AI Engineer Europe 2026
Think You Can Build a Game with AI? Think Again! The New Games Are Just Being Invented!
Loading slides…
Read the talk
Think You Can Build a Game with AI? Think Again! The New Games Are Just Being Invented!
Generating a recognizable game is only the beginning: coherent art, player judgment, runtime decisions, and platform economics determine what comes next.
From a talk by Danielle An and David Hoe
A game replaces the slide deck
Scan a QR code, join a shared game, and navigate to an NPC to reveal the next piece of the presentation. That is how David Hoe and Danielle An begin: the audience controls the movement, and therefore how chaotic the talk becomes. Hoe introduces himself as a prototyper at Meta; An is a principal engineer there. Their presentation makes participation part of the interface instead of leaving it outside the slides.
When several people steer at once, a voting system determines which input wins. The result is a small quest game whose NPC dialogue carries the presentation. An and Hoe lead AI-driven game creation work at Meta, building for casual and professional creators alike. An says they have been doing this for over a year and introduces a tour of games Hoe made in the preceding 12 hours. The intended audience includes people who play games, people who want to make them, and creators wondering how AI changes their work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Recognizable games are only the beginning
By the talk’s 2026 setting, prompting Gemini for a platformer or Tetris is already a familiar experiment. A couple of prompts can produce something recognizable, although it may not work immediately. The initial delight comes from seeing a game appear. Then the novelty wears off: if everyone asks for a little platformer resembling Mario, the results tend to resemble one another too. Producing a game does not yet give it a reason to stand out.
Meanwhile, the shared presentation exposes its own interaction tradeoff. So many attendees are moving the character that the NPC text becomes difficult to read. An suggests stopping movement long enough to see it. The voting system resolves competing inputs, but the audience still has to coordinate around the activity the interface is meant to support.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Lower barriers leave the craft intact
The larger change is who can participate in making games. A programmer who cannot draw or model in 3D can turn to Nano Banana or Meshy to get past an art bottleneck. An artist who previously had to wait for an engineer can use AI assistance to get past a coding bottleneck. These tools remove barriers between disciplines; they do not remove the distinction between a good game and a bad one.
Once the first prototype exists, familiar design demands return:
- Aesthetics: A purple box next to a green platform may communicate the mechanics, but it rarely supplies a compelling visual identity.
- Cohesion: The UI, story, and art should feel as though they belong to one universe.
- Playtesting: Real people’s reactions reveal what works beyond the creator’s own expectations.
The new access to production makes these judgments more consequential, because simply getting something onto the screen is less distinctive.
Hoe then introduces a video from Dale, an art director they work with, showing how key art can anchor an LLM’s work. The example starts with a bear-themed image. Rather than separately describing the desired look for every asset, the creator carries a shared visual reference through the process:
- Iterate on concepts until an image establishes a direction worth pursuing.
- Use that image to guide the style of the game’s assets.
- Let the image also inform possible gameplay, not just surface appearance.
- Keep using the reference across sessions so the model receives a consistent source of context.
A single image can hold several kinds of information at once. Its value is the common direction it gives subsequent decisions, helping the generated pieces belong together.
As those tools shorten the distance between a prompted prototype and a professionally presented game, polish becomes a less sufficient differentiator. An calls the next requirement taste: understanding which people would enjoy the game, why they would enjoy it, and what they are seeking from play. Delivering a functioning artifact and understanding its audience are separate accomplishments.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Runtime LLMs turn personalities into competition
So far, AI has helped produce the game. A runtime LLM introduces another role: making decisions while someone is playing. It can modify, direct, or otherwise influence the experience as it unfolds. An introduces a multiplayer demonstration that she says the team built over a couple of days. Each NPC is driven by an LLM.
Four players first specify how their NPCs should behave. A character might be a thief, honorable, or fast. Once the match starts, the NPCs independently pursue a common objective: collect as many cubes as possible. Personality affects how they approach that objective and how they treat competitors.
The resulting interactions go beyond moving toward an unclaimed cube. An points out characters stealing other NPCs’ cubes, blocking rivals, and kicking them, describing these choices as unscripted decisions made by the runtime models. The distinction is between a designer specifying each encounter in advance and giving characters behavioral direction from which interactions emerge during play. An describes the resulting matches as unique and nonrepeatable.
As the characters begin fighting, personality becomes a design input for the overall social experience. The same broad approach could support more antagonistic competition, friendlier behavior, or cooperation. An presents this as a newly practical area of experimentation, attributing its feasibility to faster inference and cheaper models. Her claim is about the team’s practical opportunity to build these games; the talk supplies no latency or cost measurements and does not establish a historical first for LLM-driven characters. She closes the demonstration by returning to lessons from roughly twelve months of leading this work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
More iteration, and more personal play
The production change starts with dependencies between disciplines. An describes a conventional sequence that moves from design through art, modeling, animation, and coding. Once later stages depend on earlier decisions, revisiting those decisions becomes expensive. AI assistance lets her teams do more of this work in parallel. An reports that her teams achieve updates and iterations in hours or days instead of months. That is a report from their workflow, without a defined comparison workload or team-size breakdown.
| Production pattern | Effect on iteration |
|---|---|
| Sequential handoffs | Upstream changes disrupt downstream work. |
| Parallel work with AI assistance | Teams can revisit decisions more readily. |
The benefit An emphasizes is more time for playtests and changes to the game idea itself. Faster production matters because it creates opportunities to discover what is fun, rather than simply accelerating execution of the original plan.
Runtime direction could also make the experience more personal. An describes herself as coordination challenged: in cooperative games with friends, she worries about holding the team back, sometimes enough to stop playing. An LLM acting as a game master could potentially adjust the experience so she can continue contributing and enjoying the game with them. This is a proposed use of adaptation, not a demonstrated difficulty system. The relevant goal is keeping friends able to play together despite differences in skill.
That possibility does not mean the technology automatically improves the product. Teams are still exploring where runtime LLMs add value. An imagines a Star Trek-style interaction in which a spoken request produces a forest and a bear-hunting experience, but explicitly places it in the future. The present task is to discover useful forms of runtime direction while tools, codebases, and expectations of quality keep changing.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Nondeterminism extends across the platform
At platform scale, the problem expands beyond the behavior of one NPC. An describes Meta’s ambition to support millions of pieces of gaming content as creation becomes more accessible. Agentic systems can participate throughout that platform: interpreting a creator’s prompt, making decisions during a match, and serving, ranking, and delivering content. Nondeterminism can appear across the whole stack.
Conventional stability practices start from a relatively familiar object: a known codebase whose behavior engineers test and debug. In an agentic system, changing a model or a prompt can substantially change behavior even when the surrounding application code stays the same. A model upgrade can therefore become a system-level change. An leaves debugging and maintaining these systems as open engineering questions; the presentation identifies the problem without prescribing a deployment architecture.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a Labubu universe to the next experiment
The longer-term aspiration is a world like OASIS in Ready Player One. An describes the industry as being at the beginning of that possibility, then asks the audience to navigate to one more NPC. The shared controls initially refuse to cooperate. When the audience finally reaches the character, the abstract vision becomes something more personal: An loves Labubu and imagines a universe where she can be whichever Labubu she wants and collect others. The appeal is a world organized around a player’s own interests.
The next experiment depends on how far a creator has already gone:
- First-time creators: Use Gemini, Manus, or another preferred tool to prompt a basic game such as Tetris or an infinite runner. Expect failures as well as moments when the result works surprisingly well.
- Prototype builders: Improve the art, UI, and surprises until the game feels deliberately made rather than like the default output of a prompt.
- Platform engineers: Work through the token economics. The system needs an economic arrangement that works for creators, players, and the platform.
Runtime generation also makes content safety part of the playing experience. If a model can generate images or other content during a session, that output must remain appropriate for its audience. Making creation accessible opens the door; sustaining those worlds requires solving both the economics of their ongoing computation and the safety of what they produce.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Google's introduction to generating and editing images in Gemini, including maintaining subjects and transferring visual styles.
Instructions for turning text prompts into 3D models, with guidance on choosing image input for greater visual consistency.
Further reading
An introduction to Canvas for generating, previewing, and revising code, including game and web-app prototypes.
A 2023 study of LLM characters whose memories, reflections, and plans produce emergent behavior in an interactive simulated town.
An official Manus Academy example of a web game deployed on the platform.
Full transcript
Loading transcript…