< Back

Musings from the AI and Games Conference 2025

Team Atelico

At this year’s AI and Games Conference, it was nice to see many developers sharing their experience building on classical game AI methods, experimenting with GenAI and building very effective hybrids. This suggests that the industry is moving from the very early pioneering of adoption of GenAI in games to a new phase focused on practicality.

The whole team went to London for the conference and we loved the conversations we had, particularly around small language models, fine-tuning, and on-device inference. It was gratifying to have these moments with colleagues, not just because the conversation has certainly evolved, but also it’s exactly where we’ve been building all along.

Top Sessions at AI and Games 2025

After Piero opened the conference by sharing some of our own learnings on building games that use local LLMs, we really appreciated the opportunity to hear from peers on many interesting topics covering both classical game AI approaches and GenAI. Here are just a few that stood out for offering real insight, practical frameworks, or a spark of creative ambition. These were the sessions that pushed our thinking forward and reinforced the direction we’re taking as a team.

How to Bully AI into Delivering Meaningful Gameplay

Thomas Keane and Ben Ackland of Meaning Machine delivered one of the most relevant sessions for anyone building AI-native storytelling games. Their breakdown of balancing authored narrative with GenAI improvisation resonated deeply and validated our own discoveries. Particularly, we liked how they:

  • Uncovered the tension between player agency, authorial intent, and LLM creativity, a triangle we navigate daily.
  • Suggested a framework to provide authors a high degree of control over the generated narrative through the use of contextual narrative snippets  + retrieval pipelines + forced story beats, which also help maintaining coherence.

In our own work, trying to find that sweet spot perched between the non-deterministic nature of GenAI and well-constructed systems for gameplay has led us to a very similar authorial approach, focused on authored game mechanics instead of narrative. Look out for a future blogpost by Ennio on this topic!

One Trillion Parameters and No Plans

This session by Jeff Orkin was a riff on his amazing Three States and a Plan talk from GDC some time ago. It explored what happens when you blend Hierarchical Task Networks (HTNs) with modern generative models. Equal parts entertaining and thought-provoking, Jeff’s talk:

  • Presented a compelling vision of LLMs as HTN co-authors, generating plans and behaviors developers can refine.
  • Surfaced a key trend: tools that help designers bridge the gap between gameplay programmers and iterate quicker.

We are thrilled to see industry stalwarts look at ways to leverage new technology and we agree that a combination of established techniques and new ones can be very effective!

Retro-AI: Dungeon Keeper

We loved this session by Ian Shaw; it was a joyful, surprisingly rich dive into how a 1997 classic invented entire categories of game AI out of necessity. This talk had us buzzing as it:

  • Showed how the team derived navigation meshes from first principles, an early example of constraint-driven innovation.
  • Demonstrated that great game AI often succeeds through simple heuristics deployed cleverly.
  • Echoed our belief that indirect-control games demand deeper systemic intelligence, an idea we evolved for the LLM era with GARP.

But Why Is It Saying That?? Making LLM Bots More Reliable

Batu Aytemiz gave one of the most practically valuable sessions of the conference. No hype, just real production lessons from running large-scale LLM NPCs in Roblox. We appreciated that the talk:

  • Detailed a full-stack approach to LLM debugging, error analysis, and reliability testing, including vibe tests and structured annotation pipelines.
  • Introduced production-ready tools every company building games with GenAI will need: prompt iterators, automated UX summarizers, and systematic behavior analysis loops.

Simulation-Based AI with LLMs for Game Agents

Simon Lucas from Queen Mary University of London gave a high-energy, imaginative exploration of combining search-based planning with generative AI. He really:

  • Showed the potential of hybrid agents that mix symbolic reasoning with LLM-driven creativity.
  • Highlighted an under-explored path in games: using LLMs not just for flavor, but to shape simulation-level decision-making.
  • Brought the kind of experimental zest that, as scientists at heart, we love to see in the field, and that aligns with some of our own early R&D explorations.

A thank-you to the organizers

Huge thanks to Tommy Thompson and the AI and Games team for bringing together such a focused, energizing event. It’s no small feat to curate a community that’s both curious and collaborative. The conversations we had, from academic veterans to indie experimenters to folks at ARM, LEGO, and Creative Assembly, reminded us just how fast this field is growing, and how important it is to stay connected.

We’re thrilled to see more fellow travelers joining the journey toward AI-native games that are alive, reactive, and endlessly creative.

Thanks for all the great conversations and see you next year!