After 16 years and 10 months -- I'm finally writing my first...X? Also, will start writing more regularly. My first post is, ironically, on why I'm posting. qy.co/new

*Fantastic* talk hot off the press by the formidable Jon Gray from @blackstone on what is happening in technology world. Actually watched the whole thing. Chock full of interesting stats: 1/ BX-linked companies went from $25M to $525M in annualized spend with Anthropic in one year (21x) 2/Hyperscaler capex is up 9x in five years; chip-company capex hasn't doubled 3/$55B per gigawatt, all-in (power + data center + chips) 4/ SK Hynix trades at ~4x earnings after a 500% run; Cisco was at 150x in 2000 5/ PE software deals are down 66% (!) piped.video/4rhShOPHl50

Does anyone have strong thoughts on this? There is this strange information cascade that is happening: everyone's spending on tokens because everyone else is spending on tokens, assuming somebody upstream did the math. I'm a bit skeptical on this 'agents run freely are creating huge value for us'. We have 1,000+ engineers and ...it's not always clear. There are some actions you can take that can cost $1 or $100 and you just don't know, and the ROI could be wildly positive or negative and require a bunch of cleanup. And the clean up matters if you're pushing to actual safety systems, which is most physical AI. We're definitely in this weird time period where the cost controls and ROI are all over the place and few people up and down the stack have incentive to rein it in.

At this year’s Agentic AI Summit hosted by @BerkeleyRDI, our Chief Scientist, @Wei_ZHAN_ shared some of the cutting edge research he has been leading in the physical AI space. As the industry races to deploy L2++ ADAS with imitation-learning-based E2E, Wei tackled a key question: what if end-to-end autonomy could be trained without imitation at all, relying only on reinforcement learning? An idea that runs against the current status quo. Here are some of his insights 🧵

Qasar doesn't post often, but when he does 🔥

When I was at Scale in 2016, self-driving was the first place where I saw how hard it is to make AI work reliably in the real world. A decade later, we've gone from autonomy demos to autonomy at national scale! @AppliedInt deploying thousands of autonomous trucks with @HUMAIN is just the beginning. Physical AI will become embedded throughout the infrastructure around us, and the Applied Intuition team will be at the forefront of this shift. Congrats @qasar & team!

I didn’t think autonomy would get nationwide-scale this fast... but here we are 📷! Our work with @HUMAIN and @TareqAmin_ will bring thousands of trucks to KSA. But this is just the start – followed by robotaxis, ports, mining, construction, and more. The size, scale and scope is unprecedented! Letssss go! 🚀🚀🚀

3. The Lenny 100 To help you prioritize your search, I’ve curated a list of the top 100 companies I’d most want to join if I were looking for a job. These are companies that IMHO have the highest talent density, are tackling the most ambitious problems, have the biggest upside potential—and are actively hiring. I’ll update this list regularly: lennysjobs.com/lenny100

preach!

the models have no moat (OpenAI, Anthropic, XAI) the IDEs have no moat (Cursor, Windsurf) the harnesses have no moat (Cognition, Factory, LangChain) the app builders have no moat (Replit, Lovable, Bolt) the wrappers have no moat (Harvey, Abridge, OpenEvidence) the inference providers have no moat (Together, Fireworks, Groq) the voice layer has no moat (Sierra, Decagon, ElevenLabs) the data labeling companies have no moat (Scale, Surge, Mercor) the AI infrastructure has no moat (Baseten, Modal, Railway) the neoclouds have no moat (CoreWeave, Lambda, Crusoe) the generative media companies have no moat (Runway, Higgsfield, Suno) apparently nobody in AI has a moat except the venture firm ☠️

Physical AI Will Be Bigger Than Digital @Qasar (Qasar Younis) (CEO) & Peter Ludwig (co-founder), @AppliedIntuition, interviewed by @pmarca (Marc Andreessen) & @ErikTorenberg (Erik Torenberg) (The @a16z Show) Summary: Qasar Younis and Peter Ludwig argue that the biggest companies of the next 25 years will come from putting AI into physical machines, not software. Applied Intuition sells the intelligence and the tools that go into cars, trucks, tanks, drones, mining rigs, and humanoids, and it just launched Dana, a platform meant to let a high schooler build an autonomous system. If they are right, the trillion-dollar winners of the AI era look more like Amazon and Apple than like companies that optimize ads, and the binding constraint is no longer the technology but how you deploy it into a physical world that punishes mistakes. 1. Physical over digital. The companies that put AI into the physical world will be bigger than the ones that put it into the digital world. Digital AI writes software, optimizes ads, and makes videos; physical AI runs manufacturing, mining, logistics, and supply chains, where the global economy actually lives. Automotive alone is about 3% of global GDP, and it is already a minority of Applied's business. Younis frames it by hindsight: the internet's real monoliths turned out to be Amazon and Apple, not the early surveying and analytics sites. 2. The jobs nobody wants. The work autonomy targets is brutal and already going unfilled. Mining is 1% of the global labor pool and 8% of work-related fatalities, with a death at most major mines once or twice a year. Long-haul truckers live about 10 years less than their peers, with melanoma on the left arm from the window and no way to sleep or eat well on the road. The average American farmer is 58, under 10% are younger than 35, and demand for food and materials keeps rising as the humans keep leaving. 3. The constraint is the human. Remove the person from the cab and the machine gets smaller, cheaper, and shaped in ways a human body never allowed. Underground, the binding constraint is simply that a human needs to breathe. The bigger prize is system-level intelligence: wire up a whole port or mine so the machines talk to each other, predict a brake failure before it happens, and keep running when one unit goes down. A human-driven fleet cannot even tell you a machine is failing, because nobody is plugged into it. 4. The horizontal chipmaker. Applied runs like a silicon company that happens not to make silicon. It sells intelligence and integration into other people's machines and lets the manufacturer badge the result, because whoever owns distribution owns the business. Self-driving trucks already carry commercial loads in Japan, but the brand on the truck is Isuzu, and once Applied is embedded, "it's really hard to take us out." Younis's example is Nvidia: Jensen's edge is knowing his customers as much as making hard chips. 5. GM is a nation state. Killing Cruise was a multivariate decision, not a failure of nerve. Three of the top five consumer lawsuits in US history are automotive, so "quality is job one" and safety is drilled into every process. Cruise ran into union negotiations, a business model built on personal car ownership rather than robotaxis, and a serious injury that had to be handled precisely with regulators. Younis's read: the people running these companies are not dumb, and a version where Cruise survives inside GM is entirely plausible. 6. Timing beats technology. Almost everyone eventually figures out the technology; the fortune is in when and how you ship it. Two years early and you are doomed, two years late and the competitors have arrived. Self-driving is now an engineering grind toward dollars per mile, and once it gets cheap enough every OEM adopts it, the way navigation went from a $4,000 option to free. Younis expects robotaxis to feel routine in the 200 biggest US cities by roughly 2028 to 2030. 7. Heartstrings versus calculator. Ludwig's line: "You buy a car with your heartstrings. You buy a truck with a calculator." Trucking is pure dollars and cents, so an autonomy vendor has to prove the savings on every mile to an unsentimental buyer who already has drivers on staff. That is why Applied picked Japan for trucking, where an acute labor shortage and shrinking population make the math obvious. The consumer robotaxi gets a market-cap premium for its story; the truck gets none, and that changes the strategy. 8. Autonomy for a ninth grader. "There's no reason autonomy should be this obscure, difficult, alchemistic technology." Dana is Applied's agentic platform for building physical AI: define the requirements, auto-generate the scenarios, pull training data, deploy to the machine, and close the loop when the robot hits a wall. Workflows that took days or weeks now run in minutes, the same collapse in effort that Claude and modern IDEs brought to software. The stated bar is a high schooler building a delivery robot for their own campus on a weekend. 9. Enabling competitors is fine. Handing rivals the tools to build autonomy is a feature. Younis points to Google, which armed the whole web with tools and still won through search and YouTube, and expects the same pattern in physical AI. Push the cost of building toward zero and you get far more kinds of machines than anyone would fund today: the leaf-picking yard bot, the dog-poop bot, humanoid entertainment, home care. The real killer apps are as hard to imagine now as Instagram was before phones had cameras. 10. The onboard model is the moat. The frontier labs have it easy, because a trillion-parameter model is allowed to be slow. Physical AI runs in real clock time with milliseconds to act and hard safety and determinism constraints, so the big models stay off-board while a small, constrained model runs on the machine. Meeting all of those constraints at once is the hard part, and Ludwig is blunt that it is also the moat. A perfectly aligned world-model simulator would, in his words, roughly solve the universe, so the value is in getting closer without ever arriving. 11. Sovereign AI is physical AI. As geopolitics fractures, every country will want this kind of AI localized, and physical autonomy draws far more resistance than social media or ride-sharing ever did. Applied operates as a global horizontal provider, with 18 offices everywhere but China, and has learned to collect proprietary data in Korea, the Middle East, and Latin America where outsiders are unwelcome. Hundreds of petabytes of proprietary data plus its own synthetic-data and neural-sim tools compound into systems others cannot easily copy. Younis reaches for Standard Oil and Aramco: serious companies have always been built around the geopolitics of their time. 12. The optimist's burden. Younis refuses the fearful crouch: no single hand can block the sun, and the sun is technological progress, so a society that will not embrace it gets left behind. If autonomy scares you, the responsibility is yours to learn the technology, because someone else builds it regardless, whether that is China or "maybe the Uzbeks." He rejects both lazy poles, that corporations are evil and that technology will be flawless, and lands in the middle: net-net it makes society better, because people stop dying on the roads and food and energy get cheaper.

Speak the truth. Peter in a rare public post!

.@AppliedInt Co-founder and CTO Peter Ludwig says "a billion machines will become autonomous or intelligent over the next ten years." "Cars, trucks, tractors, mining haulers, defense systems, warehouse robots, humanoids—the physical economy will be rebuilt around software that perceives, decides, and acts." "The prevailing assumption about how we get there goes something like this: models keep improving, world models mature, foundation models for robotics arrive, and autonomy falls out the other end. Intelligence is the whole game; scale the intelligence and the machines will follow." "The contrarian bet, then, isn't against intelligence. It's that the next order of magnitude in physical AI comes from making the engineering system as intelligent as the models it carries." Full piece on compounding systems and Applied Intuition's newest platform, Dana: a16z.news/p/the-next-ai-moat…

Dana!

Every AI company (robotics, bio, etc) is trying to master making their own data flywheel. Here's our view into how we've thought about building it up starting with autonomy

Building the autonomy flywheel—data → sim / train / eval → insights— is incredibly hard, takes deep domain expertise and years to get right. Dana brings a decade of Applied Intuition's infra, tooling, and data expertise into a single agentic platform to solve autonomy's hardest problems: 🔷 Curate datasets from petabytes of fleet data, across every domain 🔷 Reconstruct scenes with Gaussian splatting and run neural simulations 🔷 Augment behaviors with RL-trained agents and generate diverse sensor data with world models 🔷 Query results and build dashboards that turn every result into action Our PM, Gautham Sholingar, gives a peek inside ⬇️

from ur talk

We want to make it much, much easier to make physical AI

About to go live w/ @MTSlive!

Speaking the truth @pmarca!