Wonder News Morning: Japan online-game survey reports 100% AI use, MAGIC links 3D worlds

Every Japanese online-game company covered by the latest JOGA market survey is using generative AI, according to a report preview published overnight. Today’s Wonder News also covers a system that builds connected 3D game scenes, voice-driven game editing, automated playtesting, collaborative storytelling, visual-tool agents, open coding harnesses, and faster serving for a 397-billion-parameter model.

What Changed Overnight

• AUTOMATON WEST published details from the JOGA Online Game Market Research Report 2026 late July 14, including reported adoption rates for Gemini, Claude, and GitHub Copilot.
• Google published a systems report on July 14 describing 3.1× and 4.7× throughput gains for different Qwen 3.5-397B serving workloads on Ironwood TPUs.
• Develop:Brighton entered its second day on July 15 with a design programme that includes “AI Wants to Play,” a session about using AI across game production.

Industry Adoption And Events

A Japanese online-game survey reports universal generative AI use

The Japan Online Game Association and Kadokawa ASCII Research Laboratories released the JOGA Online Game Market Research Report 2026 on July 10. A preview cited by AUTOMATON WEST says 100% of the online-game companies surveyed use generative AI tools.

Gemini was the most-used tool at 94%, followed by Claude at 84% and GitHub Copilot at 76%. Companies showed the most interest in using the technology for user-preference analysis and behavior prediction. Players most often named copyright infringement and games becoming too similar as their concerns.

The public preview does not state the company sample size, so the result should not be read as a census of Japan’s entire games industry. Its scope is also narrower than a 2025 Computer Entertainment Supplier’s Association survey, which found AI use at 51% among 54 responding Japanese game companies.

Develop:Brighton puts AI inside the design programme

Develop:Brighton is running from July 14 through July 16, marking the UK game-development conference’s 20th anniversary. The organizers say the 2025 event drew 5,500 attendees and this year’s programme again spans design, art, audio, business, coding, and education.

The design track includes “AI Wants to Play: A Game Designer’s Guide to AI in Every Stage of Development.” Its presence does not show how widely attendees endorse generative AI, but it puts practical AI workflow discussion on the same schedule as level design, narrative, accessibility, and production craft.

Generated Worlds And Live Editing

MAGIC generates portals that work across 3D scenes

MAGIC turns one natural-language prompt into a runnable multi-scene game project. The system first plans a shared representation for rooms and portals, checks whether furnished interiors remain navigable, writes transition scripts, and then runs each transition during evaluation.

Across a new benchmark of 100 multi-scene cases, the authors report that MAGIC produced an executable project in every case. Its transition-identification evaluation reached 0.99 precision, 0.95 recall, and 0.96 F1. It also recovered more reference portals and produced more navigable layouts than an LLM baseline and Holodeck.

The benchmark measures whether spaces connect and can be crossed. It does not establish whether the generated projects contain compelling objectives, combat, progression, or replay value.

Voice commands edit about 100 arcade-game parameters during play

Maya Grace Torii, Takahito Murakami, and Yoichi Ochiai built a Space Invaders-style game that players can alter with spoken instructions without leaving the session. An LLM converts each instruction into structured changes across roughly 100 configuration fields covering mechanics, visuals, controls, and audio.

The user study found generally positive experiences and moderate workload, with outcomes not strongly tied to prior programming experience. Changes to immediately visible parameters aligned more with usability, while edits to core game structures aligned more with enjoyment. The system keeps changes inside a defined parameter space rather than allowing unrestricted code generation.

Playtesting And Collaborative Creation

Graph and transformer models take on Candy Crush playtesting

A new comparative study uses Candy Crush Saga boards to test whether general-purpose models can predict player behavior with less hand-built feature engineering. The researchers compare a graph attention network and a BERT-style transformer with convolutional baselines.

Both general representations performed better on difficult board configurations, according to the paper. Because the abstract omits player counts and headline accuracy figures, the result supports an architecture comparison rather than a claim that automated models can replace human playtests.

FIERO combines physical cards with generative storytelling

FIERO is a multiplayer web card game for groups writing a story together. Physical cards structure the conversation, while a digital interface generates contextual images, helps the group make decisions, maintains narrative coherence, and combines contributions.

In a study of 60 participants, FIERO produced higher idea fluency and novelty than online collaborative writing alone. The authors also report more coherent plots, while finding that the system did not reduce participants’ sense of agency. The work treats generative AI as one component of a social game rather than the sole author.

Agents Builders Can Inspect And Test

StructAgent carries explicit task state into Minecraft

StructAgent organizes a digital agent’s progress as a compact causal state, with checkpoints and verifier-backed transitions instead of relying only on raw interaction history. On OSWorld-Verified, it raised success for Qwen 3.5-9B from 27.0% to 46.9% and for Qwen 3.5-27B from 31.6% to 62.2%.

The strongest reported open model, MiniMax-M3, reached 78.9%. The authors also tested the same framework in Minecraft, though the abstract does not provide a separate Minecraft score. The transfer matters because game agents must preserve progress across long sequences of movement, tool use, and changing world state.

MM-ToolSandBox finds visual precision behind 53% of agent failures

Apple researchers introduced MM-ToolSandBox with more than 500 tools across 16 application domains. Its 258 human-verified scenarios and 50 interactive-UI variants require agents to respond to multiple images, corrections, revised goals, and state changes.

The best of 12 tested models stayed below a 50% success rate. The researchers attribute 53% of failures to incorrect extraction of information from images even when the planned workflow was otherwise sound. For visual game-building tools, reading the current scene or interface accurately can be as important as deciding which action comes next.

ToFu opens the coding-agent harness, not only the model

ToFu is an MIT-licensed agent harness that can read repositories, edit files, run commands, and connect to development tools. Its authors position it as both a research assistant and an inspectable research object, so teams can change orchestration logic and study tool behavior instead of treating the harness as a closed layer around the model.

The paper claims better token efficiency, lower cost, and stronger multilingual support than existing harnesses, without providing those comparison values in the abstract. Its MIT license and white-box design are the parts readers can verify directly from the release.

Model Serving

Google reports faster Qwen 3.5-397B inference on Ironwood

Google’s July 14 engineering report describes serving Qwen 3.5-397B-A17B, a mixture-of-experts model with 397 billion total parameters and 17 billion active per token. The team built reusable JAX and Pallas components for attention, expert routing, and communication on Ironwood TPU7x systems.

Between April and June, Google says those changes improved decode-heavy throughput about 3.1× and prefill-heavy throughput about 4.7× at the 512-concurrency tier. These are provider measurements on specialized hardware, not a general price or latency guarantee. Google says the reusable optimizations are being integrated into the open-source vLLM and SGLang serving frameworks.

Watch Next

• Whether the full JOGA report publishes its company sample size, survey questions, and breakdown of production versus player-facing uses.
• Whether MAGIC releases projects that outside developers can run, edit, and test beyond portal navigation.
• How voice editing behaves when players request conflicting rule changes or changes outside its 100-field configuration space.
• Whether Candy Crush playtesting models generalize to new mechanics and live player populations.
• Whether FIERO’s gains persist with children, classrooms, or groups that already know one another.
• What Minecraft tasks StructAgent completed and whether its explicit state remains useful in longer, open-ended play.
• Which AI-assisted games or creator tools appear at Develop:Brighton and the SIGGRAPH Games Summit on July 19.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.