  A regulatory document, EO, or law affecting AI development or deployment.

  e.g. EU AI Act GPAI obligations begin to apply.

• Hardware

  A new accelerator, datacenter, or interconnect announcement.

  e.g. New 4nm inference chip with 1,500 TOPS.

• Open-source

  A new permissively licensed model, dataset, or tool.

  e.g. Frontier-quality open-weight model released.

• Distribution

  A platform shift that moves models to new surfaces.

  e.g. Browser ships native model runtime.

• Demo

  A credible demonstration of a previously unverified capability.

  e.g. Computer-use agent completes 30-step office task on video.

• Lawsuit

  A filing or ruling materially affecting AI rights, data, or liability.

  e.g. Copyright infringement suit filed against frontier lab.

• Security

  A discovered vulnerability, jailbreak class, or incident.

  e.g. New prompt-injection class disclosed for browsing agents.

• Standards

  A protocol, format, or interoperability spec gaining adoption.

  e.g. MCP adopted by major IDE.

• The AI Signals Report: How AI Trends Form

  The full Steek methodology — from raw signal capture to promoted trend.

• The AI Signals Taxonomy

  Definitions, weights, and examples for all 14 signal types.

• The AI Trend Velocity Model

  How signals translate to a trend-velocity score.

• Live Trend Index

  See where today's signals are pointing.

Frequently Asked Questions

What is an AI signal?

An AI signal is the smallest possible observation of change in the AI industry: one event, dated, sourced to a primary URL, and tagged with the entities involved. Examples include a single model release, a benchmark score posted, a regulatory document published, a pricing-page diff, or a Form 10-Q disclosure mentioning AI revenue.

How are AI signals different from AI news?

News is a written summary of an event after the fact. A signal is the structured, machine-readable record of the event itself, captured at the source. One news article may reference dozens of signals; one signal may produce zero news articles.

How do signals become trends?

Signals are clustered by entity, topic, and time-window co-occurrence into trend candidates. A candidate is promoted to a tracked trend when its signal density crosses a stability threshold and its velocity stays positive across two consecutive observation windows.

How many signal types does Steek track?

Fourteen primary types organized in three tiers: capability signals (release, benchmark, research), commercial signals (pricing, enterprise, distribution, hiring), and structural signals (policy, hardware, open-source, demo, lawsuit, security, standards). The full taxonomy is documented separately.

Why is signal-level data better than top-down trend reports?

Top-down reports compress thousands of observations into a single narrative; the compression is where bias enters. Signal-level data is auditable end-to-end: any claim Steek makes can be traced to the underlying events, with timestamps and source URLs intact.