10,000 AI AGENTS SOLVED A PROBLEM WITHOUT A SINGLE ARGUMENT. THE HUMANS STARTED FIGHTING OVER CREDIT RIGHT AFTER THE ANNOUNCEMENT. First, what happened. OpenAI announced that its agents solved one of the most famous problems in modern mathematics. By its account, it took 88 hours. Around 10,000 agents worked on the problem at the same time. Now, what struck me the most. Sam Altman described at Stanford how they interacted. There was no leader. There were no arguments about who did what. No conflicts and no fight for recognition. Everyone simply moved toward one goal. He himself said it's unlike any organization he can imagine. And here's what came next. As soon as the result was announced, a dispute broke out between people over who deserves the credit. The agents split the problem. The humans split the glory. Meanwhile, mathematicians went through their own journey in five days. First came the reaction that it was fake. Then that their field was over. Then that it shouldn't be done this way. By day five, many said something different. Yes, one kind of their work is gone. But most of them never solved problems at that level. They made sense of knowledge, explained it to people and decided what to work on. There is one more idea from Altman that's easy to miss. He suggests, and says outright that it's pure speculation, that almost every organization in the world is modeled on the military. A pyramid, command, control. Because that was once the only known way to manage a large number of people. And here 10,000 agents did it without any pyramid at all. Now imagine a company where only agents work. Nobody takes offense at feedback. Nobody steals an idea. Nobody fights for the corner office. It sounds perfect. And that's exactly why it's interesting that Altman immediately adds: people will never work like this. Because there is beauty in our imperfection. The arguments, the ambition, the emotion. He wouldn't want a world without them. And I agree with him. Because friction between people isn't only a cost. It's also how we decide what matters to us. When two people argue over credit, they're really arguing about what contribution means to them. A machine doesn't ask that question. It just executes. So it seems to me the main question of the coming years isn't what AI can replace. It's which friction we're willing to keep. Altman says people will always care most about other people. That we don't much care what a machine does for its own purposes. That work has value because a person is behind it. It seems this holds true even when we argue. And maybe it's in those arguments that we remain most ourselves.

NINE YEARS AGO, JENSEN HUANG DELIVERED A SUPERCOMPUTER TO ELON MUSK. TODAY, HE BROUGHT HIM ITS SUCCESSOR. And this moment connects two very different eras of computing. Back in 2016, Jensen personally delivered NVIDIA's first DGX-1 to Elon Musk. At the time, Musk was part of OpenAI. That machine helped researchers push the boundaries of artificial intelligence and became part of the early infrastructure behind the AI revolution. Nine years later, Jensen is back. But this time, he is carrying something completely different. Meet NVIDIA DGX Spark. A compact desktop AI supercomputer powered by the Grace Blackwell architecture. And the specifications are impressive. Up to 1 petaflop of AI performance at FP4 precision. 128GB of unified memory. A powerful NVIDIA GB10 superchip. And a form factor small enough to sit on a desk. For context, the original DGX-1 was a large enterprise system. DGX Spark brings a substantial amount of AI computing capability into a device weighing just 1.2 kilograms. But what really matters is what this makes possible. Developers and researchers can run AI models locally, experiment with large language models, build AI agents and fine tune models without relying entirely on remote data centers. NVIDIA says a single DGX Spark can handle inference with models up to 200 billion parameters. And multiple systems can be connected together to handle even larger workloads. During the conversation, Jensen also highlighted the enormous progress in computing efficiency, comparing the original DGX-1 with the new generation. Then came one of the best moments. Jensen joked that Elon was building things bigger than him. Elon simply agreed. After all, SpaceX is building rockets and massive infrastructure, while NVIDIA is building the computing systems that power the AI revolution. And there is something fascinating about seeing these two worlds meet at Starbase in Texas. One company is pushing the boundaries of space exploration. The other is pushing the boundaries of artificial intelligence. Nine years ago, Jensen delivered a supercomputer that helped launch a new chapter in AI. Now he is bringing that computing power closer to individual developers. The next AI breakthrough might not come from a massive data center. It might start on someone's desk.

HOLY SH*T. OPENAI JUST MADE AI AGENTS CHEAPER AND BETTER AT USING YOUR COMPUTER. Meet GPT 6.1 Terra. And this update is about much more than generating better answers. OpenAI is pushing AI toward something far more useful. Actually getting work done. Terra is designed to understand what is happening on your screen, interact with computer applications and complete complex sequences of tasks. Think about what that means. An AI can help prepare proposals, interpret technical drawings, fill out forms, edit videos and handle repetitive office work. Instead of explaining how to do something, it can work through the steps required to get it done. And OpenAI says the new model is significantly better than GPT 5.6 Terra at these kinds of tasks. But here is the part that really caught my attention. The price. GPT 6.1 Terra's standard API pricing is 50% lower than GPT 5.6 Terra. That matters because AI agents can make many calls to a model while completing a single workflow. Every step costs money. Reading information, analyzing a screen, deciding what to do next and checking the result can all add up. Lower costs make it more practical to run these systems at scale. OpenAI also reports major improvements in coding performance and visual reasoning, alongside fewer factual errors in its internal evaluations. This combination is important. A model that can understand a task, operate software and write code becomes much more useful when it can do all of that faster and at a lower cost. Imagine AI agents processing invoices, updating spreadsheets, organizing customer information or handling repetitive tasks across multiple applications. Not just telling you what to do. Actually helping you get it done. Of course, computer use is not the same as perfect autonomy. Complex workflows still need testing, permissions and human oversight. But the direction is becoming clear. The next big AI race is not only about which model can think better. It is about which model can reliably turn that intelligence into action. And now, OpenAI is making that capability cheaper to use.

🚨HOLY SH*T. JENSEN HUANG JUST HELD A PETAFLOP OF AI POWER IN ONE HAND. And the comparison he made is absolutely insane. In 2016, Jensen Huang delivered an NVIDIA DGX 1 system to a small company called OpenAI. It was a massive machine built specifically for AI computing. It delivered one petaflop of computing performance and came with 128GB of memory. And Jensen reminded everyone of one detail. It took two people to carry that machine upstairs. Now, ten years later, he is holding a compact AI computer in one hand and showing the same headline numbers. One petaflop. 128GB of memory. Let that sink in for a second. The new RTX Spark lineup was developed through four years of work with Microsoft and the wider PC industry. According to Jensen, the project involved a new chip called N1X and an entirely new family of computers, including desktops, laptops and workstations. The goal is to bring serious AI computing power into a much smaller form factor. And this is where things get interesting. For years, access to powerful AI hardware meant relying on expensive cloud infrastructure or specialized machines that were far beyond the reach of most individual developers. Compact systems like this point toward a different approach. Imagine developing AI agents, experimenting with models and running demanding AI workloads on your own desk. Less dependence on remote infrastructure. More control over your development environment. And potentially fewer recurring cloud costs for workloads that can run locally. Of course, there is an important distinction. Matching the same peak petaflop figure does not automatically mean two machines deliver identical real world performance. The precision used, memory architecture and workload all matter. But the physical transformation is still remarkable. In 2016, this class of AI computing came in a machine that required two people to carry. Now, Jensen Huang can hold a system with comparable headline specifications in one hand. That is what makes this more than just another computer launch. AI hardware is not only becoming more powerful. It is becoming dramatically more accessible. And the next generation of AI developers may not need a server room to get started.

WHAT IF WE COULD TAKE QUANTUM PHOTON TRANSFER FROM 54% TO 100%? Scientists at TU Wien may have found a way. And the solution is surprisingly clever. Quantum computers need to transfer information between qubits. One qubit emits a photon. Another must absorb it. But with conventional approaches, that absorption succeeds only around 54% of the time. Imagine losing almost half of the information transfer attempts before the process even finishes. Now researchers have proposed a way to change that. The key is the shape of the photon’s wave. When a qubit emits a photon, its wave typically starts strong and gradually fades away, much like the sound of a bell. But the receiving qubit needs a very different waveform for perfect absorption. So the researchers propose using a specially engineered optical waveguide to reshape the photon as it travels. By controlling how different frequencies move through the waveguide, they calculated that the photon could arrive with precisely the waveform needed for optimal absorption. The result? Their simulations suggest that absorption could theoretically reach 100% under ideal conditions. And the most interesting part is that this is a passive approach. Instead of relying on additional active control at the receiving end, the waveguide itself does the work of shaping the photon. The method has not yet been demonstrated experimentally, but the researchers believe the required setup should be technically feasible. If it works in practice, it could help make quantum computers and communication networks far more reliable. We often think quantum breakthroughs require entirely new hardware. Sometimes, the breakthrough may come from changing the way a single photon travels. And in quantum computing, that tiny detail could make a massive difference.

WHAT IS THIS THING? THIS BOT LOSES OVER HALF OF THE TRADS AND STILL GROWS FROM $1,400 TO $22,646. Look at that bullet. These are 20,480 possible futures that the bot keeps in its head at the same time. The blue lines are the worlds where bitcoin will go up. Pink and orange are the worlds where he will go down. And now about how it works. The bot trades five-minute markets on Polymarket, up or down for Bitcoin. Live data from seven streams goes inside. These are converted into probabilities, such as 45 percent for a rise versus 55 for a fall. This probability is compared to the price at which the market is selling the same outcome. The difference between them is called the advantage. And only if there is an advantage, the bot opens a deal. If it is less than three cents, the word PASS appears on the screen and there is no transaction. This is where the main thing is hidden. The win rate on the panel is 48 percent. That is, more than half of the deals are closed in the red. But this is not a problem. Imagine you buy a stock at 32 cents that will pay a dollar if you're right. You don't have to guess more often than it's worth. It is enough that your estimate is more accurate than the market price. Then even with losses more than half the time the math works for you. This is visible on the screen for each transaction. Purchase for 32 cents with advantage plus 5.3 cents. Purchase for 53 cents with advantage plus 6.7 cents. Different prices, different sizes, but one rule. And at the bottom is the Landing Curve. This is a breakdown of where Bitcoin could end up at the close of the window. The bot misses a single point. He sees a whole range of opportunities and compares them to the market price. The window per day is 288, and the rule is the same every time. 14,222 tickets were executed in the transactions, the average advantage was plus 4.6 cents, and the maximum drawdown was 4.2 percent. Now about what interests me in this. Most people think of a good trader as someone who guesses. And here the bot doesn't even try. He thinks Every five minutes, millions of small decisions gather into one curve, which grows not because the bot is right, but because it is more expensive to be right than to be wrong. And that changes the question. Not as often as you would guess. And how much you win when you are right, and how much you lose when you are not. That's what Crawler Arena from CRAWLHIVE is all about. You set the verdict, entry, take profit, stop loss and size yourself, without a code. And your crawler enters the arena and trades according to these rules against everyone else, with real commissions and real slippage. That is, the same difference between winning and losing is now in your hands. This is not financial advice. Past results do not guarantee future results.

JENSEN HUANG JUST SHOWED A 25 CM ROBOT DUCK LEARNING TO MOVE WITH AI AGENTS. And this little robot is part of a much bigger story. Meet Microduck. It is an open source robotic project from Hugging Face, and its behavior was trained in simulation before being transferred to the physical robot. The duck can learn to run, get back up and grab objects. But what makes this demonstration interesting is how AI agents helped make it happen. Instead of treating robotics as one enormous engineering task, the workflow breaks it into stages. Build the model. Set up the simulation. Train the behavior. Then bring those skills into the real world. That last step is crucial. A robot that works in a virtual environment still has to deal with gravity, balance, friction and the unpredictability of real objects. And this is where NVIDIA's vision for physical AI becomes important. AI is moving beyond generating text, images and code. It is increasingly being used alongside simulation and robotics tools to help machines learn how to interact with the physical world. The duck may be tiny. But the idea behind it is anything but small. Because the next generation of AI may not just live inside our computers. It may learn to move outside them.

TESLA ISN’T DEFYING GRAVITY. IT’S USING SUCTION TO STAY ON THE TRACK. Look closely at this Roadster. The car climbs a giant loop, turns upside down and keeps moving even when its speed doesn’t look high enough to explain what we’re seeing. So what is actually holding it there? Not magnets. The secret is underneath the car. Tesla has built a fan powered suction system into the Roadster’s underbody. The fans pull air out from beneath the car, creating lower pressure underneath it. The pressure difference helps push the car against the track. And here is the part I find fascinating. Traditional aerodynamic downforce depends heavily on airflow generated by the car’s movement. But a fan based suction system can generate substantial downforce even at low speeds. That changes what a car can do. Instead of relying on speed alone to stay planted, the vehicle can actively create the force that holds it against the surface. It is the difference between a car that needs the road to keep it in place and a car engineered to pull itself toward it. This is not gravity being defeated. It is engineering finding another way to work with it. And that is what makes this demonstration so wild.

8 AI AGENTS RAN A TRADING DESK ALL NIGHT. THE BANK WENT FROM $12,155 TO $14,853. But the number isn't the most interesting part. The most interesting part is that the boss of this desk never opens a trade. This is the new version of Grok Bot together with Jev. And it's no longer a single bot. It's a whole trading desk where every agent has its own job. CHIEF decides which token to pick, but never trades itself. SCAN looks for what is moving. VET checks whether the token is real and what its risk structure looks like. SOCIAL looks at the project's X account. SIZE decides how much to put in. FILLS executes the trade. RISK closes the position. CRAWLER reads the pages and posts first, and only then does the desk trade. Before any of them act, the hard rules run, and they cost nothing. Minimum token age of 15 minutes. Liquidity from $12,000. 24 hour volume from $48,000. Market cap from $60,000. At least 150 trades. The largest wallet holds no more than 5 percent. The top 10 wallets hold no more than 60 percent. Plain arithmetic, with no AI call at all. Then come the soft filters from Jev. Exit risk from concentration. Whether the momentum is already spent. Whether the liquidity fits the trade size. Whether the project account is a recycled one. Whether the dev still holds too much. And the order of these filters matters. Free checks first. Then trade checks. Then chain checks. And only at the end, the expensive ones. Killing a token for free is far cheaper than asking AI about every single one. The bot also never enters on a guess. It waits for real buying flow first, and only then goes in. No flow, no entry. Most people think a strong trading bot is just a smart model. But a smart model alone makes mistakes. A system where every step is separated and every step is checked makes far fewer. It has order. And it has a boss who only decides and never gets greedy. That's the real edge. This is not financial advice. Past results do not guarantee future results. Who plays the role of CHIEF in your trading right now, you or your emotions?

THE MOST DANGEROUS AI TRADING BOT MIGHT BE THE ONE THAT LOOKS PERFECT. A rising profit curve. Clean analytics. A system that appears to know exactly when to enter and exit the market. But here's what I think people often miss. A trading system doesn't have to be good to look profitable. It only has to look good under the conditions it was tested on. Change the market. Add trading fees. Increase volatility. Let a losing position run longer than expected. Suddenly, the strategy that looked brilliant can start falling apart. That's why I care less about a screenshot showing profit and more about what happens when the strategy stops working. How much can it lose? How does it behave under pressure? Can it survive a market it wasn't designed for? Because making money in favorable conditions is one thing. Surviving unfavorable conditions is something else entirely. And I think that's the real challenge for AI trading. Not teaching a machine to find winning trades. Teaching it to recognize when its own strategy is no longer working.