# A Frozen Knowledge Base Entrenches Whoever Was Prominent at Cutoff ## Claim When a model's knowledge is fixed at a training cutoff instead of continuously updated, entities that rise to prominence after that date cannot be learned into the model at all and are visible only through retrieval, which structurally favors whoever was already established. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Applies to models deliberately kept on a static or infrequently-updated knowledge base, where the lab supplies current information via retrieval/search rather than retraining the weights. - The entrenchment is about unretrieved model output specifically — a model's own "knowledge," not what a well-built retrieval layer can surface. Good retrieval can offset the effect but doesn't eliminate the underlying asymmetry: incumbents are visible two ways (weights + retrieval), newcomers only one way (retrieval). - Doesn't establish how large the effect is in practice — the source raises it without measurement. - Breaks where a model is frequently retrained/refreshed rather than frozen, or where retrieval is comprehensive and equally weighted regardless of prominence at cutoff. ## Evidence Christopher Penn cites Gemini Flash 3.5 as retaining a January 2025 knowledge cutoff as deliberate rather than lagged — evidence that some labs have stopped optimizing models to *know* and instead optimize them to *do*, supplying current knowledge through search and retrieval at query time. He names the consequence directly, then moves past it without developing it: established players get reinforced by the model's frozen knowledge itself, an issue he expects the industry to address eventually but "not... anytime soon." ## Links - [[AI & the Future of Knowledge Work MOC]] — a structural bias in whose work becomes visible once models mediate discovery - [[3. Resources/Permanent Notes/Practical Obscurity Is an Economic Property Not a Technical One]] — because contrasts; that claim is about cost collapse widening what's findable, this one is about a frozen snapshot narrowing what's learnable — two forces pulling in opposite directions on the same discoverability question - [[3. Resources/Permanent Notes/Retrieval Quality Is Set by the Unit of Knowledge Not the Retrieval Algorithm]] — because applies-in; if retrieval is the only channel available to newcomers, the quality of that retrieval layer (not just its existence) determines whether the entrenchment gets offset or compounded - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — because applies-in; the do-vs-know training strategy is a resource-allocation decision with this discoverability consequence attached ## Tests / falsifiers - Compare how often a model surfaces pre-cutoff versus post-cutoff entities in unretrieved (no search/RAG) responses on comparable prompts — a null difference would weaken the claim. - Measure whether retrieval-augmented responses fully equalize visibility between pre- and post-cutoff entities, or whether pre-cutoff entities still get referenced more even when retrieval surfaces both — this would show whether retrieval genuinely offsets the effect or only partially. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/You Ask, I Answer - Handling AI Context Drift]] — reports the Gemini Flash 3.5 cutoff claim and the entrenchment consequence, raised but not developed by the source.