Anastasis Germanidis

Germanidis fell in love with programming at age 11 in Athens and came to the United States at 18 to study neuroscience and film, then turned back to computer science and worked at several Silicon Valley tech firms before tiring of the culture there. He recalled that during that period he was primarily an engineer, doing backend, distributed systems, and machine learning engineering at startups, while also creating interactive art that used technology as a medium; the two worlds were largely separate.⁠[1][1][2][2][2][2]

His personal website lists old projects from this period: “Welcome, Programmable Human” (2015), which simulated his own behavior with a computer program; Thingscoop, a deep-learning tool for searching and filtering videos by their content, and Videodigest, which produced automatic video summaries (both 2015); Antipersona (2016), which simulated using Twitter as someone else; Uncanny Road (2017), which used generative adversarial networks to collectively “hallucinate” a never-ending road; and Film Mirror (2018), an interactive installation using pix2pixHD that let participants perform as iconic film characters.⁠[3][3][3][3][3][3]

In 2016, Germanidis met Cristóbal Valenzuela and Alejandro Matamala Ortiz in ITP, a graduate program at New York University focused on the intersection of art and technology. The three began working together to explore creative uses of machine learning, building small tools and giving them to artists, designers, and filmmakers to try. He has said that Runway was started by “three failed filmmakers.”⁠[1][2][2][2]

The three incorporated Runway in New York in late 2018 and launched the first version of the product in January 2019.⁠[2][1]

On February 6, 2023, Runway’s research team submitted the paper “Structure and Content-Guided Video Synthesis with Diffusion Models,” which underlies Gen-1, to arXiv; Germanidis was one of its five authors. Gen-1 uses words and images to generate new videos out of existing ones; the subsequent Gen-2 is a multimodal system that can generate novel videos from text, images, or video clips. The research pages for both models carry his byline.⁠[4][4][5][6]

Germanidis announced that Runway was starting a long-term research effort around what it calls general world models. He defined a world model as an AI system that builds an internal representation of an environment and uses it to simulate future events within that environment, and suggested that video generative systems such as Gen-2 can be thought of as very early and limited forms of general world models.⁠[7][7][7][7]

Germanidis published Gen-3 Alpha under his byline, describing it as the first of a new generation of foundation models trained by Runway on new infrastructure built for large-scale multimodal training, and a step toward building general world models. On his personal website he lists Gen-3 Alpha’s selection for TIMETIMETIME is an American newsmagazine and media brand founded by Henry Luce and Briton Hadden, which published its first issue on March 3, 1923; it is headquartered in New York and known for its Person of the Year selection and the TIME100 list. Since 2018 it has been personally owned by Marc and Lynne Benioff; its current Chief Executive Officer is Jessica Sibley, and its Editor-in-Chief is Sam Jacobs.Open the full entry’s 100 Best Inventions of 2024 among related press. In October of that year he was interviewed on the podcast The Cognitive Revolution as co-founder and CTO.⁠[8][8][8][3][9][9]

In the essay he argued that rather than first “solving intelligence,” the focus should be on solving the “everything else” part first, by building systems for universal simulation: training models simultaneously on a far wider variety of data and modalities than today, including observational data from scientific experiments and raw sensor data, and then using those models to run virtual experiments much faster than is possible in physical reality, thereby accelerating scientific progress.⁠[10][10][10][10]

Runway raised a $315 million Series E round led by General Atlantic, with participation from Nvidia, Fidelity Management & Research, AllianceBernstein, AdobeAdobeAdobe (Adobe Inc.) is a software company headquartered in San Jose, California, founded in 1982 by John Warnock and Charles Geschke. Known for PostScript, PDF and creative software such as Acrobat, Photoshop, Illustrator, Premiere Pro and After Effects, it moved its main creative products to Creative Cloud subscriptions from 2013 and expanded into digital marketing and generative AI; it is listed on Nasdaq under the ticker ADBE.Open the full entry Ventures, Mirae Asset, Emphatic Capital, Felicis, Premji Invest, and AMD VenturesAMD VenturesAMD Ventures is the corporate venture capital arm of semiconductor company AMD. It backs AI models, developer tools, cloud infrastructure, silicon and interconnect technologies, and provides portfolio companies with compute, engineering and market support. Its predecessor was the AMD Fusion Fund; AMD used the Ventures name in an official 2012 announcement.Open the full entry; according to a source familiar with the matter, the company’s valuation nearly doubled to $5.3 billion. The company said the new funds would be used to pre-train the next generation of world models and bring them to new products and industries.⁠[11][11][11][11]

Runway announced that Germanidis would serve as co-CEO alongside Cristóbal Valenzuela, describing him as the driving force behind the company’s research and technical vision and saying the title change simply formalized a structure that already existed; at the same time, Kamil Sindi became chief technology officer.⁠[12][12][12][12][12]

Germanidis told TechCrunch that training models directly on observational data from the world is the next frontier of AI; language models distill existing human knowledge, and getting beyond that requires less biased data. He sees world models as scientific infrastructure and believes that building a better scientist than human scientists could accelerate progress in understanding the universe and solving problems; given enough time and resources, his moonshot goal for Runway’s technology is biological world models and anti-aging research. He also said the robotics unit Runway launched the previous year had already resulted in real-world testing and deployments.⁠[1][1][1][1][1][1][1]

In the essay he called the capability produced by repeated interaction with closed, verifiable environments “sandbox superintelligence,” such as surpassing humans at reasoning, coding, or formal problems, while “real-world superintelligence” is the ability to act above human capability in complex, partially observed systems. He argued that the former is still far from the latter, and advocated replacing hand-built deterministic environments with world models trained on large quantities of existing interactions and observational data.⁠[13][13][13][13][13]

At the Runway AI Summit held at The Masonic in San Francisco, Germanidis delivered the opening keynote with CTO Kamil Sindi, CPO Anthony Maggio, and others, presenting Runway’s latest updates with spotlights on Runway Robotics, real-time video, and Runway Dev, and then held a research Q&A with research leads Jimei Yang and Katie Lewis Grumman. Speakers that day also included NVIDIA’s Ming-Yu LiuMing-Yu Liu (NVIDIA)Ming-Yu Liu is a computer vision and generative AI researcher, vice president of NVIDIA’s Cosmos Lab, and an IEEE Fellow. He leads Cosmos, NVIDIA’s world foundation model platform for “physical AI” such as robots and autonomous vehicles; his team earlier developed image-generation technologies including pix2pixHD and GauGAN (NVIDIA Canvas).Open the full entry, Physical Intelligence’s Quan VuongQuan VuongQuan Vuong is a robot learning researcher and one of the co-founders of Physical Intelligence, a robotics foundation model company. He previously did robotics research at Google, contributed to the development of robotics transformer models including RT-1 and RT-2, and, with Pannag Sanketi, published under their bylines the cross-robot dataset Open X-Embodiment.Open the full entry, Google DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry’s Jon BarronJon Barron (Google DeepMind)Jon Barron is an American computer vision researcher and a principal research scientist at Google DeepMind, where he leads a small team in San Francisco that mostly works on neural radiance fields (NeRF). He is one of the authors of the 2020 NeRF paper and, as first author, published follow-up work including mip-NeRF, mip-NeRF 360, and Zip-NeRF.Open the full entry, and The Bot Company’s Paril JainParil JainParil Jain is a robotics and AI engineer and the co-founder and chief technology officer of The Bot Company, a home-robot company; he previously led the AI tech team at Tesla.Open the full entry.⁠[14][14][14][14][14][14][14][14][14]

Personal website: https://agermanidis.com/ (opens in a new window)⁠[3]

X: Public account: https://x.com/agermanidis (opens in a new window)⁠[3]

GitHub: Public account: https://github.com/agermanidis (opens in a new window)⁠[3]

Runway: Official website: https://runway.com/ (opens in a new window).⁠[12]