Summer Signal '26: We Turned a Museum Into a Live Multimodal AI Stack

At 3:00 PM on Monday, September 14, 2026, the Exploratorium on Pier 15 opened its doors to a different kind of exhibit. The hands-on science museum stayed fully itself, every lens, crank, and mirror ready for curious hands, and on top of it GMI Cloud layered a live gallery of multimodal AI. By 9:00 PM, 178 of the 543 registered guests from 487 companies had watched 11+ speakers move through six chapters, seen two products launch, and drifted between talks past screens of AI-generated art that rotated all evening.

The idea behind Summer Signal '26: The Modern Renaissance was simple. The first Renaissance broke down the walls between art, science, and engineering. Multimodal AI is that convergence happening again, with vision, language, sound, motion, reasoning, and now action merging into single agentic systems. We wanted people to see the possibilities in action, so we set it in a place built for exactly that kind of curiosity.

Watch our recap video here:

The highlights

• A museum as the venue. Pier 15's Exploratorium, with its hands-on science exhibits open all night, became the gallery floor for live AI art, the stage for six chapters of talks, and the hallway track between them.

• Two GMI launches, live. Alex Yeh unveiled GMI's Model-as-a-Service platform and MCP tooling, then ran the agentic workflow in front of the room.

• Frontier video on stage. MiniMax H3 generated live, GMI Prime Inference explained from the inside, and Utopai's PAI 3.0 moving from generation to direction.

• Worlds built in the room. Kyt Janae (Luma AI) built one in 15 minutes; Tripo AI turned images into 3D; Reactor and Runway showed real-time and interactive world models.

• Detection at 99.47%. Resemble AI's DETECT-World numbers across 250+ generative models, plus a sponsored gallery corner of deepfake-detection art.

• Hollywood meets AI-native. Amazon MGM Studios, Inkitt, Luma Labs, and GMI Cloud on one panel with SHÙ Studio and Real Reel.

• A community-made gallery. Roughly 50 short films from the Machine Cinema Community Challenge ran on the museum screens all night.

• SCALE Cohort 2 revealed to close the program.

Inside the Exploratorium

The Exploratorium on Pier 15 along the Embarcadero, a clear September afternoon before doors opened at 3:00 PM.

The Exploratorium sits on Pier 15 along the Embarcadero, a museum built around touching, turning, and testing things, and for one evening its screens, floor, and stage carried a multimodal AI program end to end. Guests picked up badges and walked straight into the art: roughly 50 short films from the Community Challenge, co-branded with Machine Cinema and made in the weeks before the event for $3K cash, $2K in GMI Cloud credits, and a $5K prize pool, running across the museum screens.

Between chapters the exhibits stayed open, so the hallway track happened at the science displays with the gallery art evolving on the screens overhead. It gave the night a rhythm most conferences lack: a discussion about world models, followed by interactive hands-on activities.

The Modern Renaissance Gallery: Community Challenge films and partner work playing wall to wall on the museum screens, with the crowd taking it in from below.

Launched on stage

The room settled at 4:00 PM when Alex Yeh (CEO, GMI Cloud) took the podium for "The Multimodal Stack, Rebuilt for Agents." He launched two products in one talk, GMI's Model-as-a-Service platform and MCP tooling, and then ran them live: an agent discovering multimodal models, connecting to them, and running them inside a workflow, in front of the people who will build on it. The model layer is now something agents browse, and every builder in the room can follow the same path today.

Six chapters, from seeing to convergence

From there the program moved through six chapters, each named for a capability the multimodal stack has absorbed. The model layer and agent discovery, real-time voice and characters, frontier video inference and generation, agentic filmmaking, world models, interactive video, image-to-3D, deepfake detection, execution environments for agents, live world-building, the reasoning layer after generation, and the meeting point of Hollywood and AI-native entertainment.

Igor Poletaev (CSO, Inworld AI) led a talk, "From AI Models to Living Characters", tracing how real-time voice and interaction are reshaping consumer AI, and made the case that a character people talk to is a different product from a model people prompt. Haomin (Inference PM, GMI Cloud) covered what it takes to run frontier video models at scale on GMI Prime Inference. Victor S. (Global Product Marketing Manager, MiniMax) then demoed MiniMax H3, the newest leap in AI video, generated on stage while the room watched. Jie Yang (CTO, Utopai Studios) closed with "From Generation to Direction: Agentic Filmmaking with PAI 3.0," where the model moves from producing shots to directing them.

The Exploratorium theater during Chapter III: Moving, GMI Cloud's frontier video session on screen.

Alex Yeh (CEO, GMI Cloud) opens Chapter I with "The Multimodal Stack, Rebuilt for Agents," launching GMI's Model-as-a-Service platform and MCP tooling from the Exploratorium podium.

Igor Poletaev (CSO, Inworld AI) on what drives adoption for living characters: quality, latency, cost, and multilinguality, with the whole interaction as the scorecard.

Jie Yang (Co-Founder & CTO, Utopai Studios) closes Chapter III with "Agentic Filmmaking with PAI 3.0," where the model steps from generating shots to directing them.

Jimei Yang (Director of Research, Runway) on real-time video through distillation: hit Generate and the clip starts playing within 100 milliseconds.

A gallery break followed, and the room drifted back onto the exhibit floor while the art on screen kept evolving. The best conversations of the night happened here, between the science displays, with a speaker from the last chapter and a founder from the next one comparing notes.

Gallery break: conversations moved onto the exhibit floor, GMI lanyards on and sponsor logos on every badge.

Worlds was the densest chapter, four views of generated space in 45 minutes. Ahmed Ahres (Head of GTM, Reactor) covered real-time video and world models. Jimei Yang (Director of Research, Runway) presented interactive video models across characters, worlds, and interfaces. Tripo AI showed images becoming interactive 3D worlds you could move through.

Then Tedi Papajorgji (CTO, Resemble AI) turned the chapter around with "How We Build Detection Models for a Threat That Never Stops Growing." He walked through the DETECT-World architecture and its reported accuracy: 99.47% on audio, 98.2% on video, and 95.8% on images across 250+ generative models. On a night built around celebrating generation, the detection layer got its own chapter slot and its own gallery corner, and the room took the point.

From Multimodal to Agentic asked what happens once generation works. Mingjun Sun (PM, GMI Cloud) presented "The Multimodal Renaissance Is Here. Can Your Product Keep Up?" on execution environments for multimodal agents at scale. Kyt Janae (Forward Deployed Creative, Luma AI) built a world live in 15 minutes, start to finish, in front of the room. Jerry Liu (CEO & Co-founder, LlamaIndex) closed with "What Happens After You Generate?" on the reasoning and agentic layer beneath multimodal systems, the part that decides what a generated asset is for.

Kyt Janae (Luma AI) builds a world live in Luma Scenes, a toy robot taking its first steps across a generated field, 15 minutes on the clock.

The Convergence brought Hollywood and AI-native entertainment onto one stage, in a panel presented with SHÙ Studio and Real Reel. Steve Bannerman (former Head of International VFX & Post Production, Amazon MGM Studios), Ali Albazaz (CEO, Inkitt), Gray Crawford (Interaction Designer & Artist, Luma Labs), and Yujing Qian (VP of Engineering, GMI Cloud) traded views on studio pipelines, publishing, interaction design, and inference, with Celine Zen (Founder, SHÙ) moderating.

Chapter VI: The Convergence. Celine Zen (SHÙ Studio, Real Reel) moderates Ali Albazaz (Inkitt), Gray Crawford (Luma AI), Yujing Qian (GMI Cloud), and Steve Bannerman (feature film producer and media executive) at the Exploratorium podium.

At 8:30 PM the SCALE Cohort 2 reveal brought the next wave of builders onto the stage and led straight into the closing reception.

Speakers and the GMI Cloud team on the Exploratorium stage after the closing chapter, The Modern Renaissance title card overhead.

The full program

What we took home

• The stack is one thing now. Voice, video, 3D, world models, detection, and reasoning showed up as layers of a single agentic system, and the audience treated them that way.

• Live beats slides. Every chapter had something running on stage: a launch demo, H3 generating, a world in 15 minutes, DETECT-World numbers.

• Detection belongs in the same room as generation. Resemble's session and gallery corner made that a feature of the night.

• Builders want infrastructure they can walk up to. The Model-as-a-Service and MCP launch turned the model catalog into something agents browse.

• The venue mattered. A museum built for hands-on curiosity set the tone for a crowd that wanted to touch the stack, and it made Summer Signal feel like an evening out instead of a conference.

Thanks to the partners

Thanks to the sponsors and partners on the floor: Inworld AI, MiniMax, Utopai Studios, Reactor, Resemble AI, LlamaIndex, Luma AI, Runway, Tripo AI, Alibaba Cloud (Qwen-Image), Vimmerse, and Inkitt, along with community partners Machine Cinema, Fantastic Day, SHÙ Studio, and Real Reel, and to every speaker who brought a live demo.

Built on GMI Cloud

Underneath every chapter was the same infrastructure. GMI Cloud is one of seven NVIDIA Reference Platform Cloud Partners globally, with a library of 170+ AI models built for real-time, production multimodal workloads, and every demo on stage ran on it, from GMI Prime Inference serving frontier video to the Model-as-a-Service platform and MCP tooling launched in Chapter I. Summer Signal showed what partners are building on that layer, and the SCALE Cohort 2 reveal showed who builds on it next. See you at the next one.