The Fast Company AI 20 for 2025

“Every project you see on the big screen is a result of an endless amount of creative compromises from the directors and the filmmakers,” he says.

Moonvalley, which has raised $154 million, works with four of Hollywood’s biggest studios, advising them on how to integrate AI into productions and reskill workers. Its model is trained on licensed, high-resolution content and is capable of production-grade video generation.

Over the past year, Moonvalley has shifted its focus to developing “world models,” which generate video that accurately portrays the complex physics of something like a car crash. As these models grow, says Talukdar, “they start to be able to reason on things that they haven’t seen before.” —Mark Sullivan

A onetime Google summer intern and veteran of DeepMind, the British AI startup Google acquired in 2014, Kavukcuoglu helped manage the 2023 merger of DeepMind and Google Brain, another research arm. He remains CTO of the combined entity, Google DeepMind, but now he reports directly to Pichai, who announced the promotion in a memo explaining that Kavukcuoglu’s new role would bring “more seamless integration, faster iteration, and greater efficiency” to Google’s lab-to-market pipeline. Hundreds of staffers working to apply Google’s Gemini large language model to transform its search engine are now part of his team, The Information reported. He’s also involved with everything from data center strategy to bolstering the Google Cloud web services platform.

Kavukcuoglu’s background is in the science of AI, not turning it into offerings that appeal to billions of people. Still, as Gemini-powered features increasingly show up in Google mainstays such as search, Android, and Gmail, investors have grown more optimistic that Google will be a titan of the AI era rather than a victim of it. As the company strives to keep that momentum going, Kavukcuoglu’s deep familiarity with its technical stack should be an asset. “There’s a long history of research that built up to this point,” he told Big Technology’s Alex Kantrowitz last May. —Harry McCracken

The product he spearheaded in 2024, called Automated Reasoning Checks, acts like Mr. Spock for language models, using rigid logic to catch and correct up to 99% of hallucinations.

Now Cook is applying automated reasoning to agents: autonomous, LLM-powered enterprise apps. Many businesses don’t trust them—yet. “First of all, this [agent] could lie to me,” explains Cook. “But secondly, if I let it launch rockets”—his metaphor for irreversible actions—“will it launch rockets when we’re not supposed to?”

Amazon is betting that automated reasoning, and Cook, can keep agents on a leash. —John Pavlus

Abridge is also integrated into medical platforms such as Athenahealth and Wolters Kluwer, where it can fill out forms and expedite tasks like insurance pre-authorization or writing prescriptions.

Rao, who has experience as a tech investor with UPMC, developed the idea while making his rounds. His hospital’s proximity to Carnegie Mellon, a tech hub, gave him a firsthand look at machine learning. That led him to found his company in 2018, long before ChatGPT came around. Abridge, which has raised a total of approximately $800 million, is currently in use at more than 150 U.S. health systems, including Johns Hopkins Medicine, the Mayo Clinic, Kaiser Permanente, and Duke Health. The less time physicians spend on paperwork, the more time they have to focus on their patients.

“As a doctor, I’m not compensated for the care that I deliver—I’m compensated for the care that I documented that I deliver,” Rao says. “So we are extending the documentation to help with billing.” —Yasmin Gagne

“There’s this growing sense that both digital technology and AI are happening to people, they’re not necessarily happening with us or for us,” she says.

Imbue, which emerged from stealth in late 2022, aims to help people create their own AI tools. It’s working on an AI-assisted software development tool called Sculptor, which became open to public preview in late September.

“What we’re trying to do is create a tool that lets you feel the structure of your software and understand it,” says Qiu, by enabling it to remember context across different projects and suggesting ways to refine users’ code. While other AI software development startups such as Bolt and Replit offer stand-alone products, Sculptor acts as an interface for Claude Code, allowing developers to run multiple agents in parallel. —Jared Newman

Goldman had already been immersed in these questions at eBay founder Pierre Omidyar’s impact investment firm, where she evaluated the social consequences of emerging technology. Goldman is now helping refine Salesforce’s ethical principles around the deployment and testing of generative AI and agentic tools. Her team has helped develop systems to ensure AI follows instructions, avoids toxic behavior, and stays within established ethical guidelines.

“Those types of tools are increasingly important as AI takes on more autonomy,” she says. “You want to make sure that the person that’s setting up the system is able to see in advance what it’s going to produce.”

But while cloud technology has continued to evolve, Goldman says one thing has not: establishing trust with customers. “Obviously, we are a business, and being commercially successful is very important,” she says. “Also, we know that trust is what makes that possible.” —Steven Melendez

Depending on whom you ask, the document will either secure the United States’ lead in AI or unleash reckless proliferation.

The plan focuses on accelerating innovation through deregulation, streamlining the construction of data centers, and driving the adoption of American-made AI tools abroad. It includes popular provisions like embracing open-source AI, along with divisive ones such as requiring federal agencies to work only with LLM developers whose AI models are “free from top-down ideological bias” and withholding AI funding from states that pass AI laws the administration deems burdensome.

Even as the industry has praised the document, critics have panned it for failing to curb AI’s potential harms, such as discriminatory system biases. But avoiding assumptions about AI’s future is the point, says Ball, who left the White House in August and is now a fellow at the conservative Foundation for American Innovation. “Washington’s really bad at forecasting how technology will develop,” he says. “We don’t want to make those mistakes.” —Issie Lapowsky

Following a stint as chief scientist at Uber’s self-driving car unit, Urtasun launched Waabi in 2021 to build a verifiable, human-interpretable AI model for autonomous driving. Waabi-enabled big rigs have been on public roads since 2023 and are slated for driverless operation by the end of 2025. Though many autonomous truck systems are limited to highways and depots, Waabi’s technology is designed to carry goods all the way to their final destinations on surface streets. The company has raised more than $280 million to date.

Urtasun also remains a computer science professor at the University of Toronto, where her graduate students conduct doctoral research at Waabi through a unique arrangement. Some recent research involves simulation, allowing Waabi to now let its AI practice in situations it’s never encountered in the physical world—a key advantage for its system.

Waabi’s AI has shown that it can quickly react to novel conditions, even seamlessly managing its first encounter with rain, which it had never practiced for. “It was kind of nerve-racking,” says Urtasun, who was in that vehicle with some investors. “But it was amazing to see.” —Steven Melendez

Since joining the MIT Media Lab in September, she has been expanding her research into ways of empowering individuals and communities to fight back against algorithmic overreach. This has led her to focus on “contestable systems,” which let human users “talk back” to algorithms, perhaps to contest a content moderation decision that may at first seem final. This could be through a set of preference settings to control the content of a social feed, or it might be through an AI voice or chat interface that allows a user to engage the algorithm in a plain language dialogue. If no solution is reached, the issue might be bumped up to a human moderator.

“As we build these systems, and they seem to be permeating our society right now, one of my big goals is not to ignore human intuition and not to have people give up agency,” Karahalios says. —Mark Sullivan

RDUs compete with traditional GPUs for AI inference—the application of trained models to new data that happens when we use AI apps. The goal: to slash inference power requirements, while also reducing latency. Customers with strict privacy requirements can run servers with SambaNova’s RDUs on site, or they can have the company manage them in the cloud. “We found it hard to believe that we had to rely on an architecture that was started 25 years ago, 30 years ago, and primarily focused on graphics and gaming,” Liang says.

SambaNova raised $676 million at a $5.1 billion valuation in April 2021, yet challenges remain, most notably the dominance and mindshare of large players such as Nvidia. Still, Liang believes SambaNova’s advantages will accrue with AI’s increasing power and performance demands. “All the things that we’ve designed natively into the product are going to become more and more important,” he says. —Jared Newman

The fruits of Kossnick’s labor are more and more apparent. AI features now span Figma’s portfolio, from its flagship Design app to the new Make vibe coding tool to features for creating slideshows, websites, and marketing assets. Given Figma’s inherently multidisciplinary nature—two-thirds of its users work in areas outside design—the technology can knock down some of creativity’s traditional boundaries, he asserts: “It’s easier with the help of AI to reach into a lane where you’re not as familiar with the details and bring the context, the intuition, the insight that you have.”

At the same time, the company has been careful not to mess up elements of its experiences that people liked in the first place—which means that some of its best AI is nearly invisible, at least until users know they want it. “Figma Design’s canvas is kind of like the Google homepage or Facebook newsfeed,” says Kossnick. “A single pixel of friction literally slows down millions of people every day.” —Harry McCracken

Pano’s system detects wildfires in a median of 3.5 minutes—revolutionary compared with traditional 911 alert times. It triangulates fire locations within hundreds of meters and alerts multiple agencies at once.

Kastner’s eight-person AI team has spent five years training models to spot fires and distinguish smoke from dust or clouds. “Quietly, computer vision has gotten really, really good,” she says. While enterprises (and more and more states) have embraced the system—the company has secured more than $140 million in cumulative contracts and raised a $44 million funding round in June—federal adoption remains the biggest hurdle. To that end, Kastner frequently travels to Washington to push agencies to modernize procurement. “We’re serving as a bridge between the technology sector and emergency managers on the front lines of these ever-worsening natural disasters,” she says. —Jeremy Caplan

Today, Turing supplies training data for eight of the nine companies developing the largest general-purpose AI models. The shift has also turned Turing into a quiet but central player in the artificial intelligence ecosystem, shaping what the next generation of AI systems will know. Turing is profitable and valued at roughly $2.2 billion.

As models have advanced, generic data (often scraped from the web) is no longer good enough to achieve further intelligence gains. AI researchers need a regular supply of data that captures deep subject-matter expertise across domains from STEM to healthcare, Siddharth says. “We’re able to do that because we have two engines: the talent engine that’s finding smart talent and the data generation platform that the talent works on.” —Mark Sullivan