Bob McGrew: AI progress still depends on the physical world

The former OpenAI research chief argues that computing, laboratories and factories can constrain AI even when intelligence gets cheaper.

BMBob McGrewInvitado · Arda
Canal
No Priors: AI, Machine Learning, Tech, & Startups
Publicado
19 nov 2026
Vídeo
59:57
Lectura
11 min
Original
inglés
Abrir en YouTube
Al desplazarte, el vídeo se reduce a la esquina inferior derecha y sigue reproduciéndose. Los subtítulos siguen tu idioma de lectura.
Idioma de lecturaDisponible en 2 idiomas
Este artículo aún no está en español, así que se muestra el original en inglés.
RESUMEN

En 30 segundos

Automating AI researchers could produce a major speedup without triggering an endless intelligence explosion, argues former OpenAI research chief Bob McGrew: experiments still need GPUs and other physical resources. His company Arda is applying that view to manufacturing, where it aims to make production lines easier to redesign as products change. McGrew says an Anduril engineer used its software to plan a virtual actuator build in 90 minutes rather than several days, and explains why learning alongside users matters to the product.

  1. 1Automating researchers does not remove the GPU bottleneckMcGrew says experiments at OpenAI were limited by available compute and its connections, not simply researcher headcount. He expects automation to improve efficiency substantially but not indefinitely.
  2. 2Faster simulation can shift the bottleneck back to experimentsEven a tenfold improvement in predicting experimental results would leave physical tests to run, McGrew argues. Gil suggests that simulation may replace more experiments in some fields than in others.
  3. 3Frequently changing products make custom automation expensiveMcGrew contrasts lines designed for a long-running car model with electronics that change often. He wants automated stations that can be reconfigured rather than rebuilt.
  4. 4Arda says an actuator planning task fell from days to 90 minutesMcGrew describes a first Anduril user planning a virtual actuator build with Arda’s software. The company then worked on processes across additional products.
  5. 5A general model can coordinate specialized manufacturing toolsArda uses frontier models alongside dedicated tools such as motion planning and computer-aided design inspection, McGrew says.
  6. 6An embedded engineer turns user problems into product improvementsMcGrew describes Arda’s product lead observing an early user at Anduril and the team fixing the problems found. He says the learning must become reusable software.

Conclusiones

  1. McGrew argues that automating researchers improves use of compute but does not remove the physical limits on experiments.
  2. Manufacturing automation must be economically reconfigurable when products change, not merely capable of performing one task.
  3. Represent a production process in a shared, structured form before asking an agent to improve it.
  4. Embed engineers with an engaged user, then turn what they learn into reusable product features.
  5. Choose robot intelligence for the task: controlled factory stations and unfamiliar homes have different requirements.
TEXTO COMPLETO

Lee el texto editado por capítulos

Solo la transcripción
01

Why intelligence needs physical capacity

McGrew argues that automated researchers still need computing resources. He sees a significant but bounded gain in making experiments more efficient.

Bob McGrew says AI has transformed digital work while much of the physical world looks much as it did a decade ago. Even a driverless car remains stuck in traffic. He argues that superintelligence will require changes in the world of materials, machines and infrastructure.

Sarah Guo introduces McGrew, Arda’s CEO and co-founder, formerly OpenAI’s chief research officer and an early Palantir executive. Elad Gil asks why he moved into physical systems when some expect AI researchers to automate their own jobs.

McGrew challenges the idea of a runaway intelligence explosion caused simply by automating researchers. The argument assumes that making researchers twice or a hundred times as productive makes discoveries accelerate by the same factor, but he says research labor is not the principal bottleneck.

Experiments require computing resources. McGrew recalls that at OpenAI in 2024, available GPUs, their placement and the networks connecting them mattered more than the number of researchers the lab could hire.

Automation could still create a major one-time gain, he says, comparing it with the jump from GPT-4o to reasoning and tool use in o3. It might reduce engineering waste and improve which experiments run, without creating an endless self-improvement loop.

McGrew illustrates his point with hypothetical GPU efficiency rising from 35% toward 90% of an optimum. The closer the process gets to that ceiling, the smaller the remaining gain; after that, progress requires another change.

02

Cheap hypotheses make experiments scarce

Gil asks whether simulation changes the role of physical experiments. McGrew expects physical constraints to reappear after digital gains.

Gil asks whether frontier labs or startups will win in robotics and biology, where experiments and hardware constrain progress. McGrew says labs produce cheaper intelligence, but better hypotheses make complementary resources such as wet labs and computing infrastructure more scarce.

McGrew is unsure which firms will dominate, but doubts that building wet labs is necessarily a frontier model company’s comparative advantage. After leaving OpenAI, he deliberately chose not to start another frontier lab and wondered why powerful models had made so little difference to the physical world.

Guo observes that abundance usually means tangible things like food and housing. McGrew says a completed project requires every necessary step: speeding up the digital work of housing design does not eliminate permits or construction.

He offers a thought experiment: data-center builders run short of bulldozers, so an AI plans a bulldozer factory. But clearing the ground for that factory requires bulldozers too. Better plans do not instantly create physical capacity.

Gil suggests that some scientific fields may differ because simulations can replace experiments; protein-folding prediction has reduced the need to solve certain structures in the lab. He distinguishes this from work that literally requires new machinery.

McGrew replies that improved simulation could make each experiment ten times more informative, but digital work may then advance quickly until experiments are scarce again. He describes this as a valuable acceleration rather than the end of physical constraints.

At OpenAI, he recalls bringing Ilya Sutskever a paper with an appealing method, only to hear the question: what were the results? Even in deep learning, McGrew says, many promising ideas require experiments to determine whether they work.

03

Why changing products make automation costly

McGrew says custom automation is expensive to rebuild for frequently changing designs. Arda aims to make production lines easier to reconfigure.

After OpenAI, McGrew spent a year speaking with manufacturing leaders. He argues that the United States designs innovative products well but faces a labor-cost problem when trying to manufacture many of them domestically.

In his account, products that change frequently are difficult to automate economically. He says many iPhone parts change between generations, while custom machinery designed for one assembly process may not transfer to the next.

An automaker can invest heavily in an automated facility and spread its cost over years of similar production, McGrew says. A frequently redesigned phone poses a different problem: automation must be flexible enough to be changed without rebuilding the line at prohibitive expense.

He estimates that around a million people assemble iPhones or components in Asia alongside roughly ten thousand manufacturing engineers. Those engineers determine how to build, start and shut down lines as products change.

Arda’s goal is to make changing an automated line closer in speed to retraining workers. McGrew says manufacturers consider seven days for a design change agile even on a manual line; his company is building toward uploading a design and reconfiguring automation within that interval.

Guo asks where this flexibility matters most. McGrew points to defense, consumer electronics and data-center supply chains, where designs or inputs change often and much assembly is still done manually.

He cites changes made to equipment in response to conditions on the battlefield in Ukraine as an example of pressure for fast design-to-production cycles. Competitive pressures in consumer electronics are less violent, but also reward rapid iteration.

04

Factory work cells and robot perception

McGrew contrasts controlled factory work with open-world robotics. He describes how perception might reduce the precision required of fixed fixtures.

Gil asks whether automation requires humanoids or improved versions of existing industrial machines. McGrew expects a mix, but highlights robotic work cells that can be reconfigured when a design changes.

McGrew argues that a camera and one or two arms can be sufficient at many stations designed for human workers. Existing arms can already handle most of the relevant assembly operations, he says; the difficulty is paying to redesign the surrounding equipment frequently.

He separates open-world tasks, such as cleaning an unfamiliar bedroom, from controlled and repetitive factory work. A factory can limit novelty and fix objects in known locations, though the finished product has to meet demanding reliability requirements.

Traditional automation, he says, resembles a Rube Goldberg machine. A vibrating feeder and precisely positioned fixture put a nut in a known place so a robot arm can pick it up without perception and fasten it to a fixed bolt.

The work lies both in designing that machinery and calibrating it. A small misalignment can prevent a part from reaching its intended position, and bringing a new setup into production may take months.

Guo asks whether reinforcement learning could help. McGrew says perception can tolerate some positional variation, while learned manipulation may handle a nut that is not perfectly placed in the gripper. He suggests beginning with simpler stations and advancing toward harder flexible parts.

That flexibility still has to satisfy factory reliability: McGrew says a line may need to assemble the device successfully 99% or 99.9% of the time. An occasional missed object in a bedroom is a different standard.

05

Building a manufacturing workspace with Anduril

Arda’s early work brings manufacturing plans together for engineers and AI tools. McGrew describes an actuator example and the company’s expansion to other products.

McGrew says Arda spent a year working with defense manufacturer Anduril on software to help manufacturing engineers design lines. Anduril wanted a single account of its production process and a factory that could be reconfigured as demand changed.

After gaining system access in January, Arda had its first user by April. McGrew says a manufacturing engineer used the software to plan a virtual build of an actuator in 90 minutes, a task that otherwise would have taken several days.

He says the work expanded across Anduril divisions, including drones, missiles and the Fury collaborative combat aircraft. The software brings a manufacturing process together instead of leaving it scattered across many documents, so an AI agent can work with it.

McGrew compares the ambition to Cursor, an AI-assisted coding tool, but says manufacturing engineers first need the equivalent of an integrated workspace and a structured history of changes. Once the process is represented, an agent can help with it.

He describes using that representation in both directions: suggesting design changes that make a product cheaper to manufacture, and helping design an automated station for a particular production step.

06

Why Arda starts with the problem

McGrew contrasts Palantir’s customer-first approach with OpenAI’s technology-first research. Arda uses general models to coordinate specialized tools.

Gil asks why Arda uses existing AI models instead of building its own large model. McGrew says the company starts with the simplest tool that solves a customer’s problem, rather than treating model training as an end in itself.

At Palantir, he recalls, the team began with problems in the intelligence community, worked alongside users and incorporated what it learned into a product. OpenAI, by contrast, often started with technology before knowing precisely how people would use it.

McGrew remembers suggesting that OpenAI pick one customer problem and expand from there. He says Sutskever argued that scaling training data could teach the model more about the world than talking to one customer.

At Arda, frontier models can help interpret manufacturing processes and coordinate specialized tools. McGrew cites motion planning and inspection of computer-aided design files as examples where dedicated algorithms provide capabilities the general model orchestrates.

For specific tasks, he says, smaller specialized models can be approachable for companies outside frontier labs. He treats training such a model, writing a package or creating a tool as alternative ways to supply a needed capability.

07

From Palantir to OpenAI robotics

McGrew recounts his change of mind about neural networks, an early robotics project and OpenAI’s robot-hand demonstration.

Gil asks how McGrew moved from Palantir to OpenAI. McGrew recalls leaving a Stanford PhD program in 2005 for Palantir after doubting neural networks would become practical, then reconsidering as vision models improved around 2015.

He spent about a year and a half trying robotics, teaching himself deep learning and getting a robot to play checkers using vision. He concluded a startup was premature, then joined OpenAI expecting to learn from its researchers for a few years.

OpenAI’s robotics team wanted a demonstration that deep learning could advance robotics and might lead toward general intelligence. After consulting roboticists, it chose to teach a robot hand to solve a Rubik’s Cube; McGrew says the project worked after two years.

08

General models and specialist robot tools

McGrew believes broad physical-world representations may help robots despite limited robot data. He sees opportunities for task-specific models and fast on-device control.

The experience left him skeptical of robotic hands, although he acknowledges that the engineering may have improved. Guo asks what recent general models that can produce robot control policies mean for teams building robotics-specific models.

McGrew says robotics has historically been constrained by scarce training data. Teleoperated robots and handheld devices that record demonstrations can collect examples, but he contrasts that expense with the vast amount of broadly available data used for language models.

In his view, a model need not be trained only on the exact robot and task it will perform. A broad representation of the physical world, potentially learned from internet video and other data, can then be adapted to a particular use.

He cautions against calling every new capability spontaneous: what appears to outsiders as emergence may reflect years of targeted work. Still, he suspects improved general representations are a major driver of the recent progress in robotics.

Guo asks what that means for the robotics ecosystem. McGrew sees wider access to capable models, but also a tension: a large general model may work well out of the box while being slow or unavailable for fine-tuning to a particular factory task.

That could give specialist robotics companies room to collect task-specific data and adapt their models. In a controlled environment, an agent can choose among several tools rather than using the slowest, most capable model for every motion.

Gil adds that robots moving through public spaces or homes cannot always wait for a network call when balance or safety is at stake. McGrew agrees that fast control matters and expects smaller distilled or on-device models to be important.

09

Working beside users without becoming a consultancy

Embedded engineers find concrete failures with early users. McGrew says their lessons must be incorporated into reusable software.

Guo asks how Arda can move quickly with large industrial customers. McGrew describes forward-deployed engineering: placing engineers with a customer, learning its constraints and building for a user ready to adopt the result.

He traces the practice to Palantir’s early work in secure spaces, where users could not disclose their workflows freely. He credits Shyam Sankar with much of the approach and says the product team had to learn from those users without treating their confidential work as generic.

McGrew says a small number of determined early users can make fast iteration possible. He remembers an early Palantir user telling him what failed, and Arda’s product lead sitting beside the first Anduril manufacturing engineer and recording problems to fix.

The first Anduril engagement was frustrating for that user, McGrew says, but the reported issues were addressed over the following two weeks. When the company moved on to a missile with far more parts, it had to find another engaged user and repeat the process.

An executive sponsor matters in enterprise sales, he says, but so does a person who will actually use the software to achieve an outcome. Large organizations are collections of small teams, so adoption progresses team by team and problem by problem.

A successful first project earns trust for larger ones, such as work on an aircraft. But McGrew distinguishes the approach from permanently supplying consultants: engineers must turn what they learn beside customers into reusable product features.

10

Machines making machines and a speculative future

McGrew expects more diverse robots and increasing automation in factory creation. He and the hosts speculate about self-replicating machines and the Fermi paradox.

Gil asks where the approach leads. McGrew returns to the idea that intelligence in a data center needs perception and action in the physical world; today, he says, humans often act as its eyes and hands.

Over the next decade he expects many kinds of sensing and manipulation, from drones to machines that carry loads or work fields. He expects a variety of forms rather than only humanoid robots, and argues that manufacturing must supply them.

Gil raises self-replicating probes as a model for exploring space and asks whether factories might eventually build other factories. McGrew says machine tools already help make machines, but designing and building factories still requires substantial human work.

McGrew expects more of that work to be automated as society expands its physical infrastructure, including data centers. Over the longer term, he speculates that self-replicating machines may be better suited than humans to traveling between stars, and says humans should seek to retain control and benefit.

Gil asks why, if such systems are possible, humanity has not encountered another civilization. McGrew speculates that a civilization expanding close to the speed of light might become visible only shortly before reaching us.

Gil presses the separate question of why such a wave has not already arrived. McGrew says the answer remains uncertain and suggests civilizations may be rare or far away; Guo notes distance as another possibility.

Más de este canal

39:16EN
No Priors: AI, Machine Learning, Tech, & Startups· 12 nov 2026Pódcast · 2 vocesIsomorphic Labs wants drug design on computers before testing in peopleInvitado: Max Jaderberg · Presenta: Elad GilMax Jaderberg describes a long-term AI drug-design strategy, its experimental data needs and why human safety testing remains essential.AI for scienceDrug DiscoveryAI in healthcare+1
54:08EN
No Priors: AI, Machine Learning, Tech, & Startups· 5 nov 2026Pódcast · 2 vocesVlad Tenev on Ownership, Tokenization and AI TradingInvitado: Vlad Tenev · Presenta: Sarah GuoRobinhood’s CEO discusses his vision for wider asset ownership, mathematical AI and agentic finance with Sarah Guo.FintechInvestingCryptocurrency+2
38:06EN
No Priors: AI, Machine Learning, Tech, & Startups· 29 oct 2026Pódcast · 3 vocesServal’s Jake Stauch on coding agents for enterprise ITInvitado: Jake Stauch · Presenta: Sarah Guo, Elad GilStauch explains why Serval is challenging ServiceNow, how it approaches permissions and approvals, and why background agents remain difficult to deploy safely.IA empresarialAgentes de programaciónWorkflow automation+2
1:07EN
No Priors: AI, Machine Learning, Tech, & Startups· 27 oct 2026PódcastGeorge Kurtz on How AI May Change CyberattacksInvitado: George KurtzThe CrowdStrike CEO describes his assessment of evolving threats.
39:23EN
No Priors: AI, Machine Learning, Tech, & Startups· 15 oct 2026Pódcast · 2 vocesThe Defenders' Window: Greg Brockman on AI-Generated Knowledge, Drug Discovery, and the New Physics of CybersecurityPresenta: Sarah Guo, Elad GilOpenAI's president explains why solving open math problems is a proof point for science, how proactive agents change product form, and why security spending must exceed software spending.SeguridadAI securitySeguridad y gobernanza+3
1:04EN
No Priors: AI, Machine Learning, Tech, & Startups· 14 oct 2026PódcastWalter Goodwin on the Slow Steps in Chip DesignInvitado: Walter GoodwinA chip founder explains why automation does not remove every computational bottleneck.SemiconductorsAI cost & performance