Weak signal
A pattern in raw text — names you keep mentioning, energy that's dropped, a half-thought about a sister selling in Brentwood — that a structured database structurally cannot surface. A foundation model reading the actual messages catches it.
What it does (the operator translation)
Andrej Karpathy framed this on X in 2025 inside his Software 3.0 thread — English is the new programming language, the model is the runtime. The implication for a working agent: you don't structure the data first and query it after. You hand the model the mess and ask the question.
A weak signal is what the dashboard misses by definition. Maria's Old Hickory listing is 15 days in and her replies have gotten shorter. The Andersons keep showing up in your message threads but you've never actually called them. A buyer at the open house mentioned her sister is selling — never made it into a CRM field. None of that lives in a status column. All of it lives in the raw text you already paid to capture.
Hand 14 days of messages to ChatGPT or Claude with the prompt "what's the weak-signal flag — someone I keep mentioning but never call" and the model returns the neglected lead the dashboard structurally couldn't see.
Why a working REALTOR cares (the breakpoint)
For the NAR-median 12-deal agent running a sphere-driven pipeline, weak signals are most of the actual revenue. Past clients, sphere mentions, kitchen-table comments — the stuff that drives 66% of seller-found-agent activity. A predictive lead-gen tool can't see them. A foundation model on your sphere CSV can. That's the unfair advantage on owned data.
What this is NOT (the category-flip)
A weak signal is NOT a lead score. Lead scoring ranks structured records. Weak-signal detection reads the unstructured notes between the records. Different layer of the stack.
Related terms
Foundation model vs CRM · Lead scoring · Context card · Retrieval-augmented generation
Where this comes up in The Listing Machine
The weak-signal frame anchors the CRM pillar and the Lead Generation pillar. The Tuesday morning prompt is the operator translation — read-time intelligence on raw text, $20 a month, no structured data entry required.