AI Makes Games Worse?
Piero Molino, CEO of Studio Atelico, argues that AI should be a foundational game mechanic enabling new experiences rather than a cost-cutting shortcut. He covers on-device multimodal engines built on small fine-tuned models, the economics of cloud inference, artist rights, open source in the game stack, and what skill means in an era of coding agents.
AI as a Core Game Mechanic, Not a Cost Cutter
The central argument is that generative AI in games is most valuable when it becomes a foundational gameplay mechanic that makes previously impossible experiences possible, rather than a way to cheapen art pipelines. Piero Molino draws a direct analogy to physics engines: once physics was available, games like Portal and World of Goo emerged, enabling emergent puzzles and soft-body mechanics that hard-coded movement could never produce. Studio Atelico aims to do the same with AI, giving designers a multimodal “AI engine” that unlocks new genres.
Bobium Brawlers and the Gameplay Sweet Spot
Their first title, Bobium Brawlers, lets players create creatures by describing them in natural language. Each creature gets a generated portrait and a card deck, and a dragon that spits fire yields fire attacks while an ice wolf yields ice-based ones. Because the space of creations is effectively infinite and generated in real time, this design is infeasible without AI in the loop.
The team’s prototyping revealed a clear pattern. Prototypes with minimal AI felt like a sprinkle with no real benefit, while prototypes that were almost entirely AI-governed felt gimmicky and failed as games. Games require challenge and learnable rules so players can build a mental model and get better over time, which pure AI control undermines. The most enjoyable designs blended the two: AI-governed creation followed by classical card-and-dice battles with legible rules. Molino generalizes this to products broadly: the best AI is invisible, letting users stay in flow rather than thinking about models, tokens, or implementation details.
Why Games Have Been Slow to Adopt AI
Molino notes that gaming has historically been an early adopter of new technology, yet AI has not followed that pattern. Two blockers stand out. The first is economic: free-to-play and premium models both reward engagement, but per-inference cloud costs scale with play time. At roughly 20 cents per creature generation, a million players creating five creatures a day would cost about a million dollars daily, which breaks the business model. The second is reception. Players often perceive AI-generated art as lower quality and ethically questionable because models were trained on creators’ work without consent. Molino believes the quality objection may fade, and that studios will become smarter about where AI assets are acceptable, for example reserving hand-drawn work for principal characters while using generated material for background elements.
Multimodal Engines, Small Models, and On-Device Inference
The Studio Atelico engine handles text, structured text, images, audio, speech input, embeddings, semantic search, and classification, but not yet 3D asset generation on the fly. The architecture relies on small models in the range of hundreds of millions to roughly three billion parameters. Some are custom-built, such as a text-to-speech model of around 100 million parameters, and others are open-weight models that are fine-tuned and compressed. Training from scratch at this scale is feasible because pretraining can use openly licensed data, as with NVIDIA’s Nemotron family.
The key tradeoff is a three-way balance among memory footprint, throughput and latency, and output quality. A model that is fast but low quality is unusable, and a high-quality model that exceeds device RAM or runs too slowly is equally impractical. The team targets the intersection of these constraints.
On iOS, Metal is the chosen compute layer, since it lets developers write kernels that the runtime routes to GPU, NPU, or CPU. Molino draws on his time at Stanford’s Hazy Research lab, where FlashAttention originated, to write highly efficient kernels, and describes a proprietary kernel compiler that gives the team an edge. Their first game ships on iPhone, targeting iPhone 12 Pro and iPhone 14. The image model runs in roughly 4 to 6 seconds for a 512×512 image on an iPhone 14, about one second on a four-year-old M1 laptop, and around 20 milliseconds on an RTX 5080, illustrating how the approach scales with hardware.
Post-Training and the Gap With Frontier Models
Molino cites Epoch AI’s analysis that open models runnable on device trail frontier models by roughly six to ten months depending on the task before customization. He also points to the LoRA work from his Predibase days, where around 1,200 fine-tuned models across 30 tasks reached GPT-4-level quality on their specific tasks at a fraction of the cost. The base model became an almost independent variable, and most of the value came from task data and post-training. Customization also serves control: Bobium Brawlers fine-tunes its image model on internal artist output so generated creatures match the studio’s line thickness, color palette, and overall style. A practical pipeline has large cloud models generate input-output pairs during play, which are then used to distill a small on-device model.
The Device as the New Deployment Boundary
When inference runs on the device, the device itself becomes the deployment boundary. Molino envisions a “slider of intelligence” analogous to a graphics quality setting, where weaker hardware runs smaller models and stronger hardware runs larger ones. Hardware trends favor this direction: unified memory architectures such as NVIDIA’s DGX Spark, which has 128 GB, and new laptop and Mac platforms bring workstation-class capability to consumer devices within a few years. He also notes that electricity and memory, rather than GPUs alone, are the real constraints, and mentions high-bandwidth flash as an emerging way to offload memory in ways comparable to what SSDs did for storage.
He sees a plausible future in which Apple reserves part of unified memory for an always-resident OS-level model that developers can query. Most routine tasks, he argues, only need a model that is “good enough,” with cloud models reserved for complex research or scientific work. Because OS updates can refresh the model, this hybrid local-cloud arrangement could steadily shift more work onto the device.
Artist Rights and the Open-Source Question
Studio Atelico’s artist contract compensates creators for the work they produce and gives them royalties tied to game revenue. Molino’s reasoning is that generative models can produce a thousand trees in an artist’s style from one drawn tree, so paying per asset alone would be unfair. Royalties tied to the whole game’s revenue, since individual creatures are not sold separately, are a proportional way to share value.
On openness, the company has released a linter and formatter for GDScript, the language used by Godot, and keeps its core AI engine proprietary for now. Molino expects Godot to continue gaining ground as an open alternative to Unity, drawing a comparison to Solaris versus Linux, and notes that its development is largely self-sustaining through a foundation and an ecosystem of service companies. He argues that small teams tend to be the trailblazers, while large incumbents like Tencent move slower, and that switching costs for multi-year game projects make early technology decisions sticky.
Coding Agents, Taste, and What Separates the Best
Studio Atelico relies heavily on coding agents, but Molino has observed that the same tool produces very different results depending on the developer using it. The payoff is speed for the same team, not headcount reduction. Drawing on a comparison with mathematics, where generating proofs and papers is becoming cheap, he argues that verification and human judgment remain the scarce resources. Choosing what problems to work on, developing taste, and critiquing and understanding the output are where differentiation lies. Good software design principles also remain a human contribution, since models tend toward redundant, bespoke patches and struggle to generalize across many device-specific implementations. Gaining that judgment through experience is the concern, and he suggests education may need to change accordingly.
Beyond Frame Prediction: Structured World Models
Molino is skeptical of the naive formulation of world models as past frames and controls predicting the next frame. Such systems have too little inductive bias, demand enormous data, and offer no clear way to fix bugs. He is more optimistic about hybrids that inject game-engine scene structure into world-model and style-transfer techniques. He thinks this blend could enable interactive experiences that are closer to interactive films than to traditional games.