Max Kreminski

About

I’m a researcher in artificial intelligence, human-computer interaction, and creativity. I primarily design and develop AI-based creativity support tools to assist human designers and artists, particularly in the domains of storytelling, poetry, and game design. As an interdisciplinary scholar with a background in technical games research, I’ve also worked on a wide variety of topics related to interactive narrative and digital games.

Currently, I’m a research scientist at Midjourney, where I lead the Storytelling Lab. Before that, from 2022-2024, I was an assistant professor of computer science and engineering at Santa Clara University. I hold a Ph.D. in Computational Media from the University of California, Santa Cruz, where I was a member of the Expressive Intelligence Studio. In the past, I’ve also been a resident at Stochastic Labs; a consultant at various game industry and tech companies; and a researcher in several labs, most notably the Mobile & Environmental Media Lab, at the University of Southern California.

Here’s a complete list of my papers (current count: 63; awards and nominations: 10). You can also check out my full CV for links to many of my talks and other output.

Selected projects and papers

Screenshot of a short initial “setting prompt” being expanded to a set of images and a character concept in the Patchwork interface.

Patchwork is a collaborative AI-supported worldbuilding canvas. Users work together to craft a coherent fictional setting and tell stories in that setting, with a multimodal AI system helping them gradually expand and refine their shared vision for the storyworld. We’re also using Patchwork to prototype a variety of novel control and steering techniques for generative AI, including several that have become spinoff projects (e.g., Patchview and Toyteller).

📝 Papers: UIST 2024a, UIST 2024b, HAI-GEN 2024
🗣️ Talks: Autodesk Research, South Park Commons
🗞️ Press: VentureBeat, The Decoder
🤝 With: John Joon Young Chung, Melissa Roemmele, Uchan Sun

Visualization of three sets of artifacts produced by three different users of a single creativity support tool. The sets of artifacts are being overlapped to assess the tool’s homogenization effects.

Homogenization analysis and expressive range coverage analysis are evaluation methods for creativity support tools (CSTs). By analyzing artifacts created by CST users in relation to larger design spaces, we can capture striking visual portraits of interaction dynamics; diagnose homogenization effects of generative AI tools; and generally illuminate tool impacts on creative outcomes.

📝 Papers: C&C 2024 (best paper), CHI 2024, HAI-GEN 2022, In2Writing 2024
🗞️ Press: Digital Frontier
🤝 With: Barrett Anderson, Jash Shah, Isaac Karth

Screenshot of several storytelling goals, each representing a different plot thread, in in the Loose Ends user interface.

Loose Ends is an AI-based narrative instrument that can be played to produce narrative, much like a musical instrument can be played to produce music. It aims to help players interleave multiple parallel plot threads into an effective high-level plot structure that satisfies player-specified storytelling goals.

📝 Papers: AIIDE 2022 (best artifact), In2Writing 2022, ICIDS 2022
🤝 With: Melanie Dickinson

Screenshot of the Blabrecs AI testing the nonsense word “torbible” and saying “looks good to me!”

Blabrecs is a rules modification for the wordgame Scrabble that forbids players from using actual English dictionary words. Instead, players may only use nonsense words that sound like English to an AI trained on the dictionary.

📝 Papers: NeurIPS 2023, AIIDE 2021
🗣️ Talks: !!Con 2024, GDC 2024 (skip to 1:51:00)
🗞️ Press: New Scientist, The Next Web, Dicebreaker
🤝 With: Isaac Karth

Partial example of a Felt sifting pattern that matches a sequence of events involving two impulsive betrayals.

Story sifters are computational recognizers of potentially compelling story structures. I’ve worked to frame story sifting as logic programming (Felt); developed AI-based tools that make logical sifters easier to build (Synthesifter); and enabled sifters to reason about possible future events (Winnow) and prioritize unexpected stories (StU).

📝 Papers: ICIDS 2019, INT 2020, AIIDE 2021, ICIDS 2022, book chapter
🗞️ Press: The Verge
🤝 With: Melanie Dickinson

Screenshot of the Germinate user interface, showing a design intent on the left and a generated game on the right.

Germinate is a mixed-initiative casual creator for rhetorical game design, built on the Gemini game generator. It generates games based on a user-provided design intent, then provides users with support in exploring the resulting design space and refining their intent based on the generated games.

📝 Papers: AIIDE 2020, ICCC 2021
🤝 With: Melanie Dickinson

Screenshot of an Academical scene in which an aging hippie professor discusses an ethical dilemma with a grad student.

Academical is an interactive narrative game that uses roleplaying to train new researchers in responsible conduct of research. Our user studies have found that Academical is better than existing RCR training materials at engaging users, teaching some RCR topics, and promoting moral reasoning, without falling behind in other areas.

📝 Papers: FDG 2020 (exceptional paper), CHI 2021, CHI 2020, book chapter, Frontiers in VR
🤝 With: Eddie Melcer, Katy Grasse, Nic Junius, many others

System diagram of Why Are We Like This?, showing how the system suggests what should happen next in response to player actions.

Why Are We Like This? is an AI-augmented collaborative storytelling game. Two players write a story in a pastiche of the cozy mystery genre, with support from a simulation-based AI system that suggests what might happen next based on both character and authorial goals.

📝 Papers: FDG 2020, ELO 2020, PCG 2019, AIIDE 2020, ICCC 2021
🤝 With: Melanie Dickinson

System diagram of StoryAssembler, showing how execution co-routines between a forward state-space planner and hierarchical task network planner.

StoryAssembler is an interactive narrative system that assembles choice-based stories on the fly from a library of author-provided story fragments, using planning to ensure that every story it produces meets a set of author-specified requirements.

📝 Papers: FDG 2019 (best paper honorable mention), ICIDS 2018
🤝 With: Jacob Garbe

A two-line poem created using Redactionist: “the mission is the infinite variability / that poetry can speak”.

Redactionist is an AI-based tool that helps you create poetry by erasing words from a source text. It builds on my earlier work in this direction: a browser bookmarklet that can turn any arbitrary webpage into blackout poetry.

📝 Papers: HAI-GEN 2022, ICCC 2021, PCG 2019
🗣️ Talks: !!Con West 2019, Computational Poetry Workshop 2019

System diagram of a single generated Game Boy RPG.

GBS is a genre-specific game description language for Game Boy RPGs, extracted from the GB Studio game creation tool. Using GBS as a target language for game generation enabled us to explore the generation of larger and more expressively sophisticated games than would otherwise be tractable; we believe that GBS and similar GSGDLs might serve as especially good foundations for future AI-based game design tools.

📝 Papers: CoG 2021, FDG 2021 (best game or demo)
🤝 With: Tamara Duplantis, Isaac Karth

Partial screenshot of codebook that notes dimensions of variation between Civilization VI and Stellaris retellings, from the AIIDE 2019 paper linked below.

Retellings studies are a series of investigations into the stories that players tell about their experiences in games, especially simulation-driven emergent narrative games. This work helps us better understand how to provide computational support for human storytelling practices.

📝 Papers: ICIDS 2021a (best short paper nominee), ICIDS 2021b (best short paper nominee), AIIDE 2019, UX of AI 2019

Screenshot of Epitaph, showing a procedurally generated alien civilization called the Gep.

Epitaph is a procedural narrative idlegame about existential risks and the death of civilizations. Bennett Foddy (the creator of QWOP) considers it “one of the great web games”. It was the precursor to my academic work on gardening games, which use procedural content generation to support gameplay experiences of caretaking rather than extraction.

📝 Papers: PCG 2018

Social VR environments are immersive virtual social spaces. I’ve participated in several attempts to make social VR environments better for creative collaboration – particularly by endowing users with social superpowers, or communicative augmentations that couldn’t exist in meatspace.

📝 Papers: CHI 2022, DIS 2023, SocialVR 2021, DIS 2018, HCII 2018
🤝 With: Josh McVeigh-Schultz, Katherine Isbister, many others