Tony Lu | 陆泓渊

The biggest bottleneck in robotics isn't intelligence anymore, but the software connecting intelligence to physical action. A new paper showed GPT-6 Astra improving from 22.5% to 57.7% success on simulated RoboDojo tasks, without changing or fine-tuning the model. AI could often figure out what to do, but struggled with how to execute it. In one experiment, removing the difficulty of controlling the robot arm and allowing direct object movement increased success to 76.2%. So researchers built Auriga, a software layer that gives the AI better 3D measurements, previews of planned movements, and the ability to detect errors and recover quickly. What does it mean for the robotics industry? Instead of spending billions training smarter foundation models, companies might unlock substantial improvements by building better infrastructure around existing AI. But there's a catch: The experiments were conducted in simulation, with the environment paused while the AI thinks. Furthermore, tasks still took an average of 18.8 minutes, and the system failed roughly 42% of attempts. This shows that robots are still far from commercially ready. The next breakthrough might come from helping existing intelligence work more reliably rather than training a smarter model.

An AI just debugged another AI by directly editing its weights. MIT researchers used GPT-6 Astra to modify just 0.016% of Qwen3.5-0.8B’s parameters. Its MATH-500 score jumped from 7.9% to 48.1%. Traditionally, fixing AI models involves retraining them, often adjusting millions of parameters through gradient descent. But what if AI could identify and repair specific problems directly? That's exactly what researchers demonstrated. The surprising part? The smaller model wasn't necessarily bad at math. It was getting stuck in repetitive reasoning loops. Astra identified and corrected the issue. This could have tremendous implications for AI economics. If powerful models can diagnose and repair smaller models, we could unlock significantly better performance from cheap AI without training entirely new models. The next frontier might not just be building smarter AI, but making existing intelligence dramatically cheaper.

The real bottleneck in physical AI isn’t more data. It’s better embodied data. Here’s a breakdown of different levels of data quality and what the future of robotics looks like. Level 1: Internet videos This is the lowest quality data: footage of humans doing tasks on YouTube or TikTok. This is useful for visual priors. Robot can learn what a task looks like and the sequence of it. However, this type of data is high scale but low embodiment. The robot can’t feel force, motor intent or action trajectory. Level 2: Human Teleoperation This is where a human control the robot arm or humanoid remotely. This data is better than passive video. But there’s still a gap: the human operator is using a VR headset or some type of interface. They can’t necessarily feel what robots “feel.” Level 3: Simulation This is where robot learns in a simulated environment through millions of attempts. Nvidia is heavily invest in this because real world data is scarce, expensive and slow. The only problem is that there’s still a sim-to-real gap: messy lighting, deformed objects, and edge cases in real life can throw robots off guard. Level 4: Embodiment-matched human data 
This is where humans perform tasks while wearing motion/eye tracking devices, hand sensors, etc. This reduce the gap between human motion and robot motion. However, this is hard to scale as it’s limited by the number of human training hours. Level 5: Full robot-human sensory loop This is where the industry is headed. This level creates the highest quality data where humans feel what robots feel. The human operator partially inhabits the robot body, creating the lowest possible embodiment gap. The future of physical AI belongs to the company that build the infrastructure for human intent, robot motion, sensory feedback, task outcome, and real-world failure — all captured in one loop.

Could JEPA be the next stage of AI models? For a long time, AI models have been trained by predicting the missing piece of data. In language, it predicts the next token. For images/videos, it predicts the next pixel or frame. This is where Yann LeCun believes is highly inefficient. A model can spend enormous compute learning surface-level details while only indirectly learning the deeper structure: objects, forces, causality, and consequences. This is where Joint Embedding Predictive Architecture (JEPA) comes in handy. Instead of asking what pixel comes next, it asks what the next meaningful state of the world should be. Humans do not understand a ball falling in pixels. We understand it because we see an unsupported ball pulled by gravity, hits the floor, and bounces. JEPA is trying to make AI learn that kind of abstract world model. After taking in raw data, instead of predicting every pixel, JEPA predicts the hidden meaning/state: ball position, motion, gravity, object relationship, etc. The LeWorldModel has around 15 million parameters, trains on a single GPU, learns in hours, and can plan much faster than even large foundational models. This abstract representation is called a latent space. This means the model won't memorize every detail of the data. Instead, it will learn the hidden structure that explains what's happening. This is very exciting for the future of physical AI. Robots and autonomous agents need models that can predict consequences and plan through reality. This way, robots are less scripted to perform reliably even in messy edge cases. For AI agents, they can reason cause and effect and act more like decision engines than chatbots. This is an exciting new direction.

Robotics may be close to its real breakthrough moment. Not because a humanoid can do a backflip. But robotics is finally moving from programmed motion to learned physical intelligence. The real challenge has always been manipulation: touch, force, timing, pressure, etc. That is why GENE-26.5 is interesting. They are trying to solve the biggest bottleneck in robotics: obtaining high-quality physical-world training data. Genesis is building a system where human skill can be captured, converted into robot training data, and transferred into real robotic action. A noninvasive data glove collects motion, force, and touch data while a human performs a task. A 1:1 human-like robotic hand makes that data easier to transfer because the robot hand is shaped and moves like a human hand. If they can solve human motion data, tactile feedback, and real-world transfer, they will own the data engine for physical intelligence.

Machines like this are where manufacturing is headed. The main breakthrough is reducing the need for expensive custom tooling every time you need to make a new shape. This matters most in industries where parts are high-value, complex, and low volume. For example: aerospace, defense, robotics, and prototyping. Traditional manufacturing optimizes for scale. However, the future will require more flexibility. We will want faster design changes, shorter production runs, and more customized parts. This is why manufacturing is becoming more software-defined. The manufacturing advantage comes from how quickly you can turn a digital design into a functional physical object.

Highly specialized machines like this will transform the battlefield faster than most militaries can rewrite their doctrine. The real robotics revolution in warfare will not look like sci-fi humanoid terminators. They will be narrow-purpose systems that do one important job really well, such as deploying drones, targeting vehicles, carrying supplies, or providing firepower. The point is simple: to keep humans far from the frontline while increasing the tempo of operations. Too many armies today are still built around Cold War/Gulf War military doctrine. They build large platforms, centralized command, massive military equipment, and high human exposure. However, cheap semi-autonomous systems are changing the unit economics of war. The future of the battlefield will be a competition of superior sensors, software, robotics, and technological adaptation to local geography.

This is why the future of robotics is Physical AI. Robots need to be flexible enough to handle edge cases, not just follow fixed paths in perfectly designed environments. The next unlock is not just better hardware. It is robots that can perceive, reason, coordinate, and adapt in real time. Factories no longer have to be redesigned around robots. It's robots that adapt to the factory. That is when robotics moves beyond Amazon-scale warehouses and becomes profitable even for mid-sized warehouses and manufacturers. The market will grow tremendously when robots can be reliably deployed across millions of mid-sized warehouses, factories, hospitals, and construction sites that cannot afford custom automation teams.

The defining bottleneck in AI infrastructure is not just compute. Its memory and interconnect. The expensive part is not the math itself, but moving data to where the math happens. It is like having a factory with incredibly fast machines, but the worker must walk miles to transport raw material to the line. The GPU may have enormous compute capacity, but if memory can't feed it data fast enough, that capacity goes underused. Many people worry that AI will hit an energy bottleneck. But if we reduce the energy wasted moving data between memory, chips, and servers, the same power supply could support far more compute. This is why the next frontier is not only better chips. It is better memory and better interconnect A simple comparison shows the gap: over the last decade, GPU compute improved roughly 80x, while memory bandwidth improved only about 17x. The market is noticing. The AI memory market is projected to grow from around $35B in 2025 to about $100B by 2028. Now is a great time to build.

This is the most exciting robotics development I saw today. Manufacturing is moving from fixed automation to flexible, intelligent automation. Traditional factories are rigid. You build a specific line for a specific process with specific tools. This makes them slow to change. The Functgraph reimagines the current system. It encounters a new task, prints the task-specific tools, assembles them, and then uses those tools to complete the job. As modern manufacturing moves toward more customization and faster iteration, companies don't just need cheaper production, but also adaptive production systems. The winner will be companies that can turn physical workflows into software-like modules that are programmable and reusable. This also hints at the future of robotics. Instead of one expensive humanoid trying to do everything with two hands. It may be robots that can manufacture task-specific tools on demand and use the tools to complete new jobs.

Starlink is easy to overlook if you only compare it to urban fiber. The biggest business story is that it extends the operating surface area for AI and robotics. Humanoid robots, autonomous farms, mining sites, oil fields, tankers, and remote factories all need reliable connectivity. However, a lot of physical economy sits in places where traditional infrastructure is slow or expensive to build. Robots and AI not only need better models and chips. They also need a stable deployment environment. Connectivity is a hidden bottleneck in remote settings. Starlink connects these remote assets so machines can be better coordinated, updated, and eventually become autonomous. As AI gives machines intelligence and robotics gives them agency, SpaceX's satellite internet gives them reach and connectivity.

Humanoid robot is entering the "this looks too real" phase. But the real test is not how "human" they look. It's whether they can reliably perform the boring, easy, annoying tasks that humans want to avoid: Folding clothes, handling deliveries, and cleaning rooms. We are entering the Cambrian explosion phase for robotics, when dozens of firms have impressive demos. However, the winner is the one who turns painful labor categories into repeatable robot workflows, collects real-world data, and improves faster than everyone else.

Salesperson that will thrive in the AI age must become a high-trust revenue strategist. AI increasingly handles manual prospecting, CRM, follow-up, email drafting, etc To survive and thrive, you must move toward a sales environment where the hard part is human judgment. Work in industries that value human trust, obtain as much domain expertise as possible, and then combine that with deep business insight into customer pain points. Basic SDR outbound, low ticket SaaS, demo-only sales, and inbound qualification can easily get replaced by AI in the next 5 years.

The AI “jobpocalypse” debate is asking the wrong question. Everyone asks: Can AI do this task? But jobs don’t disappear just because a technology can do a task. The real question is: Does cheaper intelligence expand demand, or destroy the business model? Software didn’t kill web developers. It made software cheaper, which created massive demand for websites, apps, tools, and workflows. AI will probably do the same in industries where the bottleneck is supply: • Healthcare • Legal access • Education • Engineering But AI will destroy businesses where the bottleneck is distribution or information asymmetry: • Traditional SEO • Generic SaaS wrappers • Basic customer support • Low-differentiation agencies My guess: AI makes the top and bottom of the economy stronger. The top layer gets more leverage: Founders, experts, operators, people with distribution. The bottom layer gets cheaper: Research, admin, support, execution. The middle gets crushed: Agencies, junior knowledge workers, and middlemen. AI will reveal which businesses were real bottlenecks, and which were just toll booths.

Cohere’s deal to take over Aleph Alpha is one of the clearest signs that AI is becoming geopolitical infrastructure. On the surface, it is a $20B AI tie-up between a Canadian startup and a German startup. But the real story is bigger: Countries outside the U.S. and China are trying to build an AI bloc of their own. For the last few years, the AI race has been dominated by American giants like OpenAI, Google, Anthropic, and NVIDIA, with China racing to build its own parallel stack. Everyone else was basically a customer. Cohere’s acquisition is a signal for change. They are betting that enterprise and government AI want secure deployment, local infrastructure, and data control. Governments do not want their tax systems, healthcare records, or public services running entirely on foreign-controlled AI infrastructure. Middle powers like Germany and Canada also do not want to choose between Silicon Valley and Beijing. Most countries will talk about “AI safety,” but what they really want is AI sovereignty. The countries that cannot build or access their own AI infrastructure may become digitally dependent on the countries that can. So the real question is: Are we seeing the beginning of a fragmented AI world where every bloc builds its own digital empire?

I believe the biggest edge in the AI age isn’t following the traditional tech path. It’s getting exposed to obscure, expensive workflows that outsiders don’t even know exist.

The Texas World Cup funding fight is a preview of where America is heading. If you are building or investing in AI, this should get your attention. Gov. Greg Abbott reportedly threatened to pull public safety funds from Houston, Dallas, and Austin unless they cooperate with ICE. Underneath the headline, this shows how higher-level governments can use grants, security, and infrastructure funding as political leverage. This matters a lot for the AI industry. The next decade will require massive physical buildout: AI data centers, power plants, transmission lines, chip fabs, and security infrastructure. All of that depends on local governments, state approvals, public utilities, zoning, and community consent. If America enters a period where more major infrastructure decisions become entangled in party loyalty, protests, lawsuits, or federal-state conflict, then business risk is no longer just economic. It becomes political. The companies that win the AI and energy boom may not just be the ones with the best technology. They may be the ones who understand how to navigate around powerholders and public opinion. The big question is: Can the U.S. build the physical infrastructure for the AI age if every city, state, and federal agency becomes a battlefield for political power?

Marco Rubio's absence during the Iran War crisis tells you everything about Trump’s second-term foreign policy. In the past, the Secretary of State built alliances and managed grand strategy. Under Trump, the new model is sending Kushner, Witkoff, Vance, or whoever he trusts personally that week. The real power flows through loyalty, family proximity, and personal relationships. And honestly, this might be perfect for Rubio. If the talks with Iran succeed, Vance gets the win. If they fail, Vance owns the failure. Rubio can stay close to Trump, keep his hands clean, and wait for 2028. So is Rubio being sidelined, or is he just smart enough not to own a crisis he can’t control?

Tesla ramped up its spending on AI to $25 billion. This shows that Musk is betting Tesla's next revenue engine will come from robotaxis, chips, and humanoid robots. Tesla is becoming less of a car company and more of a high-risk AI robotics infrastructure company. But the market will eventually demand proof. Robotaxi experiments and no real Optimus production at scale mean it could increase investor skepticism. This is bold, and if it works, the commercial upside could be tremendous. Yet, the burden of proof is high. Could Musk change industrial history once again?

Sovereign cloud infrastructure is on the rise. There will be mass demand for operators that can deliver compliant local compute at scale.