A large explorable world without loading screens is a technical achievement with specific constraints, and those constraints explain a great deal about how such games feel.
Streaming
The core technique.
The world is divided into regions, and only those near the player are loaded into memory.
As the player moves, distant regions are unloaded and approaching ones loaded, continuously.
Which means the world can be far larger than memory, and it makes storage speed a determining factor in how fast the player can travel.
Fast travel between distant points requires a loading pause because the entire nearby world must be replaced.
Level of detail
Objects far from the player are rendered with simpler models and textures.
Which reduces the work per frame enormously, since most of what is visible is distant.
Transitions between detail levels are the source of the popping people notice, and blending techniques reduce it at a performance cost.
Automated systems generating detail levels have improved substantially, reducing the manual work this once required.
The persistence problem
What happens to things the player has affected when their region is unloaded.
Storing every change indefinitely is expensive and grows without bound.
Which is why many games reset the state of distant areas, and why dropped objects disappear and defeated enemies return.
Selective persistence — recording specific meaningful changes while discarding incidental ones — is the standard compromise and it is why the rules about what persists seem arbitrary.
Density and the emptiness problem
The design difficulty that follows from scale.
A large world requires enough content to reward exploration, and authored content is expensive.
Which produces the two failure modes — worlds that are empty, and worlds filled with repeated activities that become tedious.
Procedural generation addresses volume and struggles with meaning, since generated content lacks the intentionality that makes discovery satisfying.
Games that handle this well generally use generation for the connective terrain and authored content for the destinations.
Navigation
How a player knows where to go shapes the experience more than the world design does.
Marker-heavy approaches direct attention efficiently and reduce exploration to following an indicator.
Landmark-based approaches, where distinctive features guide movement, require deliberate world composition and produce a different kind of engagement.
Several recent games have offered both, which is a reasonable acknowledgement that players want different things.
Ambient life
Non-player characters and creatures populating the world.
Simulating them all continuously is impossible at scale, so they are generally spawned near the player and removed when distant.
Which is why the world can feel like it exists only where you are looking, and it is a direct consequence of the computational budget.
Games that simulate more persistently do so at reduced fidelity, tracking simplified state for distant entities.
Traversal
How the player moves determines how the world must be built.
Fast traversal — vehicles, flight, grappling — requires streaming to keep up and reduces the detail the player perceives.
Slow traversal allows density and risks tedium across distance.
Which is why traversal design and world scale are decided together, and why changing one late in development is so disruptive.
Terrain authoring
How the landscape itself is produced.
Heightmap-based terrain is efficient and cannot represent overhangs or caves, which require separate geometry.
Voxel approaches allow full three-dimensional shaping including caves, at higher memory and processing cost.
Most large worlds combine both — heightmap terrain with placed geometry for anything the heightmap cannot express.
Vegetation and scattering
Placing millions of plants by hand is impossible, so scattering systems distribute them by rules — slope, altitude, proximity to water, biome definition.
Artists then adjust locally where it matters, which is the standard division between procedural bulk and authored detail.
Rendering that much vegetation efficiently requires instancing, where one model is drawn many times with varied transforms, and aggressive detail reduction at distance.
Testing
A large world cannot be manually tested exhaustively.
Automated systems that navigate the world looking for holes, unreachable areas and collision failures are standard, and they still miss things that players find within hours of release.
Which is why open world games ship with more geometry bugs than linear ones, and it is a scale problem rather than a care problem.
Weather and time of day
Dynamic systems multiply the lighting scenarios that must be authored and tested.
Which is why some games fix time of day in specific areas, and why night in an open world frequently looks less considered than day.
Loading and storage
Solid state storage changed what streaming systems can assume, allowing higher-detail assets to be loaded faster and closer to the player.
Which has affected world design directly, permitting faster traversal and denser environments than mechanical drives supported.
Games supporting both must design for the slower case, which constrains what the faster one can deliver.
Map design
Landmarks visible from a distance guide movement without markers, and composing sightlines so that something interesting is always visible is deliberate rather than incidental.