Trending on X · Archive
Top 6 tweets · The selection preserved as it was published.
Love you Prime but no, they did not. This model is incapable of working in a real code base without hallucinating in egregious ways. It’s not like 5% worse, it’s 10x worse.
Developers are fiercely debating the reliability of a newly released AI coding model that promised to handle large, complex codebases.
In late 2027 Google will have the largest datacenter: Goodnight - Claude, Texas either Google trains AGI there at the end of 2027 or Anthropic already took off and just buys the place
Google's planned multi-gigawatt AI data center in Texas is sparking speculation about the timeline for training next-generation models.
xAI announced today it is adding text watermarking to Grok. This appears to be rolling out globally. xAI was the last holdout, so the only model left that (as far as I know) does not watermark text is ChatGPT, and only ChatGPT in North America.
xAI's decision to watermark Grok outputs leaves OpenAI as the last major developer not universally tagging AI-generated text.
BOOM 🚨 SpaceX is about to add roughly 420,000 of Nvidia's most powerful AI chips in just two months, and Elon Musk just gave us a look at how it's happening. It's all going into Colossus 2, SpaceX's AI data center in Memphis. The site is split into halls with nicknames, and two of them, Macroharder and Minihard, are already loaded with about 220,000 GB300 chips each. Now two more are joining them. Elon says Gun Range is on track to come online by the end of November with the same setup, and Spam should follow in December with around 200,000 chips. That's about $25–35 billion worth of hardware switching on before the year is out. And here's the best part: the two new halls are built for completely different jobs. Gun Range is a training machine, where AI models learn. Training is like a massive group project, with hundreds of thousands of chips trading notes nonstop. To keep up, every chip gets a blazing 800-gigabit network connection, and the building is wired with an enormous amount of fiber. Spam is an inference machine, where AI answers people's questions in real time. Each question can be handled by a small group of chips on their own, so they barely need to talk to the rest of the building. That's why Elon says Spam needs much simpler networking and less than half the fiber. Less wiring means it can be built faster and for less money. Think of it like a classroom versus a help desk. In a classroom, everyone learns together. At a help desk, each clerk takes care of their own customer. And yes, the names are jokes. Spam's building used to store Hormel products (Hormel makes Spam), and Macroharder and Minihard take a shot at Microsoft. We usually follow the AI race through model launches and benchmark scores. This is a reminder that it's also a construction race, and right now SpaceX is moving like the deadline was yesterday.
Elon Musk revealed a massive expansion of his Memphis supercomputer, deploying hundreds of thousands of new Nvidia chips to train AI.
As futuristic as Google's project may seem, it solves one of the biggest problems currently facing the US. Besides the enormous demand for chips, there's an even bigger problem right now: energy. I often bring this up in conversations: China is currently building (or starting to build) around 46 nuclear power plants and will have installed as much solar capacity in 2025 alone as the US has in the last 10-15 years. In the early 2030s, energy and the aging power grid will become the biggest bottleneck for the US. Not chips, but energy. And here, China is currently way ahead in the race. To put it bluntly, they've "solved" the energy problem for themselves, and the US is lagging behind. Moving inference or even training into space and using satellite swarms close to the sun to capture energy (about 8 times more efficiently than on Earth) could solve all the problems at once: energy, scaling, and the grid problem.
Google's latest infrastructure investment highlights the tech industry's pivot toward nuclear power to solve the AI energy bottleneck.
Tesla invited me to be one of the first people in Latvia to try Tesla Assisted Driving 🇱🇻 That’s the European name for Full Self-Driving (Supervised), or FSD. It’s now approved in Latvia, with a software rollout coming soon to eligible cars. The subscription costs €99/month. I spent about an hour behind the wheel, and honestly, I wasn’t expecting to react this strongly. You just choose your destination on the screen, tap to go, and the car pulls out of its parking space and drives all the way to your destination while you supervise. Having your hands off the wheel while the car changes lanes, turns, brakes, accelerates and yields to pedestrians is such a strange, slightly nerve-racking, but incredibly cool feeling. I kept saying, “This is so cool!” Watching a car handle an entire route on familiar roads here in Latvia felt completely insane. We encountered awkwardly parked cars, a cyclist coming toward us and situations where we needed to give another car room to pass. The Tesla handled all of it. That’s just one drive, but I came away seriously impressed. I love driving. But for those repetitive everyday trips? I can absolutely see myself using this. And you really do have to supervise. You’re still responsible for driving. You need to watch the road and be ready to take over immediately. You can’t scroll through your phone or stop paying attention. Even when I looked at the center screen for too long, I got a gentle reminder to watch the road. Public Tesla Assisted Driving test drives have already started at Tesla’s store in Spice Shopping Center in Riga, so you can book one and experience it yourself. I’ve watched so many videos of FSD working elsewhere. Finally trying it here, on roads I actually know, made it feel much more real. I get the excitement now. This is awesome.
Tesla's supervised autonomous driving software is finally clearing regulatory hurdles in Europe, opening a major new market for the tech.
Love you Prime but no, they did not. This model is incapable of working in a real code base without hallucinating in egregious ways. It’s not like 5% worse, it’s 10x worse.
Developers are fiercely debating the reliability of a newly released AI coding model that promised to handle large, complex codebases.
xAI announced today it is adding text watermarking to Grok. This appears to be rolling out globally. xAI was the last holdout, so the only model left that (as far as I know) does not watermark text is ChatGPT, and only ChatGPT in North America.
xAI's decision to watermark Grok outputs leaves OpenAI as the last major developer not universally tagging AI-generated text.
As futuristic as Google's project may seem, it solves one of the biggest problems currently facing the US. Besides the enormous demand for chips, there's an even bigger problem right now: energy. I often bring this up in conversations: China is currently building (or starting to build) around 46 nuclear power plants and will have installed as much solar capacity in 2025 alone as the US has in the last 10-15 years. In the early 2030s, energy and the aging power grid will become the biggest bottleneck for the US. Not chips, but energy. And here, China is currently way ahead in the race. To put it bluntly, they've "solved" the energy problem for themselves, and the US is lagging behind. Moving inference or even training into space and using satellite swarms close to the sun to capture energy (about 8 times more efficiently than on Earth) could solve all the problems at once: energy, scaling, and the grid problem.
Google's latest infrastructure investment highlights the tech industry's pivot toward nuclear power to solve the AI energy bottleneck.
In late 2027 Google will have the largest datacenter: Goodnight - Claude, Texas either Google trains AGI there at the end of 2027 or Anthropic already took off and just buys the place
Google's planned multi-gigawatt AI data center in Texas is sparking speculation about the timeline for training next-generation models.
BOOM 🚨 SpaceX is about to add roughly 420,000 of Nvidia's most powerful AI chips in just two months, and Elon Musk just gave us a look at how it's happening. It's all going into Colossus 2, SpaceX's AI data center in Memphis. The site is split into halls with nicknames, and two of them, Macroharder and Minihard, are already loaded with about 220,000 GB300 chips each. Now two more are joining them. Elon says Gun Range is on track to come online by the end of November with the same setup, and Spam should follow in December with around 200,000 chips. That's about $25–35 billion worth of hardware switching on before the year is out. And here's the best part: the two new halls are built for completely different jobs. Gun Range is a training machine, where AI models learn. Training is like a massive group project, with hundreds of thousands of chips trading notes nonstop. To keep up, every chip gets a blazing 800-gigabit network connection, and the building is wired with an enormous amount of fiber. Spam is an inference machine, where AI answers people's questions in real time. Each question can be handled by a small group of chips on their own, so they barely need to talk to the rest of the building. That's why Elon says Spam needs much simpler networking and less than half the fiber. Less wiring means it can be built faster and for less money. Think of it like a classroom versus a help desk. In a classroom, everyone learns together. At a help desk, each clerk takes care of their own customer. And yes, the names are jokes. Spam's building used to store Hormel products (Hormel makes Spam), and Macroharder and Minihard take a shot at Microsoft. We usually follow the AI race through model launches and benchmark scores. This is a reminder that it's also a construction race, and right now SpaceX is moving like the deadline was yesterday.
Elon Musk revealed a massive expansion of his Memphis supercomputer, deploying hundreds of thousands of new Nvidia chips to train AI.
Tesla invited me to be one of the first people in Latvia to try Tesla Assisted Driving 🇱🇻 That’s the European name for Full Self-Driving (Supervised), or FSD. It’s now approved in Latvia, with a software rollout coming soon to eligible cars. The subscription costs €99/month. I spent about an hour behind the wheel, and honestly, I wasn’t expecting to react this strongly. You just choose your destination on the screen, tap to go, and the car pulls out of its parking space and drives all the way to your destination while you supervise. Having your hands off the wheel while the car changes lanes, turns, brakes, accelerates and yields to pedestrians is such a strange, slightly nerve-racking, but incredibly cool feeling. I kept saying, “This is so cool!” Watching a car handle an entire route on familiar roads here in Latvia felt completely insane. We encountered awkwardly parked cars, a cyclist coming toward us and situations where we needed to give another car room to pass. The Tesla handled all of it. That’s just one drive, but I came away seriously impressed. I love driving. But for those repetitive everyday trips? I can absolutely see myself using this. And you really do have to supervise. You’re still responsible for driving. You need to watch the road and be ready to take over immediately. You can’t scroll through your phone or stop paying attention. Even when I looked at the center screen for too long, I got a gentle reminder to watch the road. Public Tesla Assisted Driving test drives have already started at Tesla’s store in Spice Shopping Center in Riga, so you can book one and experience it yourself. I’ve watched so many videos of FSD working elsewhere. Finally trying it here, on roads I actually know, made it feel much more real. I get the excitement now. This is awesome.
Tesla's supervised autonomous driving software is finally clearing regulatory hurdles in Europe, opening a major new market for the tech.
Love you Prime but no, they did not. This model is incapable of working in a real code base without hallucinating in egregious ways. It’s not like 5% worse, it’s 10x worse.
Developers are fiercely debating the reliability of a newly released AI coding model that promised to handle large, complex codebases.
BOOM 🚨 SpaceX is about to add roughly 420,000 of Nvidia's most powerful AI chips in just two months, and Elon Musk just gave us a look at how it's happening. It's all going into Colossus 2, SpaceX's AI data center in Memphis. The site is split into halls with nicknames, and two of them, Macroharder and Minihard, are already loaded with about 220,000 GB300 chips each. Now two more are joining them. Elon says Gun Range is on track to come online by the end of November with the same setup, and Spam should follow in December with around 200,000 chips. That's about $25–35 billion worth of hardware switching on before the year is out. And here's the best part: the two new halls are built for completely different jobs. Gun Range is a training machine, where AI models learn. Training is like a massive group project, with hundreds of thousands of chips trading notes nonstop. To keep up, every chip gets a blazing 800-gigabit network connection, and the building is wired with an enormous amount of fiber. Spam is an inference machine, where AI answers people's questions in real time. Each question can be handled by a small group of chips on their own, so they barely need to talk to the rest of the building. That's why Elon says Spam needs much simpler networking and less than half the fiber. Less wiring means it can be built faster and for less money. Think of it like a classroom versus a help desk. In a classroom, everyone learns together. At a help desk, each clerk takes care of their own customer. And yes, the names are jokes. Spam's building used to store Hormel products (Hormel makes Spam), and Macroharder and Minihard take a shot at Microsoft. We usually follow the AI race through model launches and benchmark scores. This is a reminder that it's also a construction race, and right now SpaceX is moving like the deadline was yesterday.
Elon Musk revealed a massive expansion of his Memphis supercomputer, deploying hundreds of thousands of new Nvidia chips to train AI.
In late 2027 Google will have the largest datacenter: Goodnight - Claude, Texas either Google trains AGI there at the end of 2027 or Anthropic already took off and just buys the place
Google's planned multi-gigawatt AI data center in Texas is sparking speculation about the timeline for training next-generation models.
As futuristic as Google's project may seem, it solves one of the biggest problems currently facing the US. Besides the enormous demand for chips, there's an even bigger problem right now: energy. I often bring this up in conversations: China is currently building (or starting to build) around 46 nuclear power plants and will have installed as much solar capacity in 2025 alone as the US has in the last 10-15 years. In the early 2030s, energy and the aging power grid will become the biggest bottleneck for the US. Not chips, but energy. And here, China is currently way ahead in the race. To put it bluntly, they've "solved" the energy problem for themselves, and the US is lagging behind. Moving inference or even training into space and using satellite swarms close to the sun to capture energy (about 8 times more efficiently than on Earth) could solve all the problems at once: energy, scaling, and the grid problem.
Google's latest infrastructure investment highlights the tech industry's pivot toward nuclear power to solve the AI energy bottleneck.
xAI announced today it is adding text watermarking to Grok. This appears to be rolling out globally. xAI was the last holdout, so the only model left that (as far as I know) does not watermark text is ChatGPT, and only ChatGPT in North America.
xAI's decision to watermark Grok outputs leaves OpenAI as the last major developer not universally tagging AI-generated text.
Tesla invited me to be one of the first people in Latvia to try Tesla Assisted Driving 🇱🇻 That’s the European name for Full Self-Driving (Supervised), or FSD. It’s now approved in Latvia, with a software rollout coming soon to eligible cars. The subscription costs €99/month. I spent about an hour behind the wheel, and honestly, I wasn’t expecting to react this strongly. You just choose your destination on the screen, tap to go, and the car pulls out of its parking space and drives all the way to your destination while you supervise. Having your hands off the wheel while the car changes lanes, turns, brakes, accelerates and yields to pedestrians is such a strange, slightly nerve-racking, but incredibly cool feeling. I kept saying, “This is so cool!” Watching a car handle an entire route on familiar roads here in Latvia felt completely insane. We encountered awkwardly parked cars, a cyclist coming toward us and situations where we needed to give another car room to pass. The Tesla handled all of it. That’s just one drive, but I came away seriously impressed. I love driving. But for those repetitive everyday trips? I can absolutely see myself using this. And you really do have to supervise. You’re still responsible for driving. You need to watch the road and be ready to take over immediately. You can’t scroll through your phone or stop paying attention. Even when I looked at the center screen for too long, I got a gentle reminder to watch the road. Public Tesla Assisted Driving test drives have already started at Tesla’s store in Spice Shopping Center in Riga, so you can book one and experience it yourself. I’ve watched so many videos of FSD working elsewhere. Finally trying it here, on roads I actually know, made it feel much more real. I get the excitement now. This is awesome.
Tesla's supervised autonomous driving software is finally clearing regulatory hurdles in Europe, opening a major new market for the tech.
Ranked by engagement. Posts belong to their authors on X. Figures recorded when this selection was prepared.
The Brief
AI and technology: what changes and why it matters.
Your daily selection, in English or Spanish.
Free forever. Unsubscribe anytime.
