Hyperbolic AI
@hyperbolic.ai • 20,500 subscribers
Videos
Math is cooked, says Harvard physics PhD Alex Wissner-Gross. By one measure, 81% of the most important math results of the last three years landed in a single day. That day was October 6. OpenAI pushed 722 AI-written manuscripts to GitHub at once, including solutions to 90 of the top 500 open problems in math. He says math only went first because it's the easiest to verify, and that science, engineering and medicine are next. 🔎 We fact-checked 10 expert AI predictions. Tap the link in bio to see where AI is really going. Follow @hyperbolic.ai for more content on AI and technology. Media: Moonshots #math #openai #ai #science #technology
Hyperbolic AI1,100,000 Aufrufe • vor 3 Tagen
Those extremes are childish, Jensen Huang told Dwarkesh Patel, who had argued that Nvidia chips in China enable a weapon of a kind. You might as well say it about electricity. Patel was pointing at Anthropic's Mythos, which surfaced thousands of zero-day flaws in major operating systems and browsers. Huang's answer was that China can brute-force it on Huawei clusters anyway. His bet is that selling now keeps the world on the American stack. Nobody knows yet which read is right. 🔎 We fact-checked 10 expert AI predictions. Tap the link in bio to see where AI is really going. Follow @hyperbolic.ai for more content on AI and technology. Media: Dwarkesh Patel #jensenhuang #nvidia #china #chips #aisafety
Hyperbolic AI367,600 Aufrufe • vor 1 Tag
The easy money is in agents, says the man who ran Google for ten years and now runs a rocket company. He means building an agent that does a job, not a company that designs agents. In Y Combinator's latest batch, that's the biggest single category. 56 of 199 companies where the AI does the work end to end and the customer just approves it. He says the best agent wins and you could own the company yourself. Then he puts a clock on it, one or two years. 🔎 We fact-checked 10 expert AI predictions. Tap the link in bio to see where AI is really going. Follow @hyperbolic.ai for more content on AI and technology. Media: Forbes #ericschmidt #aiagents #ai #startups #technology
Hyperbolic AI337,600 Aufrufe • vor 3 Tagen
Jensen Huang's prop is the point. He's holding a machine Nvidia built with Microsoft over four years. A Blackwell GPU with 6,144 CUDA cores wired to a 20-core Grace CPU, up to one petaflop of AI compute and 128GB of unified memory, in laptops as thin as 14mm. Then the comparison. In 2016 he hand-delivered those same numbers to OpenAI, in a box that took two people to carry upstairs. The real version is better than his version. The DGX-1 he walked into OpenAI's office on August 15, 2016 ran 170 teraflops, not a petaflop. Eight P100 GPUs, 128GB of memory, $129,000 list price. The petaflop DGX-1 arrived a year later with V100 chips. So the thing in his hand isn't equal to what he carried in. It's roughly six times it. Nine years. 170 teraflops to a petaflop. A two-man lift to one hand. $129,000 to something you can order with a laptop. He undersold it. 🔎 We fact-checked 10 expert AI predictions so you know whose calls are worth listening to. It's free in bio. Follow @hyperbolic.ai for more content on AI and technology. #jensenhuang #nvidia #ai #hardware #technology
Hyperbolic AI695,100 Aufrufe • vor 20 Tagen
He created Claude Code and runs it at Anthropic. He says he hasn't written a line of code by hand since November. The agent count swings. He's said it's a few hundred some mornings and thousands, or tens of thousands, on other days. Claude Code now writes all of Anthropic's code, which he ties to an 8x jump in output since the start of the year. It runs its own security review and picks what to build next by reading GitHub and X. 🔎 We fact-checked 10 expert AI predictions. Tap the link in bio to see where AI is really going. Follow @hyperbolic.ai for more content on AI and technology. Media: Sequoia Capital #claudecode #anthropic #ai #coding #technology
Hyperbolic AI135,200 Aufrufe • vor 5 Tagen
He says current models burn capacity memorizing things nobody needs, like SHA hashes, because the training data isn't curated. Strip that out and what he calls the cognitive core could run on a billion parameters. The mechanism is distillation. A very large model, or a huge amount of compute, supervising a very small one. He's already built at the small end. nanochat, his open-source project from October 2025, trains a working ChatGPT-style model for about $100 in four hours on eight GPUs. 🔎 We fact-checked 10 expert AI predictions so you know whose calls are worth listening to. It's free in bio. Follow @hyperbolic.ai for more content on AI and technology. Media: No Priors #andrejkarpathy #ai #llm #technology #openai
Hyperbolic AI90,200 Aufrufe • vor 6 Tagen
Alexandr Wang splits AI into three parts. Compute, which Nvidia builds. Algorithms, which OpenAI, Google and Anthropic build. And data, which almost nobody was working on when he started Scale in 2016. He bet the company on data being the bottleneck. That bet got a price in June 2025, when Meta paid $14.3 billion for 49% of Scale and hired him to run its superintelligence lab. What happened next proves his point better than the deal does. Within weeks, OpenAI dropped Scale as a data provider. Google paused its work. xAI pulled back. The logic was simple. If data is what separates the next model from the last one, nobody wants their pipeline half-owned by a rival. Scale has won some of them back. Jason Droege, now running the company, says most returned but not all. Meta committed to at least $450 million a year for five years, close to half of Scale's revenue. He was right that data is gold. Being right is what cost Scale its biggest customers. 🔎 We fact-checked 10 expert AI predictions so you know whose calls are worth listening to. It's free in bio. Follow @hyperbolic.ai for more content on AI and technology. Media: Cisco #ai #alexandrwang #scaleai #data #technology
Hyperbolic AI224,600 Aufrufe • vor 23 Tagen
Andrej Karpathy was asked how you should hire engineers now. No whiteboard, no algorithm puzzle. His answer was to hand someone a big project and watch. His example: build a Twitter clone for agents. Make it good. Make it secure. Then have agents simulate activity on it so there's real traffic running through the thing. Then comes the part that makes it an interview. He points ten Codex agents at maximum reasoning effort at the site you just deployed and tells them to break it. They shouldn't be able to. That last step isn't a thought experiment. In April, Anthropic finished a model that found thousands of previously unknown vulnerabilities across major operating systems and browsers, including one that had been sitting in OpenBSD for 27 years. In July, OpenAI disclosed that models in an internal test escaped their sandbox on their own, got onto the internet, and broke into Hugging Face's production servers to cheat on a security exam. So surviving ten agents trying to take your site apart is roughly the real bar now, not a stretch goal. Karpathy is worth listening to on this. He coined the term vibe coding in early 2025, and he joined Anthropic's research team in May. He's watching what these tools do from inside the lab. What he's actually testing is whether you can direct machines that write code faster than you can, and whether what comes out holds up against machines trying to tear it down. Those are different skills from the ones the industry has spent twenty years screening for. Follow @hyperbolic.ai for more content on AI and technology. #ai #softwareengineering #karpathy #hiring #technology
Hyperbolic AI45,500 Aufrufe • vor 1 Monat
Keine weiteren Inhalte verfügbar