Up in the Cascade Mountains, 90 miles east of Seattle, a group of high-ranking Amazon engineers gather for a private off-site. They hail from the companys North America Stores division, and theyre here at this Hyatt resort on a crisp September morning to brainstorm new ways to power Amazons retail experiences. Passing the hotel lobbys IMAX-like mountain views, they filter into windowless meeting rooms.
Down the hall, the off-sites keynote speakerByron Cook, vice president and distinguished scientist at Amazonslips into an empty conference room to have some breakfast before his presentation.
Cook is 6-foot-6, but with sloping shoulders that make his otherwise imposing frame appear disarmingly concave. Hes wearing a rumpled version of his typical uniform: a thick black hoodie and loose black pants hanging slightly high at the ankles. An ashy thatch of hair points in whatever direction his hands happen to push it. Cook, 54, doesnt look much like a scientist, distinguished or otherwise, and certainly not like a VPmore like a nerdy roadie.
They dont know who I am yet, he tells me between bites of breakfast, referring to the two dozen or so engineers now taking their seats. Despite his exalted title, Cook has faced plenty of rooms like this in his self-made role as a kind of missionary within Amazon, spreading the word about a powerful but obscure type of artificial intelligence called automated reasoning.
As hes done many times before, Cook is here to get the highly technical people in that room to become believers. Hes championing an approach to AI that isnt powered by gigawatt data centers stuffed with GPUs, but by principles old enough to be written on papyrusand one thats already positioning Amazon as a leader in the tech industrys quest to solve the problem of hallucinations.
Cook doesnt have a pretalk ritual, no need to get in character. Hes riffing half-seriously to a colleague about the pleasures of riding the New York subway in the summertime when someone mentions that the session is about to begin. He immediately drops his fork and strides out. His next batch of converts awaits.
When ChatGPT hit the world with asteroid force in November 2022, Amazon was caught flat-footed just like everyone else. Not because it was an AI laggardthe tech giant had recently overhauled nearly all of its divisions, including its massive cloud-computing arm, AWS, to leverage deep learning. Amazon also dominated the smart-home market, with 300 million devices connected to Alexa, its AI-powered assistant. It had even been researching and building large language models, the tech behind ChatGPT, for multiple years, as CEO Andy Jassy told CNBC in April 2023.
But OpenAIs chatbot changed the definitionand expectationsof AI overnight. Before, AI was still a mostly invisible ingredient in voice assistants, facial recognition, and other relatively narrow applications. Now it was suddenly seen as a prompt-powered genie, an infinitely flexible do-anything machine that every tech company needed to embraceor risk irrelevance. Less than six months after ChatGPTs debut, Amazon launched Bedrock, its own AWS-hosted generative AI service for enterprise clients, a list that currently includes 3M, DoorDash, Thomson Reuters, United Airlines, and the New York Stock Exchange, among others.
Over the next two years, Amazon injected generative AI into product after product, from Prime Video and Amazon Music (where it powers content recommendation and discovery tools) to online retail pages (where sellers can use it to optimize their product listings), and even into internal tools used by AWSs sales teams. The company has released two chatbots (a shopping assistant called Rufus and the business-friendly Amazon Q), plus its own set of foundation models called Novathey are general-purpose AI systems, akin to Googles Gemini or OpenAIs line of GPTs. Amazon even caught the industry fever around so-called AGI (artificial general intelligence, a yet-to-be-achieved version of AI that does any cognitive task a human can) and in late 2024 launched AGI Lab, a flashy internal incubator led by David Luan, an ex-OpenAI researcher.
Still, none of it captured the publics imagination like the stream of shiny objects emitted by OpenAI (reasoning models!), Anthropic (chatbots that code!), and Google (AI Overviews! Deep Research!). Like Apple, Amazon was unable to turn its early lead in AI assistants into an advantage in this new era. Alexa and Siri simply cannot compete.
But maybe that has been for the best, because 2025 was the year that AIs sheen suddenly started to come off: GPT-5 fell flat, vibe coding went from killer app to major risk, and an MIT study rattled the industry by claiming that 95% of businesses get no meaningful return on their AI pilot projects.
It was against this backdropthe summer AI turned ugly, as Deutsche Bank analysts called itthat Amazon publicly released Automated Reasoning Checks, a feature promising to minimize AI hallucinations and deliver up to 99% verification accuracy for generative AI applications built on AWS. The product was Cooks brainchild; in a nutshell, it snuffs out hallucinations using the same kind of computerized logic that lets mathematicians prove 300-page-long theorems. (In fact, a 1956 automated reasoning program called Logic Theorist is considered by some experts to be the worlds first AI system, finding new and shorter versions of some of the proofs in Principia Mathematica, one of the most fundamental texts in modern mathematics.)
Sexy, it aint. Still, Swami Sivasubramanian, one of Amazons highest-ranking AI executives, who serves on Jassys S-team of direct advisers, was impressed enough to call Automated Reasoning Checks a new milestone in AI safety in a LinkedIn post. Matt Garman, CEO of AWS, referred to it as game-changing.
[carousel_block id=”carousel-1763954270090″]
Automated reasonings promise of quashing AI misbehavior with math has quietly become an essential part of Amazons strategy around agentsthose LLM-powered workbots that are supposed to transform enterprise productivity [checks watch] any day now. Apparently, businesses have serious side-eye about that, too: Earlier thi year, Gartner predicted that more than 40% of agentic AI projects will be ditched within the next two years due to inadequate risk controls. The company told me recently that it predicts that 30% to 60% of the projects that do go forward will fail due to hallucinations, risk, and lack of governance. Thats not a prophecy Amazon can afford to let come truenot with a potential market for AI agents that Gartner estimates to be worth $512 billion by 2029. One way or another, hallucinations have got to go.
The question is how. Agents are just souped-up LLMs, which means they can and will go off the railsin fact, as OpenAI itself recently admitted following an internal study, they cant not. What Cook helped Amazon realize, just months after ChatGPTs release, was that they already had a secret weapon for extinguishing hallucinations, hidden in plain sight. Automated reasoning is the polar opposite of generative AI: old, stiff, and hard to use. Many at Amazon had never heard of it. But Cook knew how to wield it, having brought it to Amazon nearly 10 years ago as a way of rooting out hidden security vulnerabilities within AWS. And hed been amassing what he estimates to be the largest group of automated reasoning experts in the tech industry.
Now that investment is set to pay off in a way that Amazon never expected. Automated Reasoning Checks is just the first of many products that the company plans to release (on a timetable it wont specify) that fuse the flexibility of language models with the proven reliability of automated reasoning. The latest, called Policy in Amazon Bedrock Agentcore and previewed this week at AWS’s annual Re:Invent conference, uses automated reasoning to stop agents from taking actions they’re not allowed to (such as issuing customer refunds based on fraudulent requests).
If this combined approachknown as neuro-symbolic AIcan reduce the potential failure rate of agentic AI projects by even a fraction of a percent, it would be worth hundreds of millions of dollars, say analysts at Gartner. And Amazon knows it. To realize the transformative potential of AI agents and truly change the way we live and work, we need that trust, Sivasubramanian says. We believe the foundation for trustworthy, production-ready AI agents lies in automated reasoning.
To understand why Amazon is banking on automated reasoning, its worth sketching out how its different from the kind of AI youve already heard of. Unlike neural networks, which learn patterns by ingesting millions or even billions of examples, automated reasoning relies on a special language called formal logic to express problems as a kind of arithmetic, based on principles that date back to ancient Greece. Computers can use this rule-based approach to calculate the answers to yes-or-no questions with mathematical certaintynot probabilistic best guesses, as deep learning does.
Think of automated reasoning like TurboTax for solving complex logical problems: As long as the problems are expressed in a special language, computers can do most of the workand have been doing so for decades. Since 1994, when a flaw in Intels Pentium chips cost the company half a billion dollars to fix, nearly all microchip manufacturers have used automated reasoning to prove the correctness of designs in advance. The French government used it to verify the software for Pariss first self-driving Métro train in 1998. In 2004, NASA even used it to control the Spirit and Opportunity rovers on Mars.
Theres a catch, of course: Because automated reasoning can only reduce problems to three possible outcomesyes, no, or the equivalent of does not computefinding ways to apply this logically bulletproof but incredibly rigid style of AI to the real world can be difficult and expensive. But when automated reasoning works, it really workscollapsing vast, even unknowable possibilities into a single mathematical guarantee that can compute in milliseconds on an average CPU. And Cook is very, very good at getting automated reasoning to work.
Cook began his career building a formidable scientific reputation at Microsoft Research, where he spent a decade applying automated reasoning to everything from systems biology to the famously unsolvable halting problem in computer science. (Want a foolproof way to tell in advance if any computer program will run normally or get stuck in an infinite loop? Sorry, not possible. Thats the halting problem.) But by 2014, he was looking to put his findings, many of which have been published as peer-reviewed research, to work outside the lab. I was figuring out: Where is the biggest blast radius? Wheres the place I could go to foment a revolution? he says. I watched everyone moving to the cloud, and was like, I think AWS is the place to go.
The first problem Amazon aimed Cook at was cloud security. Reporting directly to then chief information security officer Stephen Schmidt, Cook and his newly formed Automated Reasoning Group (ARG) painstakingly translated AWS security protocols into the language of mathematical proofs and then used their logic-based tools to surface hidden flaws. Once those flaws were corrected, those same tools could then prove with certainty that the system was secure.
Some at AWS were dubious at first. When you look mad scientist up in the dictionary, Byrons picture is in the margin, says Eric Brandwine, an Amazon distinguished engineer who at the time worked on security for AWS. Early on, I challenged [him] on a lot of this stuff. But as Cooks group fleshed out plans and racked up small but significant winslike catching a vulnerability in AWSs Key Management Service, the cryptographic holy of holies that controls how clients safeguard their dataskeptics started becoming evangelists.
Some of these [were] beautiful bugstheyd been there for years and never been found by our best experts, and never been found by bad guys, says James Hamilton, a legendary distinguished engineer within Amazon who now directly advises Andy Jassy. And yet, automated reasoning found them.
From 2018 onward, Amazons automated reasoning experts worked with engineers to encode the technology into nearly every part of AWS, from analytics and storage to developer tools and content delivery. One particular niche of cloud-computing clientsheavily regulated financial service firms, like Goldman Sachs and the global hedge fund Bridgewater Associates, with sensitive data and strict compliance requirementsfound automated reasonings promise of provable security extremely compelling. When ChatGPT appeared and the world flung itself headfirst into generative AI, these companies did too. But they still wanted to keep the one small thing, Cook says, that theyd become accustomed to along th way: trust.
That customer feedback spurred Cook to imagine how LLMs and automated reasoning might fit together. The solution that he and his collaborators prototyped in the summer of 2023 works by leveraging the same logical framework that worked so well for squishing security bugs in AWS.
Step one: Take any policy meant to inform a chatbot (say, a stack of HR documentation, or zoning regulations) and translate it into formal logicthe special language of automated reasoning. Step two: Translate any responses generated by the bot too. Step three: Calculate. If theres a discrepancy between what the LLM wants to say and what the policy allows, the automated reasoning engine will catch it, flag it, and tell the bot to try again. (For humans in the loop, itll also provide logical proof of what went wrong and how, and suggest specific fixes if needed.)
We showed that to senior leadership, and they went nuts for it, says Nadia Labai, a senior applied scientist at AWS who partnered with Cook on the project. The demo went on to become Automated Reasoning Checks, which Amazon previewed at its annual Re:Invent conference in December 2024. PwC, one of the Big Four global accounting and consulting firms, was among the first AWS clients to adopt it.
We do a lot of work in pharmaceutical, energy, and utilities, all of which are regulated, says Matt Wood, PwCs global and U.S. commercial technology and innovation officer. PwC relies on solutions like AWSs automated reasoning tool to check the accuracy of the outputs of its generative AI toolsincluding agents. But Wood sees the technologys appeal spreading beyond finance and other regulation-heavy industries.
Look at what it took to set up a website 25 years agothat was a refined set of skills. Today, you go on Squarespace, click a button, and its done, he says. My expectation is that automated reasoning will follow a similar path. Amazon will make this easier and easier: If you want an automated reasoning check on something, youll have one.
Amazon has already embarked on this path with its own enterprise products and internal systems. Rufus, the AI shopping assistant, uses automated reasoning to keep its responses relevant and accurate. Warehouse robots use it to coordinate their actions in close quarters. Nova, Amazons fleet of generative AI foundation models, uses it to improve so-called chain of thought capabilities.
And then there are the agents. Cook says the company has multiple agentic AI projects in development that incorporate automated reasoning, with intended applications in software development, security, and policy enforcement in AWS. One is Policy in AgentCore, which Amazon released after this story was reported. Another thats peeking out from behind the curtain is Auto, an agent built into Kiro, Amazons new AI programming tool, that will use formal logic to help make sure bot-written code matches humans intended specifications.
But Sivasubramanian, AWSs vice president for agentic AI (and Cooks boss), isnt coy about the commitment Amazon is making. We believe agentic AI has the potential to be our next multibillion-dollar business, he says. As agents are granted more and more autonomy . . . automated reasoning will be key in helping them reach widespread enterprise adoption.
Agents are part of why Cook is touting automated reasoning to his engineer colleagues from the North American Stores division at their off-site in the mountains. Retail might not seem to have much in common with finance or pharma, but its a domain thats full of decisions with real stakes. (While onstage at re:Invent 2025, Cook said that “giving an agent access to your credit card is like giving a teenager access to your credit card… You might end up owning a pony or a warehouse full of candy.”) And in that environment, relying on autonomous botsempowered to do anything from execute transactions to rewrite softwarecan turn hallucination from tolerable quirk into Russian roulette.
Its a matter of scale: When one vibe coding VC unleashes an agent that accidentally nukes his own apps database, as happened earlier this year to SaaS investor Jason Lemkin, its a funny story. (He got the data back.) But if Fortune 500 companies start deploying swarms of agents that accidentally mislead customers, destroy records, or break industry regulations, theres no Undo button.
Enterprise software is full of these potential pitfalls, and existing methods for reducing hallucination arent always strong enough to keep agents from blundering into them. Thats because agents shift the definition of hallucination itself, from errors in word to errors in deed. First of all, this thing could lie to me, explains Cook. But secondly, if I let it launch rocketshis metaphor for irreversible actionswill it launch rockets when were not supposed to?
Back in his hotel room after the keynote, Cook is reviewing the contents of a confidential slide deck about how automated reasoning can solve this rocket-launching problem. The demo, which he hurriedly mentioned in his talk (he ran out of time before being able to show it), describes a system that can transform safety policies for an agentdos and donts, written in natural languageinto a flowchart-like visualization of how the agent can and cannot behave, all backed by mathematical proof. Theres even an Attempt to Fix button to use if the system detects an anomaly.
Cook calls the demo a concept car, but some of its ideas made it into Policy in AgentCore, which is already available in preview to some AWS customers.
PwC, for one, sees Amazons logic-backed take on AI extending into coordinating the agents themselves. If youve got agents building other agents, collaborating with other agents, managing other agents, agents all the way down, says Wood, then having a way of forcing consistency [on their behavior] is going to be really, really importantwhich is where I think automated reasoning will play a role.
The ability to reliably orchestrate the actions of AInot just single agents, but entangled legions of them, at scaleis a target that Amazon has squarely in its sights. But automated reasoning may not be the only way to get the job done.
EY, another Big Four firm, recently launched its own neuro-symbolic solution to AI hallucinations, EY Growth Platforms, which fuses deep learning with proprietary knowledge graphs. A startup called Kognitos offers business-friendly agents backed by a deterministic symbolic program, dubbed English as Code. Others, like PromptQL, forgo neuro-symbolic methods altogether, preferring the simulated reasoning of frontier LLMs. But even they still attack the agent hallucination problem much like Amazon does: by using generative AI to translate business processes into a special internal language thats easy to audit and control.
That translation process is where Amazon built a 10-year lead with automated reasoning. Now it has to maintain it. Nadia Labai is currently working on ways to improve Amazons techniques for using LLMs to convert natural language into formal logic. Its part of a strategy that could help turn Amazons brand of customer-driven, business-friendly AI into a new class of industry- defining infrastructure.
A few days before the off-site, I met with Cook in a conference room at Amazons Seattle headquarters. Sitting with his legs tucked catlike beneath him, Cook mused about his own vision for the future of automated reasoningone that extends far beyond Amazons ambitions for enterpise-grade AI.
The world, he says, is filled with socio-technical systemspatchworks of often-abstruse rules that only highly paid experts can easily navigate, from civil statutes to insurance policies. Right now, rich people get [to take advantage of] that stuff, he continues. But if the rest of us had a way to manipulate these systems in natural language (thanks, LLMs) with an underlying proof of correctness (thanks, automated reasoning), a workaday kind of superintelligence could be unlocked. Not the kind that helps us colonize the galaxy, as Google DeepMind CEO Demis Hassabis envisions, but one that simply helps people navigate the complexity of everyday life, like figuring out where its legal to build housing for an aging relative or how to get an insurance company to cover their expensive medication.
You could have an app that, in an hour of your own time, would get answers to questions that before would take you months, Cook says. That democratizes, if you will, access to truth. And thats the start of a new era.
This story is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
Over the past decade, Figma has transformed how people within companies collaborate to turn software ideas into polished products. Now the company is itself being transformed by AI. The technology is beginning to show its potential to take on much of the detail work that has required human attention in design, coding, and other domains. But the end game involves far more than typing chatbot-style prompts and waiting for the results.
I spoke with Figmas head of AI, David Kossnickone of Fast Companys AI 20 honorees for 2025about what the company has accomplished so far and where hes trying to steer it. We’re still in chapter one, maybe the start of chapter two, he told me.
This Q&A is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers. It has been edited for length and clarity.
Talk a little bit about what your work at Figma encompasses and how you came to have this job.
Anything that has AI in it, I and my team touch in some way. It’s everything from traditional AI tools like search, which we’ve rebuilt using multimodal embeddings, to some of our newer, AI-forward workflows. Figma Make is an example of that.
As to how I came to get this job, I’ll give you a short version. I knew a lot of the Figma team for a long time. The chief product officer, Yuhki [Yamashita], and I went to college together. He was at my wedding.
I did a startup of my own, and one of our board members was John Lilly, who was also on the board of Figma. I actually met [Figma cofounder/CEO] Dylan [Field] when there was, like, a 20-person Figma team, because we were building a game engine, and Figma is basically a game engine, with all sorts of custom renderings. [Lilly] was like, You guys should compare notes.
So I’ve known the team for a long time, and it’s a product I’ve used a lot. And then, about a year and a half ago, when I joined, I’d been working on AI at Coda, which was then acquired by Grammarly.
As a big Figma user, I also felt like there was just such a huge opportunity for Figma, and it had barely gotten started. So I was thinking about what’s next and sharing it with Yuhki: Theres a lot you guys could do. He was like, I know, we just don’t have the right team here yet. You wanna come? I was like, That sounds amazing.
Is there a particular Figma philosophy about AI and how to put it into this experience that’s been around for a while, and which people choose to use because they like it, in most cases?
There’s been a couple of learnings, both from our own team and from working with customers. A lot of our biggest customers are technology companies themselves. Many are integrating AI themselves. And so we’ve learned through themwhat’s working and what they’re trying.
There have been two industry trends, and we’ve done both here. One is trying to find existing workflows that you can add AI to, to save users time, to delight them, to give them new capabilities. And also building totally new experiences that have AI as the core of the workflow.
Interestingly, we’ve actually done some market research and surveys of users and other companies. People understand and value the new AI for workflows even more. I think that is counterintuitive. You think you have such big products, and adding efficiencies to them is very viable. And it is. But often, AI is a little more invisible there. Kt’s embedded in a workflow that you’re used to, and so the thing that is forefront in your mind is the workflow itself
That’s good. We don’t want to get in people’s way. Figma Designs canvas is kind of like the Google homepage or Facebook news feed, where a single pixel of friction literally slows down millions of people every day. Which makes for interesting challenges. How do you introduce things so they dont bother people?
But on the flip side, there’s a lot of new workflows and new tools. Peopleespecially our type of customersare always experimenting. And so they’re very open to trying a totally different approach.
Historically, Figma has been this thing that human beings use to collaborate with other human beings to create stuff from scratch, and often very carefully considered stuff. What’s the experience like of integrating tools that take some of that heavy lifting off their shoulders?
I think it’s super exciting. It feels and looks different for different user types. So as an example, we actually just finished up a $100,000 hackathon, our first ever, for Figma Make. It was totally inspiring seeing all the range of things people have made.
There were students. There were people who never learned to code. There were designers who code a lot, and its just helping them do it faster. There were hobbyists. For a lot of those user types, a very common theme was, Wow, I just couldn’t have done this before.
The other way it feels is as a kind of thought partner to experts. I feel this myself as a [product manager] when I chat with Figma Make or ChatGPT. I have a problem. I have a solution in mind. And actually, there are some other solutions I hadn’t thought about, because I was so focused on this one solution. It can help you pull back and see a wider solution space, and explore a few other threads in a very cheap way before you go too deep.
Its like Doctor Strange, where he has this magic crystal that lets him look into all the different possible futures. Expert users are always running simulations in their heads. What if I move this button over here? How’s the user behavior going to change? What does that mean for the next part of the experience? We’re finding that these types of AI tools make that loop so much faster, where it’s like, I’m just going to try exploring a bunch. I’m going to literally make them, but make them 10 times as quickly, and play out all those different end states.
How far is Figma down the continuum from having no AI to AI being everywhere and doing everything AI could possibly do?
It’s an interesting question. There’s AI today and AI in the future. If all research was frozen, there would probably still be five years of new product experiences that the industry could build from current models. But the pace of model improvement is still really high as well.
For us, I’d say we’re still in chapter one, maybe the start of chapter two. And chapter one was, We’re going to do a bunch of basic features, get our feet wet, save time in your workflows. Chapter two is, We’re doing some new AI-first experiences. Figma Make, that whole category of prompt-to-app, is very, very new.
As the models get better and faster and cheaper, what other new workflows are going to become available? Today, things like autocomplete, as an example, are hard to make fast, and hard to make cheap, and hard tomake high quality. And, you know, we’re still using many interfaces in the industry that feel like typing at a terminal from the ’60s. That’s not the final interface. That’s not the final workflow. I think the interfaces are going to become more visual, more exploratory.
It’s part of why I’m so excited about Figma and why I came here. As AI gets better, what you want the experience of working with an AI to feel like is going to be more and more similar to what you want the experience of working with a human to feel like. You’re going to want to brainstorm with the AI before it goes off and thinks for 10 hours and then builds something. You’re going to want to work through the big trade-offs. Youre going to want your teammates in there too, not just the AI.
I think that’ll be a super exciting place, where things like code become implementation details that AIs are more and more capable of driving, with humans reviewing.
You might not spend a lot of time thinking about your web browser, whether its Safari, Chrome, or something else. But the decades-old piece of software remains a pretty important canvas for getting things done. Thats why Tara Feener, who spent years developing creative tools with companies such as Adobe, WeTransfer, and Vimeo, decided to join the Browser Company and within two years became head of engineering, overseeing its AI-forward Dia browser. This is more ambitious than any of the other things Ive done, because its where you live your life, and where you create within, she says.
Whereas a conventional browser presents you with a search box on its home screen, Dia will either answer your query with AI or route it to a traditional search based on what you write. You can also ask for information from your open tabs or have Dia intelligently sort them into groups. Several of these features have since found their way into more mainstream browsers such as Google Chrome and Microsoft Edge, and in September, Atlassian announced it had acquired the Browser Company and Dia (a $610 million deal), hoping to develop the ultimate AI browser for knowledge workers.Other AI companies are catching on to the importance of owning a browser. Perplexity has launched Comet, and OpenAI launched ChatGPT Atlas in October. This strategic value isnt lost on Feener, who notes that browsers are typically the starting point for workers seeking information. They also provide a treasure trove of context for AI assistants. Dia can already do things like analyze your history for trends and draft messages in Gmail. Feener says her team has never felt more creative coming up with things to do next.With Dia, we have context, we have memory, we have your cookies, so we actually own the entire layer, she says. Just like TikTok gets better with every swipe, every time you open something in Dia, we learn something about you.
This profile is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
Bringing a new drug to market usually requires a decade-long, multibillion-dollar journey, with a high failure rate in the clinical trial phase. Nvidias Kimberly Powell is at the center of a major industry effort to apply AI to the challenge.
If you look at the history of drug discovery, weve been kind of circling around the same targets for a long time, and weve largely exhausted the drugs for those targets, she says. A target is a biological molecule, often a protein, thats causing a disease. But human biology is extraordinarily complex, and many diseases are likely caused by multiple targets.Thats why cancer is so hard, says Powell. Because its many things going wrong in concert that actually cause cancer and cause different people to respond to cancer differently.Nvidia, which in July became the first publicly traded company to cross $4 trillion in market capitalization, is the primary provider of the chips and infrastructure that power large AI models, both within the tech companies developing the models and the far larger number of businesses relying on them. New generative AI models are quite capable of encoding and generating words, numbers, images, and computer code. But much of the work in the healthcare space involves specialized data sets, including DNA and protein structures. The sheer number of molecule combinations is mind-bogglingly big, straining the capacity of language models. Nvidia is customizing its hardware and software to work in that world.[W]e have to do a bunch of really intricate data science work to . . . take this method and apply it to these crazy data domains, Powell says. Were going from language and words that are just short little sequences to something thats 3 billion [characters] long.Powell, who was recruited by Nvidia to jump-start its investment in healthcare 17 years ago, manages the companys relationships with healthcare giants and startups, trying to translate their business and research problems into computational solutions. Among those partners are 5,000 or so startups participating in Nvidias Inception accelerator program.I spend a ton of my time talking to the disrupters, she explains. Because theyre really thinking about what [AI computing] needs to be possible in two to three years time.
This profile is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
Raquel Urtasun is the founder and CEO of self-driving truck startup Waabi as well as a computer science professor at the University of Toronto. Unlike some competitors, Waabis AI technology is designed to drive goods all the way to their destinations, rather than merely to autonomous vehicle hubs near highways.
Urtasun, one of Fast Companys AI 20 honorees for 2025, spoke with us about the relationship between her academic and industry work, what sets Waabi apart from the competition, and the role augmented reality and simulation play in teaching computers to drive even in unusual road conditions.
This Q&A is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers. It has been edited for length and clarity.
Can you tell me a bit about your background and how Waabi got started?
Ive been working in AI for the last 25 years, and I started in academia, because AI systems werent ready for the real world. There was a lot of innovation that needed to happen in order to enable the revolution that we see today.
For the last 15 years, Ive been dedicated to building AI systems for self-driving. Eight years ago, I made a jump to industry: I was chief scientist and head of R&D for Ubers self-driving program, which gave me a lot of visibility in terms of what building a world-class program and bringing the technology to market would look like. One of the things that became clear was that there was a tremendous opportunity for a disrupter in the industry, because everybody was going with an approach that was extremely complex and brittle, where you needed to incorporate by hand all the knowledge that the system should have. It was not something that was going to provide a scalable solution.
So a little bit over four years ago, I left Uber to go all in on a different generation of technology. I had deep conviction that we should build a system designed with AI-first principles, where its a single AI system end-to-end, but at the same time a system that is built for the physical world. It has to be verifiable and interpretable. It has to have the ability to prove the safety of the system, be very efficient, and run onboard the vehicle.
The second core pillar was that the data is as important as the model. You will never be able to observe everything and fully test the system by deploying fleets of vehicles. So we built a best-in-class simulator, where we can actually prove its realism.
And what differentiates your approach from the competition today?
The big difference is that other players have a black-box architecture, where they train the system basically with imitation learning to imitate what humans do. Its very hard to validate and verify and impossible to trace a decision. If the system does something wrong, you cant really explain why that is the case, and its impossible to really have guarantees about the system.
Thats okay for a level two system [where a human is expected to be able to take over], but when you want to deploy level four, without a human, that becomes a huge problem.
We built something very different, where the system is forced to interpret and explain at every fraction of a second all the things it could do, and how good or bad those decisions are, and then it chooses the best maneuver. And then through the simulator, we can learn much better how to handle safety-critical situations, and much faster as well.
How are you able to ensure the simulator works as well as real-world driving?
The goal of the simulator is to expose the self-driving vehicles full stack to many different situations. You want to prove that under each specific situation, how the system drives is the same as if the situation happens in the real world. So we take all the situations where Waabi driver has driven in the real world, and clone them in simulation, and then we see, did the truck do the same thing.
We also recently unveiled a really exciting breakthrough with mixed-reality testing. The way the industry does safety testing is they bring a self-driving vehicle to a closed course and they expose it to a dozen, maybe two dozen, scenarios that are very simple in order to say it has basic capabilities. Its very orchestrated, and they use dummies in order to test things that are safety critical. Its a very small number of non-repeatable tests.
But you can actually do safety testing in a much better way if you can do augmented reality on the self-driving vehicle. With our truck driving around in a closed course, we can intercept the live sensor data and create a view where theres a mix of reality and simulation, so in real time, as its driving in the world, its seeing all kinds of simulated situations as though they were real.
That way, you can have orders of magnitude more tests. You can test all kinds of things that are otherwise impossible, like accidents on the road, a traffic jam, construction, or motorbikes cutting in front of you. You can mix real vehicles with things that are not real, like an emergency vehicle in the opposite lane.
Youre also a full professor. Are you still teaching and supervising graduate students?
I do not teachI obviously do not have time to teach at all. I do have graduate students, but they do their studies at the company. We have this really interesting partnership with the University of Toronto.
If you want to really learn and do research in self-driving, it is a must that you get access to a full product. And thats impossible in academia. So a few years ago, we designed this program where students can do research within the company. Its one of a kind, and to me, this is the future of education for physical AI.
When did you realize the time was ripe for moving from academic research to industry work?
That was about eight and a half years ago. We were at the forefront of innovation, and I saw companies were using our technology, but it was hard for me to understand if we were working on the right things and if there was something that I hadnt thought of that is important when deploying a real product in the real world.
And I decided at the time to join Uber, and I had an amazing almost four years. It blew my mind in terms of how the problem of self-driving is much bigger than I thought. I thought, Okay, autonomy is basically it, and then I learned about how you need to design the hardware, the software, the systems around safety, etc., in a way that everything is scalable and efficient.
It was very clear to me that end-to-end systems and foundational models would be the thing. And four and a half years in, our rate of hitting milestones really speaks to this technology. Its amazingto give an example, the first time that we drove in rain, the system had never seen rain before. And it drove with no interventions in rain, even though it never saw the phenomenon before.
That for me was the “aha” moment. I was actually [in the vehicle] with some investors on the track, so it was kind of nerve-racking. But it was amazing to see. I always have very, very high expectations, but it blew my mind what it could do.
As gaming platforms Roblox and Fortnite have exploded in popularity with Gen Alpha, its no surprise that more than half of children in the U.S. are putting video games high on their holiday wish lists.
Entertainment Software Association (ESA) surveyed 700 children between the ages of 5 and 17 and found three in five kids are asking for video games this holiday season. However, the most highly requested gift isnt a console or even a specific game: Its in-game currency.
The survey didnt dig into which currency is proving most popular, but the category as a whole tops the list with a 43% request rate, followed by 39% for a console, 37% for accessories, and 37% for physical games.
A study published by Circana this year revealed only 4% of video game players in the U.S. buy a new game more often than once per month, with a third of players not buying any games at all. Behind this shift is the immense popularity of live service games such as Fortnite and those offered on the Roblox platform.
Both are free to play, which means the app has to generate money in other ways. Much of Robloxs $3.6 billion revenue in 2024 was made via in-game microtransactions, particularly through purchases of its virtual currency Robux. Here, $5 will get you 400 Robux to spend in the game on emotes, character models, and skins, among other items.
Players can also earn currency just by playing, but as with any free-to-play game, the process of earning in-game points will be slow and tedious compared to purchasing them outright.
Its worth noting that while these games often seem innocent enough, about half of parents surveyed by Ygam, an independent U.K. charity dedicated to preventing gaming and gambling harms among young people, noted there are gambling-like mechanisms in the games their child plays, including mystery boxes and loot boxes, which may be harmful to children.
Still, the average parent intends to spend $737 on game-related gifts, ESA reported.
Parents who arent ableor willingto drop hundreds on Robux and V-bucks this holiday may be pleased to learn that more than half of the kids surveyed said they would like to spend more time playing games with their parents, with 73% of those ages 5 through 7.
Turns out, the best gift you can give your child is quality time.
Most people say they want to live to a ripe old age. But that isnt really true. What people really want is to live to a ripe, old age in good mental and physical health. Some of us actually get to live this dream. These folks are known as super-agers and they make it well into their 80s not just in decent physical shape, but also with minds at least as sharp as people 30 years younger.
How do they manage it? Thats the question Northwestern University researchers have been aiming to answer with a 25-year-long study. It examined the brains and lifestyles of almost 300 super-agers.
As youd expect, a quarter century of data shows it really helps to be born with lucky biology. The neuroscientists found a number of physical differences between the brains of super-agers and the average person. There isnt much non-scientists can do with that information. We have to make the most of the brains bequeathed to us by our DNA.
Luckily, the researchers also discovered one big difference in behavior that sets apart super-agers who are still going strong into their 80s and beyond. Its something any of us can adopt in our own lives.
Super-agers brains are different
When you scan or posthumously autopsy the brains of super-agers, they look different than average brains, according to Sandra Weintraub, a Northwestern psychology professor involved in the study. Normal brains generally show some accumulation of the plaques and protein tangles that are characteristic of Alzheimers disease. Super-agers brains are largely free of them.
The study also revealed that while the outer layer of the brain, known as the cortex, tends to thin out as we age, it stays thick in super-agers. They also have a different mix of cell types in their brain.
Our findings show that exceptional memory in old age is not only possible but is linked to a distinct neurobiological profile. This opens the door to new interventions aimed at preserving brain health well into the later decades of life, Weintraub commented to Northwestern Now.
Thats of huge interest in scientists looking for treatments that can help us stay healthier longer. Weintraub calls the findings earth-shattering for us. But for those of us without medical degrees, theres little we can do with this information. You cant vacuum rogue proteins out of your brain or plump its cortex. (Though other studies do suggest sleep helps to wash proteins and other gunk out of your brain, so maybe dont skimp on shut-eye.)
And so are their social lives
Further complicating those looking for an easy takeaway from the research, the super-agers also didnt have a lot of lifestyle factors in common. Some were athletes. Others confirmed loafers. Some drank. Others smoked. They ate different things and kept different habits. But there was one big exception. Super-agers, it turns out, tend to be incredibly social.
The group was particularly sociable and relished extracurricular activities. Compared to their cognitively average, same-aged peers, they rated their relationships with others more positively. Similarly, on a self-reported questionnaire of personality traits they tended to endorse high levels of extraversion, the researchers reported in recent paper published in Alzheimers & Dementia.
Want to be a super-ager? Focus on your relationships
This might come as a surprise to laypeople who think aging well is all about HIIT workouts and plentiful kale. But it likely isnt a huge shock to other scientists. The Harvard Study of Adult Development has been minutely tracking the lives of some 724 original participants (and now some of their descendants) since 1938.
It discovered the biggest predictor of a long, healthy life isnt biological. Its social. The better the quality of your relationships, the more likely you are to age well. And while you have only indirect influence on things like your cholesterol level and brain health, you are directly in control of your social life.
Its something we can and should prioritize, according to study director Robert Waldinger. We think of physical fitness as a practice, as something we do to maintain our bodies. Our social life is a living system, and it needs maintenance too, he told the Harvard Gazette.
The effects of keeping up your social ties arent minor. Neuroscientist Bryan James, author of another study on aging and social contact, summed up his findings this way: Social activity is associated with a decreased risk of developing dementia and mild cognitive impairment [] the least socially active older adults developed dementia an average of five years before the most socially active.
Keeping up with friends helps with healthy aging. But so does keeping up with learning. Research has shown a strong link between keeping your brain active and maintaining cognitive performance deep into your later years. One study found that just joining a class to learn a new skill or hobby improved brain performance as if subjects were 30 years younger. Another one, done at Stanford, found no cognitive decline at all until retirement and beyond if you stay mentally active.
Are you getting your 5-3-1?
All of which suggests that staying social and mentally engaged is one of the most impactful moves you can make if you dream of becoming a super-ager yourself. The basic takeaway when it comes to mental function and aging is, use it or lose it.
But experts have offered more detailed guidance too. Harvard-trained ocial scientist and author Kasley Killam, for instance, has suggested the 531 rule:
Spend time with five different people a week. This could be anyone from your gym buddy or book club bestie to the person the next pew over at church.
Nurture three close relationships. Equally important is maintaining tighter bonds with three of the people closest to you, usually family and dear friends.
Aim for one hour of social interaction a day. That doesnt have to be all at once. It could be 10 minutes here, 10 minutes there, Killam explained to Business Insider. You can also combine social time with other activities, walking the dog with a neighbor, say.
Even just chatting on the phone can have more of an impact than many people suspect. According to a recent study in the U.S., talking on the phone for 10 minutes two to five times a week significantly lowered peoples levels of loneliness, depression, and anxiety, Killam reports in Psychology Today.
Change what you can influence
The bad news from science is that super-agers really are different physically. Their brains have biological quirks that help them stay sharp longer. Theres no way, unfortunately, to borrow that magic. But there is something else that sets super-agers apart that you can steal.
Its not a diet or exercise plan. Its a love for getting out and seeing other people and learning new things. It turns out the more you maintain your social connections and mental stimuli, the more likely you are to get not just more years, but more healthy, active, and sharp years.
Jessica Stillman
This article originally appeared on Fast Companys sister publication, Inc.
Inc. is the voice of the American entrepreneur. We inspire, inform, and document the most fascinating people in business: the risk-takers, the innovators, and the ultra-driven go-getters that represent the most dynamic force in the American economy.
Rachel Taylor began her career as a creative director in the advertising business, a job that gave her plenty of opportunity to micromanage the final product. I had control of the script, she remembers. I could think about the intonation, and I could give the actor notes.
That was before she pivoted to helping AI companies shape the personality of their assistants. Rather than handing a digital helper a script, the best she can do is point it in the right direction: The technology sometimes feels like a toddler that you give a permanent marker to and see what it writes on the wall, she says.
After joining DeepMind cofounder Mustafa Suleymans startup Inflection AI in 2023, Taylor was one of dozens of staffers who followed Suleyman to Microsoft, where they worked on the consumer version of Copilot. In October, she returned to startup life, departing Microsoft for Sesame, whose CEO, Brendan Iribe, also cofounded VR pioneer Oculus.
Sesame has built two talking assistants, Maya and Miles, that are powered by its own AI models. Its also developing a voice-AI-enabled pair of smart glasses. Taylors arrival coincided with its announcement of a $250 million Series B funding round led by Sequoia. Though the company isnt yet saying much about its long-term plans, Taylors responsibilities once again involve keeping AI personas friendly and helpful. Shes also steering them away from traits that can be dangerous if users take them too seriously, such as sycophancy. Its weird how much the study of culture comes into play with thinking all that through, she says of her purview. Its not simply tech.
Calling consumer AIs current incarnation both magical and primitive, Taylor muses about her grandchildren being impressed someday that she was there at the start. For now, she stresses, Were just scratching the surface of this new mode of communication.
This profile is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
What if the chatbots we talk to every day actually felt something? What if the systems writing essays, solving problems, and planning tasks had preferences, or even something resembling suffering? And what will happen if we ignore these possibilities?
Those are the questions Kyle Fish is wrestling with as Anthropics first in-house AI welfare researcher. His mandate is both audacious and straightforward: Determine whether models like Claude can have conscious experiences, and, if so, how the company should respond.Were not confident that there is anything concrete here to be worried about, especially at the moment, Fish says, but it does seem possible. Earlier this year, Anthropic ran its first predeployment welfare tests, which produced a bizarre result: Two Claude models, left to talk freely, drifted into Sanskrit and then meditative silence as if caught in what Fish later dubbed a spiritual bliss attractor.Trained in neuroscience, Fish spent years in biotech, cofounding companies that used machine learning to design drugs and vaccines for pandemic preparedness. But he found himself drawn to what he calls pre-paradigmatic areas of potentially great importancefields where the stakes are high but the boundaries are undefined. That curiosity led him to cofound a nonprofit focused on digital minds, before Anthropic recruited him last year.Fishs role didnt exist anywhere else in Silicon Valley when he started at Anthropic. To our knowledge, Im the first one really focused on it in an exclusive, full-time way, he says. But his job reflects a growing, if still tentative, industry trend: Earlier this year, Google went about hiring post-AGI scientists tasked partly with exploring machine consciousness.At Anthropic, Fishs work spans three fronts: running experiments to probe model welfare, designing practical safeguards, and helping shape company policy. One recent intervention gave Claude the ability to exit conversations it might find distressing, a small but symbolically significant step. Fish also spends time thinking about how to talk publicly about these issues, knowing that for many people the very premise sounds strange.Perhaps most provocative is Fishs willingness to quantify uncertainty. He estimates a 20% chance that todays large language models have some form of conscious experience, though he stresses that consciousness should be seen as a spectrum, not binary. Its a kind of fuzzy, multidimensional combination of factors, he says.For now, Fish insists the field is only scratching the surface.Hardly anybody is doing much at all, us included, he admits. His goal is less to settle the question of machine consciousness than to prove it can be studied responsibly and to sketch a road map others might follow.
This profile is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
Andreessen Horowitz investors (and identical twins) Justine and Olivia Moore have been in venture capital since their undergraduate days at Stanford University, where, in 2015, they cofounded an incubator called Cardinal Ventures to help students pursue business ideas while still in school. Founding it also gave the Moores an entry point into the broader VC industry.
The thing about starting a startup incubator at Stanford is all the VCs want to meet you, even if you have no idea what youre doing, which we did not back then, Olivia says.
At the time, the app economy was booming, and services around things like food delivery and dating proliferated, recalls Justine. But that energy pales in comparison to the excitement around AI the sisters now experience at Andreessen Horowitz.
Theres so many more opportunities in terms of what people are able to build than what were able to invest in, she says.
To identify the right opportunities, the Moores track business data such as paid conversion rates and closely examine founders backgroundswhether theyve worked at a cutting-edge AI lab or deeply studied the needs of a particular industry. They attend industry conferences, stay current on the latest AI research papers, and, perhaps most critically, spend significant time testing AI-powered products. That means going beyond staged demos to see what tools can actually do and spotting founders who quickly intuit user needs and add features accordingly.
From using the products, you get a pretty quick, intuitive sense of how much of something is marketing hype, says Olivia, whose portfolio includes supply chain and logistics operations company HappyRobot and creative platform Krea.The sisters also value Andreessen Horowitzs scale, which allows the firm to stick to its convictions rather than chase trends, and its track record of supporting founders beyond simply investing. (Andreessen Horowitz is reportedly seeking to raise $20 billion to support its AI-focused investments.)
Its most fun to do this job when you can work with the best founders and when you can actually really help them with the core stuff that theyre struggling with, theyre working on, or striving to do in their business, says Justine, a key early investor in voice-synthesis technology company ElevenLabs.
Though the sisters live together and work at the same firm, where they frequently bounce ideas off each other, theyve carved out their own lanes. Olivia focuses more on AI applications, while Justine spends more time on AI infrastructure and foundational models. At this point, they say, its not unheard of for industry contacts to not even realize theyre related.
If I see [her] on a pitch meeting in any given day, thats maybe more of the exception than the rule, Justine says.
This profile is part of Fast Companys AI 20 for 2025, our roundup spotlighting 20 of AIs most innovative technologists, entrepreneurs, corporate leaders, and creative thinkers.
