HN Debrief

The Low-Tech AI of Elden Ring

  • AI
  • Gaming
  • Programming
  • Developer Tools

The post argues that Elden Ring’s enemy behavior is driven by a compact, low-level decision system that pushes and pops actions instead of traversing a heavyweight tree every tick. It presents this as a pragmatic alternative to “real AI” and as part of why FromSoftware can make enemies feel sharp without relying on anything like modern generative models. People familiar with game development landed on a blunter read: this is standard game AI. Several said the mechanism described is basically a behavior tree or a close cousin of one, just implemented with a goal stack or continuation-style execution rather than always restarting from the root. The article was even updated to acknowledge that overlap.

If you build interactive systems, the useful lesson is not that Elden Ring found a novel AI architecture. It is that strong behavior can come from deterministic, inspectable logic when it is paired with level design, animation, and careful authoring.

Discussion mood

Mostly positive about the article’s core idea but impatient with the framing. Game developers and experienced players saw the system as familiar behavior-tree-style game AI, not a breakthrough, and they appreciated the reminder that clever design and execution still beat hype-heavy notions of AI.

Key insights

  1. 01

    Goal stack is behavior tree adjacent

    The article’s “stack of goals” description lands much closer to mainstream behavior trees than it first admits. The interesting distinction is implementation style. Instead of re-walking a whole tree from the root every update, a system can keep its place like a call stack and resume a subtree, which preserves the same authoring model while reducing overhead and complexity in the runtime.

    If you are evaluating AI architectures, focus less on the diagram shape and more on execution semantics. You may be able to keep designer-friendly behavior trees while changing the runtime model to get the performance and debuggability you want.

      Attribution:
    • drunken_thor #1
    • AndrewDucker #1
    • andrewflnr #1
  2. 02

    Simple verbs hide expensive engine work

    A top-level action like “attack” sounds trivial until you unpack what it invokes. In practice it bundles navigation in 3D space, facing, animation transitions, hit timing, and reactions to blocks or parries. Boss behavior also reads as richer because moves are authored to chain into specific follow-ups, and because aggro and path-following are tuned against hostile level layouts. The apparent intelligence sits in the integration layer, not just the decision policy.

    When reviewing gameplay AI, inspect the whole action pipeline instead of the planner alone. A modest controller can feel excellent if movement, animation, and encounter geometry are doing their share of the work.

      Attribution:
    • nitwit005 #1
    • 8note #1
    • havblue #1
  3. 03

    Quest jank supports FromSoftware’s world design

    The obscure NPC progression in Elden Ring is not just sloppy scripting bolted onto the combat system. It reinforces a world where characters do not wait around as quest kiosks and where some storylines are intentionally missable or unresolved for many players. One commenter tied Hyetta’s quest directly to the game’s broader narrative structure, arguing that the friction is part of how FromSoftware makes side characters feel independent rather than utility nodes.

    If your product leans on immersion or atmosphere, removing every bit of friction can flatten the experience. Some rough edges are worth keeping when they reinforce the fiction users are supposed to feel.

      Attribution:
    • vector_spaces #1
    • bigyabai #1
    • bee_rider #1
  4. 04

    Game AI still counts as AI

    The strongest pushback was against the idea that only learning systems deserve the label. In game development, behavior trees, state machines, and other symbolic systems have been called AI for decades. Commenters connected this to older “good old-fashioned AI” work and pointed out that once a technique becomes normal and useful, people often stop calling it AI at all. That is a naming drift, not a technical boundary.

    Be precise when you talk about AI in products and hiring. If you mean large language models or generative models, say that directly instead of erasing the older symbolic and rule-based systems your team may actually need.

      Attribution:
    • vector_spaces #1
    • badsectoracula #1
    • raincole #1
    • bitwize #1
    • empath75 #1

Against the grain

  1. 01

    Quest opacity is not just technical debt

    The reload-and-repeat NPC flow was framed by some as an engine limitation, but others pushed back hard. Sekiro shows FromSoftware can make character progression much more direct when it wants to, even if a few repeated interactions remain. That makes the obscurity look like a deliberate studio preference rather than a constraint they are stuck with.

    Do not assume awkward UX is always inherited from the stack. Sometimes the team is consciously preserving a style, which means improvement requires a product decision, not an engineering fix.

      Attribution:
    • tancop #1
    • mostlysimilar #1
    • SerpentJoe #1
  2. 02

    The performance argument is underspecified

    The post leans on the claim that its approach avoids decision-tree overhead, but commenters noted that behavior trees can be compiled or executed in multiple efficient ways already. Given that most gameplay cost often sits in native engine calls rather than in the control structure itself, the article does not really prove that the stack model is meaningfully faster.

    Treat architecture claims with suspicion when they are not paired with profiling data. Before rewriting control logic for speed, measure whether the real cost is in traversal, scripting, or the underlying gameplay systems.

      Attribution:
    • badsectoracula #1
    • BigTTYGothGF #1

In plain english

aggro ↩
Game shorthand for how enemies detect, target, and keep pursuing a player.
AI ↩
Artificial intelligence, a broad term for systems that perform tasks associated with decision-making, reasoning, or pattern recognition.
behavior tree ↩
A common game AI structure that organizes character decisions as a tree of conditions and actions.
NPC ↩
Non-player character, a game-controlled character rather than one controlled by a human player.
pathing ↩
The logic a game uses to move characters through the world toward a target or destination.
tick ↩
A single update step in a game loop where game logic is processed.

Reference links

Game AI papers and references

Game examples and mods

AI terminology and history

FromSoftware writing and translation