Amir
Sadeghian

In 2023, Amir Sadeghian's brother Ali - a machine learning researcher at Google - showed him something. A model. A demo. Proof that AI could generate not just images or text, but interactive environments. The kind of thing that had always required teams of developers, years of iteration, and millions of dollars. Ali had built it at Google Research. The two brothers looked at each other and made a decision that would reshape both of their careers.

That moment became Astrocade.

Amir Sadeghian arrived at this junction carrying a very specific kind of authority. His 2018 Stanford PhD wasn't in abstract machine learning. It was in computer vision and robot perception - the discipline of teaching machines to understand the physical world in real time. He had published landmark papers on pedestrian trajectory prediction (SoPhie, CVPR 2019) and multi-object tracking (JRMOT, IROS 2020). His GIoU metric - Generalized Intersection Over Union - became a standard in object detection pipelines worldwide. He has 9,682 citations and counting. For most researchers, that's a career. For Sadeghian, it was preparation.

Before Astrocade, he co-founded Aibee Inc. as Founding Director of AI. Backed by Sequoia Capital, Aibee built AI systems for VR, AR, and real-time tracking, raising $200M and reaching unicorn valuation. Sadeghian oversaw the technical direction of the AI organization through its rapid scaling. He left in 2022 with a concrete understanding of what world-class AI systems look like when they ship.

Astrocade describes itself as the world's first "wish-to-game" platform - and that framing is precise. A user types a sentence. The platform returns a fully playable game. Not a template. Not a starter kit. A game, with art, animation, sound, music, and mechanics - all coordinated by multiple specialized AI agents working in parallel.

The technical backbone is what Sadeghian calls an agentic AI architecture: a system where distinct AI models handle distinct creative roles - one for gameplay logic, another for visual assets, another for audio - all orchestrated together. This is not a single generative model doing everything at once. It's closer to a production studio, staffed entirely by specialized AI.

The contextual editor - the interface users interact with after generation - is itself partially AI-generated, built around the specific vision the creator described. You don't manipulate a generic toolset. You work within an environment shaped by your own prompt.

Remix is a first-class feature. Users can take any game on the platform, fork it, alter it, extend it, and publish a new version. This creates compounding creative networks - games begetting games - that no traditional development studio could sustain.

Eight months after public launch: 20 million engaged users, 5 million monthly actives, 140 million monthly plays, 75,000+ games, and creators in 80 countries. Those are not metrics you can buy with a clever marketing campaign.

In June 2024, AME Cloud Ventures (Jerry Yang's fund) and NVIDIA Ventures led a $12M seed round. The angels who followed read like a who's-who of the gaming and tech worlds: Eric Schmidt, Niccolo De Masi, John Riccitiello, David Baszucki, and Mike Abbott. Each of those names had a different reason to say yes - and each of them did.

The May 2026 raise was in a different weight class. Singapore's Sea - the conglomerate behind Garena and Shopee, with deep distribution roots across Southeast Asia - led the Series A. Sequoia Capital led the Series B. Between them: $56M. Google's AI Futures Fund, NVIDIA, LG Technology Ventures, Dentsu Ventures, Conviction Embed, Chaac Ventures, and Rogue VC joined the round. Total funding: $68M.

Sequoia partner David Cahn wrote a public investment memo explaining the firm's thinking. The insight he highlighted wasn't about AI or gaming in the abstract. It was about the user. Astrocade's fastest-growing demographic - women ages 20-40 - suggested the platform had escaped the ceiling of traditional gaming entirely. It wasn't competing for the same players everyone else was fighting over. It had found a new category of creators who had never been invited in before.

Part of the new capital is going toward a $10 million creator fund, with monetization available for games reaching 100K significant plays in 30 days. Top creators earned more than $50K on the platform last year.

Sadeghian's public framing of the long-term vision is worth taking seriously. He argues that the next Pixar - the next defining force in interactive storytelling - will need to be both a world-class creative studio and a top AI company simultaneously. Not one then the other. Both at the same time, built into the DNA of the organization from the start.

This isn't a metaphor for efficiency. It's a structural claim about how entertainment will be made. The studios that survive will be the ones where AI isn't bolted on but constitutive - where the creative process and the model are the same thing.

Astrocade's design reflects this. The platform is not a tool for professional developers. It was built for the 8 billion people who have never opened a game engine. The comparison Sadeghian reaches for is TikTok - a platform that didn't improve video editing for professionals, but eliminated the barrier entirely for people who had never made a video before. The question is whether interactive entertainment can undergo the same collapse of entry costs.

The 75,000 games already created on the platform - and the 5 million people playing them every month - suggest the answer is yes. The speed matters too: Astrocade reached these numbers in 8 months. Most platforms would spend 8 years trying.

Before Amir Sadeghian built products, he built knowledge. His Stanford PhD research focused on a specific and consequential problem: how do machines understand how people and objects move through the world - not just where they are, but where they're about to be?

The work he produced during those years, in collaboration with the Stanford Computational Vision and Geometry Lab and research stints at EPFL and Nokia, has been cited over 9,682 times. The most widely deployed result is GIoU - Generalized Intersection Over Union - a metric published at CVPR 2019 that improved the accuracy of bounding box regression in object detection. It's now a standard component in detection pipelines across academia and industry.

The trajectory prediction papers - SoPhie and Social-BiGAT - tackled the problem of predicting how pedestrians move in social settings, incorporating both physical constraints and social norms. This is the kind of work that feeds directly into robotics, autonomous vehicles, and eventually game AI. The research isn't a detour from where Sadeghian ended up. It's the foundation.