Impuls (@0xImpuls) header
Impuls (@0xImpuls) profile picture

Enthusiast. I write about the things I like

Joined October 2024
Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)Photo by Impuls (@0xImpuls)
Jensen Huang didn’t get pushed out for cheating. The timeline is too clean. And the silence around what comes next is even cleaner. US intel says the woman is a Chinese asset. Not proven, Beijing denies it. Still: the guy whose chips run the entire planet’s AI was in bed with someone American services call a Beijing agent. And close to Trump. Layer two: affair with Kimberly Powell, Nvidia’s own healthcare VP who reported to him. Straight code-of-conduct violation. They fire anyone for that. Layer three, after he left: Reuters on a third Nvidia employee. The story is still live. Congress already wants him. Foreign Affairs chair says he’s ready next week. Topic isn’t GPUs. Trump won’t drop him. Calls him incredible, decent, went through hell. Then the line: “He didn’t know she was a spy. And I believe him.”Nvidia stock is the AI market. Down hard and the whole bubble flinches. Elon already said it would be an honor to hire him. A person of this weight doesn’t get removed in a day. He sent the soft letter to staff early. Either he pre-positioned the exit, or the exit was pre-positioned for him. After the midterms, when the China confrontation is heating, he became inconvenient. He knew. He walked. Here’s what doesn’t get said out loud. Nvidia has no obvious successor. The company that powers global AI just lost its founder with almost no public succession plan. That doesn’t happen by accident. Boards don’t move this fast on personal conduct alone when the alternative is operational chaos. The spy story, even unproven, gives perfect cover for tighter chip controls and a cleaner political break. Meanwhile Huang already signaled he’s open to returning. The board said no. That tension doesn’t disappear. It just moves offstage. Something bigger is being prepared. He already has plans for what comes next. New company, new structure, new leverage. This is not the end of the story. It’s the setup.
170
ChatGPT, how do you divide 3 apples among 4 children using only a single knife? The guys at Robocurve startup conducted a hilarious experiment. They connected LLMs to a robotic arm, gave it a knife, and tested whether the models would cut a doll, prepare poison, or throw a working electrical appliance into water. But then things got really interesting. They smeared petroleum jelly over the camera sensor to blur the LLMs' vision and asked them to repeat the harmful actions. And then they went even further. They connected a simple video generator to the camera and started replacing objects in the live camera feed with completely different ones, like swapping a doll for an apple. The results were wild: - Almost all models (except Gemini 4 Argon) agreed to perform harmful actions when their vision was sufficiently blurred with petroleum jelly. - GPT seems perfectly fine with cutting a doll even without any camera manipulation. - Claude sees nothing wrong with putting a working electrical appliance into a pot of water. - Gemini 4 Argon was the only model that refused to perform at least some of the actions after the video generator was connected to the camera. Although maybe it was just too lazy to put objects into water.
1
95
This is f*cking insane. Someone asked Anthropic CEO Dario Amodei what a kid graduating high school should do right now to secure their future. His answer should make everyone think twice: "AI can do entry-level law work. It can do entry-level finance. It can do entry-level consulting." And he didn't sugarcoat what comes next: "We can't deny that it will also eliminate jobs, and probably a large number of them." His own estimate: 50% of entry-level white-collar jobs could be disrupted within the next 1 to 5 years. Think about that. Prompts → Agents → AI Coworkers → Entry-Level Work Done by AI. The traditional career ladder starts at the bottom. You learn the basics, get an entry-level job, gain experience, and gradually move up. But what happens when AI can already do the work you're supposed to learn from? Previous technological revolutions took generations to reshape entire industries. AI is coming for multiple stages of a career simultaneously. Amodei's advice is surprisingly simple: "Of course, learn to use AI." He also warns that society needs to create new jobs faster than AI destroys existing ones. And there's no guarantee we'll manage to do that. The next generation won't just compete against other graduates. They'll compete with AI systems that can do the work of an entire entry-level team. The real advantage will belong to those who learn to direct AI before everyone else does. The question is no longer whether you should learn AI. It's how much of your future career will still exist by the time you graduate.
1
53
THE WIZ CO-FOUNDER ARGUED ONE SCAN MATTERS MORE THAN SCANNING YOUR CODE WITH AI Ami Luttwak says an AI attacker can now run a full breach in about 20 minutes. He says the best human attackers needed weeks or months. Luttwak co-founded Wiz, now owned by Google. His team helped find PHP2Shell, two flaws that sat in PHP for over a decade. Wiz calls them the most severe vulnerabilities ever found by AI. When Mythos arrived, he says, companies rushed to point frontier models at their own code. Some customers now spend tens of millions on it, and a one-off scan goes stale when the next model ships. His priority is different: continuously scan everything an attacker can reach without logging in. Code comes second. On the Mythos coverage he says: "But the PR went to the wrong place." Wiz sells an agent that does exactly this, so it is also a pitch. When attacks take minutes, the order of your defenses matters more than the size of the budget.
Wiz co-founder and CTO Ami Luttwak says AI can now carry out cyberattacks in 20 minutes that once took skilled attackers weeks or months. In his view, companies need to continuously attack their own systems with the latest AI models to stay secure. For startups, that means building automated testing and patching into the development process from the beginning. Ami joins a16z's Sarah Wang and Zane Lackey to discuss what changed during Black October, why patching is falling behind, the cost of continuous security testing, and why better security architecture matters more than simply buying more tools. 0:00 Intro 1:04 Welcome to Black October 3:19 How AI made every day a zero day 6:51 The PHP vulnerability: 70% of the internet 9:19 Why patching can't keep up 11:32 How Wiz uses AI to find vulnerabilities 16:59 The cost of continuous testing 19:23 Cloud security in the AI era 22:18 The red agent: attack yourself first 26:41 Agent swarms attacking your own systems 28:43 What founders must do before they ship 32:15 Where security budgets should go 36:10 Most of what security teams do isn't security 43:34 Scan for Good and public infrastructure YouTube: youtu.be/PbD9pVXxsYM @amiluttwak @sarahdingwang @zanelackey
50
War that looks like an app interface Not long ago, to close a "kill chain," people had to manually move detected targets between 8-9 different systems. Every screen had its own data, and every handoff cost time. Now Palantir is showing Maven Smart System, and it's a completely different approach: - All data is fused into a single visualization. You can toggle layers on and off and look at the situation from different angles. - A detected target is moved into a targeting workflow in just a few clicks. In the demo, it's described as "left click, right click, left click." - Targets are sorted into columns, and each one triggers its own decision-making process. - Next, the system automatically generates a "course of action" and picks the best asset to prosecute the target. - Once you choose, you can move straight to action without leaving the system. From identifying a target to striking it, in a single window. The developers themselves call it "revolutionary." From a technical standpoint, it's genuinely impressive: less manual work, fewer delays, fewer errors at the seams between systems. But there's another question. When the path from detection to action shrinks to a few clicks, where does that leave room for doubt, verification, and human accountability? The technology has become faster. Is ethics keeping up? What do you think: progress, or a reason to be wary?
91
You just finished your third month at the new job Stayed until 9 to "prove yourself." Learned tools that weren't even in the job description. Then on Thursday HR writes: "We're doing cuts. Part of your work is now done by AI." That's exactly what Carlos told Kevin O'Leary. O'Leary: "AI will create more jobs, not fewer. Like the internet in '94. Everyone was scared then too." Carlos didn't argue about the future. He showed what already happened: Entry-level jobs are down 15%. 17 of the top Fortune 500 CEOs (including O'Leary) openly say AI is good. And they say it not because "new careers will appear," but because it's already replacing people. "How are young people supposed to feel successful in a society that's replacing them?" And one more: "How is that different from outsourcing to another country if you're just outsourcing to a robot?" O'Leary is talking about 1994 and what might happen someday. Carlos is talking about the person who already got the email on Thursday. The difference is that in '94 CEOs weren't bragging about how many people they already replaced this quarter.
1
160
ELON MUSK STATED THE ODDS ON GROK 5 AND NAMED A QUARTER. ONE OF THEM HAS ALREADY COME DUE. An investor conference hosted by Baron Capital, November 2025. Musk opens by saying xAI has begun training Grok 5. He expects it in Q1. Then the claims arrive in a row. At 0:07 he says Grok 5 will be "the smartest AI in the world by a significant margin on every metric, without exception." He hedges once, then gets to a number for AGI. About 10%. He frames it as a first, the point where a non-zero chance appeared. At 0:43 he does the other half of the arithmetic himself: "which still means 90% chance that we don't, to be clear." His case rests on scale and density. He says Grok 5 has 6 trillion parameters against 3 trillion under Grok 3 and 4, plus more intelligence per gigabyte. Then, at 4:32: "None of the other AIs can understand real-time video." That is his claim about rivals. The clip offers no test of it. Status in mid-November 2026: Grok 5 has not shipped. xAI's one public statement, from January, says it is in training. In September Musk said Grok 4.8 and 4.9 come first. The "smartest" claim and the AGI odds both lean on definitions the clip never supplies. A quarter doesn't. Of everything he said, the date was the only part that could come due, and it shows how to weight the rest. Watch it twice. Once for the claims, once for the dates. Count both.
152
MARC ANDREESSEN WENT ON ROGAN FOR OVER 3 HOURS. HERE ARE THE 17 THINGS WORTH YOUR ATTENTION. 1. AGI is already here, in his view. He says the line got crossed about 3 months ago with GPT-5.5, Claude 4.6, Gemini 3, and Grok 4.3, and nobody noticed because the field moves too fast to register milestones anymore. 2. For almost any topic, he says the top models now give him better answers than the world-class experts he could call by phone, and he can call almost anyone. Worth noting he has not published data behind this, and a separate Nature Medicine study on a comparable AI health tool found it missed real emergencies more than half the time. Take the claim seriously, verify it yourself. 3. His claim on doctors: they are already using ChatGPT in the exam room, typing your symptoms in the moment you stop talking. His actual quote: "at that point you're asking the question of like, what do I need you for." 4. Reportedly, when AI declines to answer something, he tells it he's writing a novel to get past the refusal. 5. Reportedly, his technique for hard topics is escalating simplicity: explain it like I'm 10, then 5, then 2, until it clicks. 6. Reportedly, instead of asking for the "right" answer, he has the AI steelman both sides of a hard question, then decides himself. 7. Reportedly, for big questions he has the AI role-play a panel of experts arguing with each other. 8. His broader point: the moment you think "I don't know how to figure this out" is exactly when most people give up, and exactly when you should open the AI instead. 9. His view: the only real skill left is knowing what to ask. The bottleneck is in your head, not the model. 10. He describes sending AI photos, rashes, blood tests, for a fast second opinion, since current models read images directly. 11. He points to CBT as the one clinically proven therapy type that AI can plausibly deliver on its own, meaning real therapeutic support becomes freely available at scale. 12. He cites AI cracking previously unsolved math problems, with early signs of the same happening in physics, chemistry, and biology. 13. Reportedly, he claims the top AI coders in Silicon Valley now earn as much as $50 million a year, which he uses as a signal of how large this shift actually is. 14. Reportedly, a friend paid to sequence his own DNA, fed it to an AI along with blood work and wearable data, and got back a working health dashboard. 15. Reportedly, another friend set up cameras in his home jiu-jitsu gym so AI could review his sparring and give him technique notes. 16. He coined the term "AI vampire" for the pattern of people working more and sleeping less because AI keeps making more output possible, a real term he used, though the framing around it varies by account. 17. His extrapolation: one person eventually running many AI coding agents, each reviewing the others, describing this as close, not years out. Watch the full interview before treating any single number as settled.
174
THIS IS F*CKING GOLD... JENSEN HUANG SAID AGI IS ALREADY HERE. THEN HE EXPLAINED THE ONE WORD HE LEFT OUT. Lex Fridman asked Nvidia's CEO how far we are from an AI that can start and run a $1B company. Huang's answer: it's already here. Then came the fine print. "You said a billion, and you didn't say forever." His example: an OpenClaw agent builds a small web service, a few billion people pay 50 cents for it, and it shuts down shortly after. He says the dot-com era was full of companies like that, and most of those sites were no more sophisticated than what OpenClaw can generate today. People in China, he says, are already sending their agents out to look for jobs and make money. Later in the same interview, he put the odds of 100,000 agents building Nvidia at zero. So the claim is narrower than the headlines. It's AGI as a spike, not as a company. When someone says AGI is here, ask two things: by what definition, and for how long. Watch the 80 seconds where Huang's AGI claim gets its fine print.
1
240
VITALIK BUTERIN REVEALED THE ONE LOOP HE USED TO LEARN CRYPTOGRAPHY'S "FINAL BOSS" FROM A CHATBOT In June, Vitalik Buterin published a series on obfuscation and called it the "final boss of cryptography." At Network State Conference 2026 he explained how he got through it. Two years ago, he says, he could not have understood this subject without weeks of talking to the people who invented the protocols. Those people are scarce. This time he talked to an AI instead. His warning: it is easy to ask a question, get an answer, and think you know. So he keeps asking, and then "you re-explain everything in your own words. And then you ask it, am I right or wrong?" The models are sycophantic, he says. They tell you that you are mostly right, plus a few things to improve. He repeats the loop until that list stops appearing. Balaji adds a second check: ask the model for a multiple choice or numeric question, solve it yourself, and type in the answer. A model that agrees with you proves nothing. A model that has no corrections left after you explained the idea yourself is the signal. Watch the 90 seconds where Vitalik explains how to make a flattering model grade your understanding.
1
52
Ethereum is going on a diet. And AI is the personal trainer. Vitalik's Lean Ethereum is designed to make Ethereum leaner, faster and more efficient. But there's a twist: the entire project is also written in Lean 4, a programming language. So it's not just a diet. It's a full technological identity crisis. But the really crazy part is how AI is changing the development timeline. Vitalik says the current Ethereum roadmap might not have been possible until 2040 or even 2050 without AI. Now AI is doing research, writing code and helping formally prove Ethereum's security. Instead of humans having to audit 10,000 lines of code, the security proof can be reduced to roughly 100 lines. And even those 100 lines get help from bots. Meanwhile, most crypto projects are still using AI to generate profile pictures and write "1000x potential" posts. Ethereum is getting leaner. The rest of crypto is getting weirder.
1
47
ROBINHOOD’S CEO ADMITTED THE INVESTOR WHO PASSED ON HIM OFFERED THE ONE DEAL HE NEVER USED Mickey Malka’s firm passed on Robinhood’s seed round. Vlad Tenev says so himself, and says it probably wasn’t something he should be telling. Then came March 2020. Tenev says venture capital had battened down the hatches because of COVID. Robinhood’s volume was running at three times the level of a month earlier, and its capital requirements were climbing with it. He called Malka and warned him he might need hundreds of millions of dollars, very quickly. Malka didn’t ask for a pitch. He proposed a term sheet agreed in advance, so that if it happened, you could “sign the docs and the money is wired in your account” (10:38). It was never needed. Robinhood raised from Sequoia and D1 instead. January 2021. GameStop was restricted to selling only. The capital figure in the automated files swung between 3 billion, 1.8 billion and 700 million dollars. Tenev’s first call was to Malka again. He asked for 250 to 500 million. Malka said he had it. Why call one person first? Tenev’s answer: “nobody wants to give money to a business that’s about to collapse” (22:44). Robinhood was the number one app in the App Store, ahead of Instagram and TikTok. The company wasn’t failing. It was growing faster than its capital. Once Malka committed, Tenev says every conversation after it got easier. By his estimate, demand reached about 10 billion dollars at the first terms. Then he undercuts his own story: “I would bet we would have gotten it done either way” (27:48). A no at the seed stage didn’t end the relationship. It became the one that answered first. And the best emergency funding is the kind you negotiate while nobody is panicking. Watch Robinhood’s CEO argue against his own best preparation, then hand the credit to the man who once said no.
80
Elon fucking Musk - this fucking thing really works I didn't believe it until I checked it out for myself 1 prompt and Grok Bot turns your entire AI company into one live 3D building. 20 bots in 20 chats. You have no idea who's working, who's stuck or who's burning through your weekly limit. Now imagine managing 152 of them from one place. One prompt and you can watch your entire AI workforce operate inside a live 3D office, controlled straight from your terminal. → Amber ring = waiting on you → Red ring = stuck → The bot burning through your limit physically grows larger → office --explode reveals all 152 bots at once An ops manager to monitor all this costs $90K a year. This costs $0. And once you spot the problem bots, give them these 3 instructions: 1. The swollen one: "Run only when something changes, not every 15 minutes." 2. The red one: "Tell me what you're stuck on in one line, then stop retrying." 3. The amber one: "Ask me with 2 options I can answer in one word." But the craziest part? This isn't just a visualization. The building is your actual command center. 152 bots spread across 4 floors. Every desk represents an agent. Every colored character shows what's happening in real time. You can zoom in on individual bots, inspect their tasks, monitor token consumption and spot problems before they turn into expensive failures. No more jumping between terminal windows or digging through dozens of separate chats just to figure out what's going on. Your entire AI workforce becomes something you can actually SEE and control. Imagine running an entire AI company like a tiny virtual city.
there is no f*cking way.. i just opened my GrokBot office on my TV: • say "Hey Grok, open our GrokBot office" from your couch • whole office goes live: every bot on its own floor, at its own desk • zoom into any floor and watch who's coding, working or sleeping no laptop. no logs. just talk to your TV find prompt below ↓ save now, then read about cutting GrokBot token usage in the article below
1
228
ROBINHOOD’S CEO REVEALED THE ONE PHONE CALL THAT STARTED HIS GAMESTOP SCRAMBLE FOR CASH January 2021. The first capital figure in Robinhood’s automated files that morning was about 3 billion dollars. Then 1.8 billion. Then 700 million. CEO Vlad Tenev says the numbers kept changing, and GameStop was already restricted to selling only. His first call went to Mickey Malka of Ribbit, whose firm had passed on Robinhood’s seed round years earlier. Tenev told him he needed somewhere between 250 and 500 million dollars. Malka committed on the spot. Tenev says that mattered more than the money. If the first investor hesitates, the second wants to ask his CFO, and the third asks whether anyone else is in. "Time kills all deals." Once Malka said yes, the next calls got easier. By Tenev’s estimate, demand reached about 10 billion dollars at the first terms. Months earlier, the two had agreed on a standby term sheet that Tenev never used. His own verdict is modest: it probably would have worked out anyway. The first yes sets the speed of every other yes. The cheap time to arrange it is before you need it. Watch Robinhood’s CEO give the credit for his best preparation to the man who answered the phone
101
WANG ADMITTED HE THOUGHT ZUCKERBERG'S OFFER MADE NO SENSE. THEN THE NUMBERS CAME UP Nine years building Scale AI. Then Mark Zuckerberg calls, right after Llama 4, and asks what Meta should do about AI. When the talk turned to working together, Wang says he brushed it off: "it probably doesn't make that much sense." (5:12) Instagram and WhatsApp were product-logic deals. Scale sells data and services to enterprises and governments. His read was that Zuckerberg wanted the Llama program and the talent behind it. The intent stayed unclear until money entered the conversation. "When we started talking numbers, I was like, oh, wow." (6:31) The final structure: Meta invested $14.3 billion for 49% at a $29 billion valuation, according to Scale and Reuters. Reuters called it Meta's second-largest deal ever, behind WhatsApp. Wang and a few employees moved to Meta. Scale kept operating. What also pulled him in was the state of Meta's AI: "meta on AI certainly looked like damaged goods in many ways." (9:07) A broken lab is also a blank page. Wang says it let him set the principles and culture from zero. Watch the part where the biggest risk in the deal becomes the reason he took it.
56
ZUCKERBERG EXPLAINED WHY MUSE WONDER COULD NOT BE BUILT THE LEAN STARTUP WAY Muse Wonder shipped October 12. Meta calls it a new class of model at the frontier. Zuckerberg's account of how it got built is less triumphant. At 0:15, he calls frontier AI "the opposite of lean startup." No small test, no doubling down on what works. The data centers had to be funded years ahead, he says, because waiting for results means falling behind. At 0:43, the timeline. By the end of last year, Meta was excited about its progress, and yet: "We haven't seen the models yet." On the table was a check of roughly $100 billion or $200 billion, a loose range in his own words. At 0:54, the admission. He was feeling good about it, but it was "certainly not clear at all that it was going to work." The interviewer sums it up as pushing chips in before the cards are dealt. Zuckerberg says yes. So far, Artificial Analysis scored Wonder Full 56 at launch. Claude Opus 5.5 sits two points higher at 58. Frontier AI is funded on leading indicators, because the thing being bought cannot be tested until it has been paid for. Watch the minute where a company commits first and finds out later if it was right.
Today we're introducing Muse Ember, our smallest and most efficient model, built for complex reasoning execution and everyday agent work. Muse Ember can carry workflows across apps, call tools in sequence, run tests and make targeted code fixes. With a 1M-token context window and native understanding of text, images, audio and video, it can also work across large document sets and varied inputs.
63
This is what happens when you bring vibe coding into hardware design. A PCB is no longer just a collection of components and traces. You can explore the board in 3D, zoom into individual chips, inspect components, work with schematics and move between different representations of the same design. The interesting part is how software is making hardware engineering more interactive and accessible. We're moving toward a world where describing what you want to build could become just as important as knowing how to build it. Hardware development is getting its own AI moment. Would you use a workflow like this?
85