The Knowledge Gap
This is the real-world consequence of the simulation in The Context Window. The 6 additive rules were fundamentally unpredictable from temporal history alone because the information they needed was spatial, not temporal. The same principle applies to knowledge in the physical world.
You are driving to the Tallinn ferry terminal. There are 8 lines at Terminal A β 4 for Viking Line, 4 for EckerΓΆ Line. No website, no sign, no app tells you which ones are open today. You pick a lane, queue up, and hope. If you chose wrong, you wasted the queue and have to start over. Only people who have driven through that week know which lines are actually open.
(All 4 were open this time. But not at the same exact time. Some people got hesitant, switched lines, and ended up worse off because the lanes opened just two minutes apart. The knowledge gap is not just about knowing which lane β it is about knowing when.)
This is not a failure of documentation. It is a fundamental class of knowledge that exists only in embodied, situated experience.
Situated Operational Knowledge
Call it situated operational knowledge. It is:
- Ephemeral. It changes by the hour. The lane that was open this morning may be closed this afternoon. The knowledge has a half-life measured in hours, not years.
- Location-bound. It exists only in the specific place at the specific time. The knowledge of which lane is open at the Tallinn ferry terminal is useless at the Helsinki terminal.
- Context-dependent. It depends on factors that are not captured in any documentation. The lane that is open depends on the weather, the time of day, the phase of the moon, the whim of the terminal operator.
- Unrecorded by design. No one writes it down because it changes too fast. The documentation would be obsolete before it was published. The knowledge exists only in the minds of people who have been there recently.
This is not a niche category. Every physical place has thousands of such gaps.
The Scale
Which cafe has working outlets? Which trail is muddy after the rain? Which shortcut through the construction site is passable? Which bus stop is closed for construction this week? Which restaurant has a quiet corner for a phone call? Which park bench has the best view of the sunset? Which supermarket aisle has the item you need when the shelf label is wrong? Which door in the office building is propped open today? Which public toilet is actually clean? Which staircase in the metro station is closed for repairs? Which self-checkout machine is broken but not marked? Which charging station for electric cars actually works and is not ICEd? Which path through the park is lit at night? Which pedestrian crossing was repainted and is now safer? Which bus driver lets you board with a bike and which one doesn't?
Some of these are partially captured by phone geo-location and user feedback. Google Maps shows how busy a cafe is. Waze knows about construction. But most are not. The working outlet in the corner of the cafe is invisible to every sensor. The propped-open door is too ephemeral to map. The broken self-checkout machine will not appear in any dataset until a human walks up to it and discovers it. The knowledge exists only in the moment of encounter.
Each of these is a piece of knowledge that is systematically invisible to web search. It is not on Wikipedia. It is not on Yelp. It is not on Google Maps. It is not on any website, any database, any archive. It exists only in the embodied experience of people who have been there.
The scale is staggering. There is no way to count it β that is the point. The knowledge is invisible to measurement by design. But every physical place generates thousands of such gaps every day. Every cafe, every bus stop, every ferry terminal, every park bench. The gaps are not exceptions. They are the rule.
A Large Language Model (LLM) cannot reach them. Not because the technology is not good enough. Not because the training data is not large enough. Not because the algorithms are not sophisticated enough. But because the knowledge does not exist in text. It exists in bodies, in places, in moments.
The Connection to the Simulation
The context window simulation showed that 6 additive cellular automaton (CA) rules are fundamentally unpredictable from temporal history alone. The Markov predictor β the simplest possible learning algorithm, which works by counting patterns in past data β could not learn them because the information it needed was spatial (the states of neighboring cells) rather than temporal (the history of the center column). Even with a window of 100 time steps, it barely beat random guessing.
The knowledge gap is the same phenomenon at a different scale. An LLM trained on text is a temporal predictor of the same kind. It sees a fixed history of tokens and predicts the next one. If the knowledge it needs has never been written down β if it exists only in the physical world at the moment of use β the LLM cannot access it, no matter how large its context window or how much training data it has consumed.
The gap is not about scale. It is about modality. Temporal history cannot substitute for spatial presence.
The Philosophical Background
This idea has a name: tacit knowledge, a term coined by the philosopher Michael Polanyi in his 1966 book The Tacit Dimension. Polanyi's famous formulation: "We can know more than we can tell." He was describing the knowledge that cannot be fully articulated β the knowledge of how to ride a bicycle, how to recognize a face, how to navigate a familiar city. It is knowledge that lives in the body and the context, not in language.
Polanyi was writing about humans, but the principle applies even more strongly to machines. An LLM can only know what has been told. It has no body, no physical presence, no access to the world outside its training data. The tacit dimension of human knowledge is doubly inaccessible to it: first because the knowledge is not written down, and second because even if it were, the LLM would lack the embodied context to understand it.
The Knowledge Gap
The gap between temporal and spatial prediction is the knowledge gap β the information that exists in the world but not in the text. The simulation made this concrete: the spatial predictor (which could see the full CA state) achieved 100% accuracy on the additive rules, while the temporal predictor (which could only see history) never beat 55%. The same gap exists between an LLM and a human standing at the ferry terminal.
The gap is not a bug. It is a feature of how knowledge works. Most of what humans know is not written down. The vast majority of human knowledge is tacit, embodied, situated. It is the knowledge of how to ride a bicycle, how to read a room, how to navigate a ferry terminal. It is the knowledge that cannot be captured in a manual, cannot be encoded in a database, cannot be transmitted through text.
The Warning
The knowledge gap is a reminder that knowledge is not data. It is something that happens between people, in bodies, in places, in time.
The gap is not a problem to be solved. It is a feature to be respected. It is the reason that asking a local is better than reading a review. It is the reason that being there is better than reading about it. It is the reason that experience is not the same as information.
The gap is also a warning. If we mistake data for knowledge, we will build systems that are confident and wrong. We will trust the map more than the territory. We will forget that the most important knowledge is the knowledge that cannot be written down.
Next
What happens when an LLM is asked a question whose answer is not in its training data? It does not say "I don't know." It generates a plausible-sounding answer that is confidently wrong. This is not a bug. It is a feature of how the architecture works. And it has consequences that go far beyond the ferry terminal.
Coming next.
Code and data for this entry: codeberg.org/halhour/knowledge-gap