The Researchers Shaping Today's AI: Who's Who
Five names, five short profiles — written because the people behind the papers matter as much as the papers, and because almost every claim below traces back to the researcher's own page, their own paper, or an official bio. Where a fact couldn't be verified that way, we left it out.
Karol Hausman — Physical Intelligence
Karol Hausman is co-founder and chief executive of Physical Intelligence, the San Francisco startup building generalist AI models for robots. He has described the goal plainly: "It's a single generalist brain that can control any robot," he told the New York Times when the company announced a $400 million raise in November 2024. Before founding the company in 2024 he spent six years at Google — as a staff research scientist at DeepMind and, earlier, Google Brain, where his own CV lists him as robot manipulation lead — and he taught as an adjunct professor at Stanford, according to his own site. He took his PhD at the University of Southern California and his master's at the Technical University of Munich.
Three pieces of work define his research record. SayCan ("Do As I Can, Not As I Say," CoRL 2022, which won the conference's Special Innovation Award) paired a large language model with a robot's learned skills, so a planner could only propose things the machine could actually do — co-authored with Chelsea Finn and Sergey Levine, the two others profiled below. Open X-Embodiment (2023) pooled robot-learning data from many institutions into shared datasets and retrained Google's RT-1 and RT-2 models on the mixture, producing RT-1-X and RT-2-X — generalist models that run across different robot bodies rather than one custom rig. Q-Transformer (CoRL 2023) scaled offline reinforcement learning for manipulation tasks through autoregressive Q-functions. He also co-authored Robotic Skill Acquisition via Instruction Augmentation with Vision-Language Models (RSS 2023), which used language models to expand the instructions robots could be trained on — with Levine again in the author list. Before the foundation-model era his work ran in a different direction: active perception, jointly estimating state and controlling flight, including quadrotor self-calibration and navigation in the International Journal of Robotics Research, as his own research list records.
Why they matter now: Physical Intelligence raised $600 million at a $5.6 billion valuation in November 2025, Bloomberg reported, and was in talks the following March for roughly $1 billion more at over $11 billion, per Bloomberg via TechCrunch. The company's π0.5 model, released in April 2025, claims open-world generalization — controlling a robot in a kitchen it has never seen. Hausman is the executive betting that language-model scaling recipes transfer to the physical world.
Chelsea Finn — Stanford and Physical Intelligence
Chelsea Finn is an assistant professor in computer science and electrical engineering at Stanford, where she runs the IRIS lab — a group studying intelligence through robotic interaction at scale — and a co-founder of Physical Intelligence. Her Stanford bio lists her as the William George and Ida Mary Hoover Faculty Fellow and credits her work with pioneering end-to-end deep learning for vision-based manipulation, meta-learning for few-shot learning, and scaling robot learning to broad datasets. She studied electrical engineering and computer science at MIT, took her PhD at UC Berkeley under Pieter Abbeel and Sergey Levine, and later worked as a research scientist at Google Brain and Google DeepMind. Her stated interest, on her own page, is the capability of robots and other agents "to develop broadly intelligent behavior through learning and interaction."
Her best-known contribution is Model-Agnostic Meta-Learning (MAML), published at ICML 2017 — an algorithm that lets a model adapt to a new task from a handful of examples without retraining from scratch, and model-agnostic in the paper's own sense: it works with any model trained by gradient descent. The ICML proceedings list more than 20,000 citations, and it became the standard reference point for few-shot learning across the field, not just robotics. At Google she co-authored MT-Opt (2021), a multi-task learning system that trained a repertoire of manipulation skills at scale, and she was part of the SayCan team. Recent work continues on both fronts: Hi Robot, a hierarchical vision-language-action system published at ICML 2025 (with Levine), lets a robot handle compound instructions and mid-task corrections — evaluated on three robot platforms doing tasks from cleaning messy tables to grocery shopping — and MemER (2025) from her Stanford lab gives robot policies a memory buffer to retrieve from. Her bio also lists the Presidential Early Career Award for Scientists and Engineers, a Sloan Fellowship, and the ACM doctoral dissertation award, and her teaching record includes Stanford's graduate courses in meta-learning (CS330) and deep reinforcement learning (CS224R).
Why they matter now: Finn is one of the few people running both a leading academic robot-learning lab and a founding role at a generalist-robotics company, so her lab's agenda — memory, feedback, interaction at scale — is a live read on what today's commercial robot models still lack.
Nicholas Carlini — Anthropic
Nicholas Carlini is a research scientist at Anthropic, where — in his own words — he studies "what bad things you could do with, or do to, language models." One correction is worth making up front: he is no longer at Google DeepMind. He left in March 2025 after seven years — Google Brain from 2018, DeepMind after that — writing that he had come to believe DeepMind's leadership "no longer … supports the type of high-impact security and privacy research I like to do," and that getting papers approved had become considerably harder after the Brain–DeepMind merger. He joined Anthropic initially for a year; his current homepage still lists him there. He holds a PhD from UC Berkeley under David Wagner.
His research record is essentially a catalogue of ways machine-learning systems break. The Carlini & Wagner attack, described in a 2017 IEEE S&P paper that won a Best Student Paper award, became the standard method for evaluating whether a defended model is actually robust. Along the same lines, his obfuscated gradients paper (ICML 2018, one of the best-paper awards he lists) showed that a run of published defenses looked robust only because they hid gradient information from attackers — the critique reset how the field judges defenses. Audio adversarial examples (2018) showed hidden voice commands could fool speech-to-text systems — the New York Times and BBC both covered it. Extracting Training Data from Large Language Models (USENIX Security 2021, with ten co-authors) showed that large language models leak their training data, recovering hundreds of real training examples from GPT-2. His group later bought expired domains matching image URLs inside web-scale datasets and poisoned the training data of anyone who had scraped them — an attack the paper's own title calls practical. A 2024 paper on extracting neural-network parameters won a best paper at Eurocrypt.
Why they matter now: on his site, Carlini says his recent talks focus on language models that find thousands of vulnerabilities in real code and chain them into complex exploits — the clearest available barometer for whether shipping coding agents is a security decision or just a feature decision.
Sergey Levine — UC Berkeley and Physical Intelligence
Sergey Levine is an associate professor in electrical engineering and computer sciences at UC Berkeley, where he holds the Harvey E. Wagner Presidential Chair in Artificial Intelligence and leads the Robotic Artificial Intelligence and Learning (RAIL) lab, per the EECS faculty page. He took his BS, MS and PhD at Stanford (2009–2014) and joined Berkeley's faculty in 2016. His own research page frames his interest broadly: general-purpose methods that let any autonomous system learn to solve any task — his formal bio on that page puts it as end-to-end neural policies that combine perception and control, mapping sensory input straight to action. He teaches UC Berkeley's CS285: Deep Reinforcement Learning, which has become a standard free curriculum for the field, and his talk archive includes a lecture making the case for real-world reinforcement learning.
Levine's signature contributions run through the modern RL toolkit. Soft Actor-Critic (ICML 2018, with Tuomas Haarnoja and Pieter Abbeel) added an entropy term to off-policy learning, trading exploration for stability, and became one of the field's default algorithms. With Aviral Kumar and colleagues he wrote Conservative Q-Learning (NeurIPS 2020), which learns useful policies from static datasets by penalizing over-optimistic value estimates, and the D4RL benchmark that gave offline RL a common yardstick — his own site now teaches the whole family, CQL, AWAC and IQL, side by side. Earlier, with Finn and Abbeel, he co-authored MAML; he appears across the Google robotics lineage too (SayCan, Q-Transformer, MT-Opt, where he worked alongside Hausman and Finn) and was a co-author of Hi Robot at ICML 2025. His through-line has stayed constant since his Stanford PhD: give any autonomous system a general way to learn a task from data, rather than hand-coding one skill at a time.
Why they matter now: Levine co-founded Physical Intelligence alongside Hausman and Finn, and his framing of the company's ambition is the one everyone quotes: "Think of it like ChatGPT, but for robots," he told TechCrunch in January 2026, when the company had about 80 employees and had raised just over $1 billion.
Alex Atallah — OpenRouter
Which Alex Atallah? The one behind this profile is the chief executive and co-founder of OpenRouter and, before that, co-founder and chief technology officer of OpenSea — an infrastructure entrepreneur, not a security researcher. A search pairing the name with OpenAI security turns up no such figure at that company, so if you arrived here from that direction, this is the Alex Atallah in question.
His own site lists the résumé: Stanford, Palantir (natural-language search and cyber tooling), Y Combinator and the HF0 AI residency, then OpenSea, the NFT marketplace, where by his LinkedIn account he helped grow it past $4 billion in monthly volume — before stepping back from day-to-day work in July 2022, as he announced at the time. After ChatGPT arrived he built Window AI in April 2023, a browser extension that let users choose which model a web app used, and that experiment turned into OpenRouter, which he co-founded later that year with Louis Vichy — a single API and marketplace that routes requests across hundreds of models and providers. He is also a researcher in a more literal sense: he co-authored State of AI: An Empirical 100 Trillion Token Study with OpenRouter with Andreessen Horowitz collaborators (published December 2025), one of the largest empirical looks at how people actually use models — covering open-weight adoption, coding and roleplay workloads, and the rise of agentic inference. He writes about the job occasionally too; a 2024 essay on his site argued that running AI systems well looks more like running experiments than writing code.
Why they matter now: Stripe announced on August 19, 2026 that it agreed to acquire OpenRouter; Bloomberg reported the price at more than $7 billion (per Yahoo Finance's report). The layer Atallah built — neutral routing between model providers — is now part of payments infrastructure.
If one name here changed jobs tomorrow, which profile would you want rewritten first? Tell us in the comments.
