Charting a Course for the Next Decade of Gaming With AI

Overview

A panel discussion on how AI is reshaping game development across three broad axes:

1. Game creation — building content and workflows faster and better with AI.
2. Rendering and runtime improvements — for example, DLSS5 and related runtime optimizations.
3. Player-facing experiences — new kinds of immersion and gameplay enabled by agents, world models, and smarter NPCs.

Key themes and takeaways

• AI is becoming central across the game pipeline: design, tooling, asset creation, runtime behavior, and QA.
• No single paradigm dominates: traditional engines and world models will coexist. World models enable novel experiences and lower the barrier to experimentation, while classical systems remain valuable.
• Pragmatism matters: start with approaches that reduce risk and deliver player value. Use classical systems where they make sense and hybridize with ML/LLMs where they add unique value.
• Hands-on experimentation across teams is critical — letting developers play with LLMs/agents helps discover productive workflows.
• Upskilling and returning AI tools/tech to game teams (rather than centralizing all AI work) is important for adoption and long-term success.

Use cases and examples mentioned

• Rendering

  • DLSS5 (NVIDIA demoed recent advances).

• NPCs and in-game companions

  • Craftton’s PUBG Ally — on-PC stack (ASR → LLM → TTS) running locally for immersive companions.
  • ReLU Studio’s “Uncover the Smoking Gun” — early cloud LLM-driven detective NPCs (hallucination was acceptable given gameplay).
  • Enzo / ZOE — live simulation with LLM-driven NPC agents; players can mod character prompts.

• World models and generative interactive simulations

  • DART — interactive real-time world models (Oasis One demo: 10M users in two weeks).
  • World models used for rapid prototyping and democratizing development.

• Game AI & advisors

  • Creative Assembly (Total War series) exploring in-game AI advisors to improve onboarding and retention.

• Bots and ML-driven gameplay

  • Riot deploying ML bots in early PVE experiments; exploring imitation learning + RL for better tutorial/PVE bots.

• Infrastructure & tooling

  • Craftton building large training clusters (~1,000 GPUs) and training specialized speech-to-speech models for lower latency; focus on on-device inference.
  • DART running expensive training (scaling toward hundreds of millions/year in compute) while aiming to reduce inference costs.

• Evaluation

  • Creative Assembly built an evaluation system combining designer/QA “golden” questions with LM-as-judge metrics (usefulness, correctness, tone/personality).

Strategies, practical tips and recommendations

• Start small and practical

  • Put off-the-shelf LLMs and models in the hands of teams to discover useful workflows before investing in custom models.
  • Use Retrieval-Augmented Generation (RAG) and off-the-shelf inference where appropriate before training from scratch.

• Hybrid design is often the best near-term approach

  • Combine classical AI (behavior trees, utility systems) as “system 1” with LLM/agent “system 2” orchestrating or selecting scripted behaviors.
  • Execute LLM-decided high-level plans by triggering existing robust game scripts.

• Prioritize player value

  • AI features should improve the game for players or enable new experiences — avoid adding AI purely for novelty.

• Evaluation & control

  • Build evaluation suites with designer-curated test cases and metrics (usefulness, correctness, tone). Treat an LM as a judge in early testing.
  • Ensure explainability and designer control so players perceive fairness and predictable outcomes.

• Cost & infrastructure pragmatics

  • Start with cloud or rented GPU capacity for experimentation; benchmark costs and identify near-term productization paths (for example, internal dev tools) to justify compute spend.
  • On-device inference is valuable for latency and cost but may require smaller or specialized SLMs.

• Upskilling & adoption

  • Create internal training, safe spaces, and curated toolkits to lower friction for non-AI specialists.
  • Let central AI teams hand off prototypes and tooling to dev teams for long-term ownership and learning.

• User testing & iteration

  • Ship experimental features to learn; waiting for “perfect” will fall behind competitors.
  • Collect player feedback and iterate — responses can vary widely by audience.

Technical & infrastructure notes

• Model trends

  • Small language models (SLMs) are improving quickly; the gap to large cloud LMs is shrinking.
  • Emergent capabilities can appear with scale — certain tasks can suddenly become tractable above specific model sizes.

• On-device vs cloud

  • Local multi-model stacks (ASR → LLM → TTS) can run well on powerful PCs and are desirable for privacy and latency.
  • Cloud models speed prototyping but increase ongoing cost and ops complexity.

• Cost realities

  • Training world-scale models is expensive (tens–hundreds of millions per year in compute).
  • Inferencing, optimization, and engineering for efficiency are central technical challenges.

• Tooling and resources

  • Agents and automated summarization agents are already used internally to speed R&D.
  • GPU/accelerator supply (e.g., NVIDIA Blackwell family) can be a bottleneck; teams should plan purchases and scaling accordingly.

Risks and challenges

• Hallucinations

  • Acceptable in some game contexts (e.g., detective games) but dangerous in others; require design-aware mitigation.

• Difficulty tuning and fairness

  • Strong models may exploit unintended mechanics or produce unfun results; tuning to achieve fair and “just-barely-beatable” difficulty is difficult.

• Explainability and designer control

  • Designers need predictable, controllable outcomes — pure generative systems can frustrate that need.

• Organizational friction

  • Central AI teams can be ineffective unless tightly integrated with game teams. Ownership and capabilities should sit with dev teams where possible.

• Economic constraints

  • High training costs demand near-term value extraction (internal tools, dev productivity gains) to sustain investment.

Concrete example best practices

• Evaluating LMs for in-game characters

  • Create designer/QA golden question sets and use automated LM-based judging for usefulness, correctness, and tone.

• Handling NPC hallucination

  • Design gameplay so hallucination is tolerable or enhances experience (for example, NPCs that lie during interrogation).

• Hybrid scripting + agents

  • Use agent orchestration to select high-level actions and trigger scripted behaviors for execution and control.

• Public experimentation

  • Test features publicly or in betas to learn player responses (example: Craftton’s PUBG LA beta testing).

People, companies, projects and sources mentioned

• Panel hosts / speakers

  • John Spitzer — VP of Developer Performance & Technology, NVIDIA (host/panelist)
  • Dean — founder/CEO of DART (world-model work)
  • Konguk — Chief AI Officer at Craftton; CTO at Ludo Robotics
  • CH Kim — CEO of Craftton
  • D (aka Dugu) — leads AI and R&D at Creative Assembly (SEGA)
  • Troder Cashion (Troder) — Lead AI Product, Riot Games

• Other contributors mentioned

  • Wesker — leads Riot’s central AI research
  • Yanni/Yiani — aided Creative Assembly with evaluation tooling
  • Ike — initial introducer/moderator (brief mention)
  • Jensen Huang — referenced (NVIDIA CEO)

• Companies, projects and tech referenced

  • NVIDIA — DLSS5, Blackwell GPUs
  • DART — world models, Oasis One demo
  • Craftton — PUBG Ally, PUBG LA, Enzo/ZOE, ReLU Studio projects
  • Ludo Robotics — related robotics group
  • Creative Assembly — Total War series, AI advisors
  • Riot Games — ML bots, League of Legends, Valorant
  • ReLU Studio — LLM NPC experiments
  • Anthropic, OpenAI — model vendors referenced
  • Gemini, Gemini CLI — referenced as used tools
  • GitHub Copilot — example of developer productivity change
  • Oasis One — DART’s public demo
  • DLSS5 — NVIDIA rendering advancement
  • Blackwell 300 — GPU family referenced
  • RAG (retrieval-augmented generation) — architecture approach discussed

End (no further conversation)