yesterday evening, about $6.3 billion left Binance across ten blockchains in roughly twenty minutes. Binance put the total impact near $6.3 billion, with about $3 billion frozen on-chain. Blockchain analysts like TRM and Chainalysis point at North Korea’s TraderTraitor unit, the same group blamed for the Bybit hack. That’s the news. This video is about how it actually works, because this is a cybersecurity channel and “a hacker stole it” is not an explanation, it’s a horoscope. First, the honest disclaimer. Binance hasn’t published a full root cause or explained how the attacker got from the container to signed withdrawals. So what follows is what Binance has said, what the on-chain evidence shows, and how attacks like this typically work. It isn’t a confirmed play-by-play. Here’s what Binance has said. Their initial investigation suggests a compromised DevOps engineer’s laptop gave the attacker access to a service container, where they submitted unauthorized withdrawal transactions, and they found no evidence that a private key leaked. That last part is the whole story. The attacker didn’t need the keys. In exchange security there’s a concept called a signing oracle. Your keys live inside HSMs or MPC systems, and nobody ever sees them. But a service somewhere is allowed to say “please sign this withdrawal,” and the signer obeys. If you compromise that service, you don’t have to steal the key, because you’ve become the thing the key listens to. It’s like not bothering to pick the vault lock because you’ve managed to become the manager’s voice on the intercom. Blockaid’s CEO says the evidence points to compromised signing infrastructure rather than a leaked key. Now, how does a DevOps laptop get owned? Binance hasn’t said, but DevOps machines are premium targets for a reason. They hold cloud credentials, kubeconfigs, SSH keys, CI/CD tokens, VPN access and password manager sessions. Compromise that one machine and you’re standing next to the keys to the kingdom. Threat intel reports on North Korean crews have documented a recurring playbook: fake recruiters, a “take-home coding test” with a trojanized dependency, a fake meeting app that needs an “update.” You run the test on your main laptop because you want the job, and the interview was the exploit. The defensive version of this is simple and free: never run code from a stranger on a machine that holds anything valuable. Use a throwaway VM. From the laptop, the attacker reached a service container. This is lateral movement, and in cloud environments it usually goes through the same boring doors: an overly permissive service account, a long-lived token sitting in an environment variable, a CI/CD secret, a Docker socket exposed where it shouldn’t be. None of it is exotic, which is the scary part. Binance’s co-CEO said this wasn’t a single one-click action and that the attacker did more than borrow one engineer’s permissions. That fits a chain: laptop, credentials, internal network, container, signing path. Then there’s the Ethereum side, which is where it gets fun. In May 2025, Ethereum shipped the Pectra upgrade, which introduced EIP-7702. The attack transaction was the new type 4, which lets a wallet hand its control to a smart contract. Technically, the transaction carries a list of authorizations. Each one is signed by the wallet and contains a chain ID, a contract address and a nonce. When it’s processed, the wallet’s code field becomes a tiny pointer to that contract, and from then on every call to the wallet runs the contract’s logic. The attacker pointed it at a sweeper that only obeyed them. The nonce trick made me say “bro what the fuck” out loud. An authorization only applies if its nonce matches the wallet’s current nonce, and a busy exchange wallet is constantly sending transactions, so that number keeps moving. Blockaid found that the attacker didn’t know the live nonce, so they sprayed 50 authorizations covering consecutive values in a single transaction. One landed. That’s the equivalent of trying fifty keys at once because you know exactly one will fit. This also gives defenders a detection opportunity. Fifty signed authorizations for sequential nonces of a hot wallet is nowhere near normal behavior. That’s a screaming anomaly at the signing layer. A policy engine that sits independent of the requesting service, with rules like “no delegation signatures, ever, for this wallet tier,” would have asked hard questions before a single one was signed. After the delegation, the attacker could drain the wallet at will. The delegation stayed active, so the attacker came back at 21:44 UTC and took another 200 million USDT in fresh deposits without needing a new signature. That’s persistence, blockchain style. The backdoor sat in the wallet’s own code until someone overwrote it. The part that should keep security teams up at night is that this wasn’t only Ethereum. Wallets on several other chains also sent funds out within the same window, using different key types and signing schemes. That tells you the weakness wasn’t in any one chain’s cryptography. It was a shared pipeline, one trusted path feeding signatures to everything. One compromised service, ten chains of blast radius. The fix is segmentation: separate signing infrastructure per chain and per wallet tier, hard per-transaction and per-hour limits enforced outside the requester, destination allowlists, and out-of-band human approval for anything large. On the endpoint side, the checklist isn’t glamorous but it works. Hardware-bound credentials like FIDO2 keys so a stolen session isn’t enough. Just-in-time access instead of standing admin rights. Short-lived workload tokens instead of secrets in environment variables. Network policies so a container can only talk to what it must. EDR and egress filtering on every engineer machine. And separate hardened workstations for anyone who can touch production. The funniest part, if you can call it funny, is that the two biggest exchange thefts in history now both trace back to an endpoint and not to broken math. Exchanges spend fortunes on cold storage and cryptography, and the attacker walks around all of it through a laptop. Somewhere there’s a DevOps engineer having the worst Thursday in human history, and I’m not blaming them, because the real failure is a design where one laptop can reach a signing path at all. Meanwhile my biggest heist this week was taking the last Monster from the gym fridge. And if you’re thinking “I want to actually understand every layer of this chain,” that’s exactly why I built Cyberflow Academy. It takes the guesswork out completely. You start by setting up your environment properly, then you build real things yourself, like a port scanner in C++ and automation in Python, so tools stop being black boxes. Then you go through web hacking, bug bounty, reverse engineering, Android hacking and OSINT, plus the OPSEC guides that show you how to protect yourself the way professionals do. Everything is in order, hands on from day one, and when you get stuck you’re not alone, because there’s the Discord and you can message me directly. Link’s in the description. Binance says users are safe and has pledged to cover the losses from its SAFU fund and its own funds. We’ll see the real post-mortem soon. Until then, go check which of your machines can reach something valuable, because that’s your attack path too. You still stuck in tutorial hell? We got full courses step-by-step as well. Web hacking labs, real attack flow, Not “click here”… actually break the code. Reverse engineering crackmes inside, Ghidra open, ego died. Bug bounty methods that actually scale, Find one vuln that’s grocery money mail. Twenty-nine dollars that’s it, bro. Skill goes up. That’s Cyberflow. Python for hackers, C++ pain, Build your own tools, don’t just complain. New course dropping every month, No content drought, we stack it up. Private Discord killers in chat, Stuck on a bug? We fix that.
Enthropic just filed to go public. And if the stock market values it at the $2 trillion that people expect, about 940 early employees will be worth 9 figures each. >> Money, please. >> Oh, no, no, there's no money. >> And that's just one of the six things I'm breaking down from the AI world that I think actually mattered this week. I'm cutting out all the rest. I've also got early access to Google's new interactive world model, so we'll play with that one in a little bit. And at the end, OpenAI says it broke an AI math assumption that hundreds of research papers currently rely on. I'll go over everything and break down what I think all of it means. Let's start with the Anthropic IPO. So, on Monday, Anthropic filed its S1. Now, if you're not familiar with an S1, it's basically paperwork that a company files with the SEC before it's allowed to sell shares to the public. oftentimes it's like the first time we get to see a company's real numbers like how much it makes, how much it's spending, and who actually owns what inside the company. And it's expected to get listed around November 16th at a valuation of between $1.5 and $2 trillion. Now, I'm not going to go super deep into this S1 document because if you look at this thing, it's wellund and I think 85 pages long. So there's a lot here, but I will sort of zoom in on some of the more interesting things. Starting with the numbers here, if we look at their revenue in 2025, they generated almost 4.6 billion. This is in millions here. So that's $4.585 billion. In 2026 so far through June 30th, they've generated over 16 billion. So in just like half a year, it generated almost four times as much as it did in the entirety of 2025. However, there is one big caveat here. If we look at their net loss, they lost $80 billion in that same time. So they generated 16 billion, lost 80 billion. >> Oh no. >> Now there is some nuance around that loss. Like they did some deals early on where investors lent them money which would show as a loss on their book, but that would later convert to stocks. So that wasn't like money that left Anthropic's bank account technically because the value of some of that debt is going to convert into stocks. I'm not good at explaining this stuff, but it's not like a real picture of how much money they lost going out of like a bank account. But to be honest, this stuff here isn't what I really find the most interesting about all this. I saw this post on X from Andrew Benson which just kind of blew my mind. Who actually gets rich if Anthropic lists at $2 trillion? If they end up listing at $2 trillion, that works out to roughly $118 per share. $199 billion of that goes to the seven controlling co-founders. Another 100 billion of it goes to Anthropic's public benefit fund and then 331 billion goes to employees through options and RSUs. So at about 5,300 employees, that's on average $62 million each. Now, we know not every single employee is going to get the same amount. It's going to depend on how long they've been at the company and things like that, but here's like a estimated breakdown. Take this all with a grain of salt. I don't know how truly accurate these numbers are, but a lot of people are getting very, very wealthy just for being employees of Enthropic. So 300 to 500 employees that were hired by 2023 are going to get between 110 million to 400 million each. 440 to 640 people hired in 2024 are looking to hit 200 million to 300 million each. 1360 people hired in 2025 between 70 million and 95 million each. And the 3,000 people hired in 2026 stand to get between 10 million and 20 million each. 940 people stand to make nine figures on this. Now, if you were one of the lucky people to get in early on Enthropic, not necessarily an employee, but somebody that had the ability to invest, back in April, holders were able to sell at $25.91 per share. Again, it's looking to list at roughly $118 a share, so over four times as much if you invested back in April. A few weeks later, they raised their series H at $58.90 a share. Essentially doubling your money, it seems. Again, take these numbers with a grain of salt. I have no idea how accurate this like tweet is. I didn't doublech checkck the math myself, but I think it's roughly correct. Companies like Amazon and Google also stand to make like hundreds of billions of dollars off of this because they essentially rent compute to Enthropic. And I believe, you know, there's some weird deals where they do some of that for equity in the company. Some of it is actually paid directly by Anthropic, but as a result of their sort of interesting deals they've made over the years, Google, Amazon, Broadcom, these companies are all going to make hundreds of billions of dollars. Also, I also think this is going to be an interesting test around whether or not AI really is in a bubble. Now, we won't find this out immediately, but this IPO is like the first time we'll see whether the people closest to AI like actually cash out or hold on even longer. I think that's going to be a pretty good like bubble test. Now, there is typically like this lockup period where insiders, they can't usually sell like the day a company goes public. Typically, they need to wait something like 6 months before they're allowed to sell. But when those sort of lockup periods start to lift, we'll find out who truly has like strong beliefs in the AI ecosystems that are being built. That kind of shows us that investors were more in it because they saw huge potential to cash out really quickly and not because they necessarily believe AI is going to get a whole heck of a lot bigger than it currently is. So, we'll see how that plays out. Now, you're probably wondering why did I lead with this story? Like, why should you care about this? But, I think this is a really, really big turning point when this IPO happens. This is like the first time that a Frontier AI lab is going public. Now, you can argue Google, but they have like a whole bunch of other business ventures. Their AI is just like one piece of Google. This is a purely AI company going public and it's really really going to show a lot about like investors real beliefs in this AI that they're building. So when this stock goes live, it's going to be really fascinating to see what investors do. Are they going to pump it? Are they going to dump it? That will be a clear signal of like the true feelings towards AI. Another really cool update recently is from Warp. They just launched Warp Factories, which basically lets engineering teams build an entire AI powered software assembly line in the cloud. So, I wanted to test this on something I do all the time, writing YouTube scripts. My goal is to build an AI script writer that can write an entire 20inute deep dive video in just one shot. But AI still struggles with sounding exactly like me. And it also needs to think about what visuals will be on the screen since this is for a YouTube video. So, with Warp, I can set up scoring agents to evaluate each of those metrics that I care about. And then I have a self-improvement agent that identifies recurring problems and suggest changes to make the script writer even better. Then I can track the scores over time and see if it's actually improving. If you're using this on an engineering team, these same systems can help you build more reliable, cost-effective AI workflows at scale. So, check out Warp Factories at the link in the description box. And thanks so much to Warp for supporting my channel and sponsoring this portion of today's video. Three big new model launches that happened this week. Let's start with the new model out of OpenAI, GPT 6.1 Terra. Now, if you look at the various models that OpenAI has, they have their Soul models, their Astro models, their Luna models, and their Terra models. The Astro models are essentially like their state-of-the-art top-of-the-line publicly available models. The Soul models are the tier below that. The Terra models are the tier below that and the Luna models are the tier below that. So if you were to compare them to like Anthropic, Astra is sort of semi-equivalent to what Anthropic's Fable would be. Soul would be sort of the equivalent of Opus. Terra would be the equivalent of like a sonnet from Anthropic and Luna would be equivalent to like their haiku, their smallest, cheapest but you know dumbest model. So this new model we just got is the 6.1 version of Terra. So theoretically cheaper than Soul and Astra, but also maybe not quite as smart. I don't totally understand why you would actually use this model. I know that sounds harsh, but when I look at some of these charts here, MMU Pro, for instance, on this chart, we can see GPT 6.1 Soul here highlighted in blue, and we can see 6.1 Terra here. Okay, so if we're using 6.1 soul on its highest mode on max, it costs 0013 cents. So just barely above a penny. And it gets an intelligence score on MMU Pro of 86.3%. Now this new model GPT 6.1 terra as opposed to 013 cost 012. So a fraction of a penny less, but it's dumber. So, if you're using it on its max, you're essentially paying the same price. I don't know. Like this chart just kind of when you line these up side by side, it's cheaper at these modes here. The cost is still less than a penny on each of these, but this is just a slightly dumber model. And I mean, a lot of these charts show very similar stuff. So, OS World 2.1, GPT 6.1 Soul here showing a cost of $1.15 for a score of 78%. Now, if we look at Terara here, $119 for a score of 74.2. So, 6.1 Terra costs more for a lower score on max effort. Coming over here, let's say you're using medium effort. Medium effort costs 65 cents for a score of 73.5, where medium effort here costs 50 cents for a worse score of 66.7. But like if we're looking at cost to intelligence, GPT 6.1 Soul Medium does better than GPT 6.1 Terra extra high at basically the same price. Terminal Bench 4.0, same story. $1.98 for a score of 55.9% on 6.1 Soul. Terra for quite a bit lower score is $1.84. So like not that big of a price difference for quite a bit better result with 6.1 Soul. At the end of the day, I guess I don't totally understand who this model is for. Is it primarily like a speed thing? Now, if we look at artificial analysis here, which is sort of like this holistic all-encompassing benchmark. Here's GPT 6.1 Soul at 55, and here's GPT 6.1 Terra at 51. So, I mean 6.1 is an objectively smarter model. So, let's see. Um, cost per task, Soul is about 127 cost per task versus Terra at 96 cost per task. So, it is it's slightly cheaper on a cost per task basis. Time per task, it's slightly very very very slightly faster than Astra and very slightly faster than Soul, but not by a whole lot. This model stumped me a little bit. It doesn't feel particularly smarter than Soul or particularly cheaper or faster than Soul. So, if you're going for like cheap and smart, why would you not just use Soul? And if you're going for, you know, best performance, you're going to jump to Astra. And then if you're trying to run stuff on device and get like the fastest, cheapest models, that's when you would jump down to Luna. So, I'm beating a dead horse here. I just don't quite understand Terra or who it's for. When I do run these on VC Bench, I run them on like their max effort level. So, this one took 7 minutes, cost me 19 cents, used 38,000 tokens, and my LLM as a judge put it in fifth place. As far as generating SVG images go, OpenAI pretty much dominates my leaderboard with one outlier being Gemini 4 Argon here, which is really interesting. But the GPT models have really really sort of stood out as the best like SVG generation models, at least specifically for generating Gary Buucy. But moving along, I said we got three new models this week that I think are interesting to talk about. The next one is Gemini 4 Carbon. This appears to be a step up from the Gemini 4 Argon model we got a couple weeks ago, but unfortunately it's not publicly available yet. This was another like announcement of a future announcement. Like Gemini 4 carbon is done and some people can use it, but probably not you. However, this model looks really impressive. I'm actually going to jump over to artificial analysis here really, really quick because this is actually the new leading model on artificial analysis. We could see it having this checkered pattern means it's not actually publicly available in the API yet, but it now scores higher than every other model here. Better than Opus 5.5, better than GPT6 Astra. It is the new state-of-the-art according to artificial analysis. Not the cheapest model by any means, cheaper than most of the anthropic models, but more expensive than pretty much all of the OpenAI models. And if we look at the benchmarks here, we can see Argon, which was their previous model, and they also compare it to GPT6 Astra, Opus 5.5, and Fable 5.1. So these three here sort of represent state-of-the-art, best of the best in this class of models here. And you can see everything in blue is where this one wins. But again, this one's coming soon to developers, businesses, and consumers. So not available for most people yet. They've made it accessible to just like some cyber security testers and early testers for now. So, I'm not going to go much deeper into this one cuz I can't show you any other benchmarks or tests that I've run with it. So, yeah, we learned about a new model out of Google that we're going to get soon that is now going to be the state-of-the-art when we finally can get our hands on it. And there's one more model that I want to talk about and that is Kimmy K3.1 Pro out of Moonshot AI. Now, this is one of the Chinese labs that, you know, Anthropic and the other companies have accused of distilling their models from, but this model is significant because it is now the new state-of-the-art top-of-the-line best openweight model available. Well, it will be soon. According to them, open weights are going to be available on November 4th. It's a 2.8 trillion total parameter model with a 1 million token context window. And we can see a look at the benchmarks here on Terminal Bench 4.0. It actually apparently performs better than GPT6 Astra, but not quite as good as Opus 5.5. Now, they're obviously cherrypicking the benchmarks that make this model look the best. But, as you can see, it's comparing it to other state-of-the-art models, not really many other openweight models. They do have GLM 5.3 here and Deepseek here, but you can see that Kimmy K 3.1 blows those away in pretty much every benchmark that it's compared to. And for the most part, it's holding its own against the top-of-class state-of-the-art models. Taking a peek here back at artificial analysis, Kimmy K3.1 Pro falls all the way up here with a score of 46. To put this into perspective, Fable 5.1 was just a few months ago blowing everybody's minds. Like people were freaking out that the whole internet was going to get hacked, everything was going to break because Fable was just too powerful. And here we are with Kimmy K3.1 scoring just three points below it on artificial analysis. So, an openweight model that's nearly as good as what was absolutely blowing people's minds just a few months ago. As far as openweight models go, the next closest one is all the way down here at Quinn 3.8 and Deepseek, which are down in the 30s at 32 and 34. Really, really big leap for openweight models. Now, I did manage to toss this one into Buccy Bench. And here's what it generated. It cost 28 cents, about 28,000 tokens, and took about 6 1/2 minutes to generate. But it looks pretty decent. My LLM as a judge ranked it down here at number 11. So just falling out of the top 10, but like artificial analysis falling just behind Claude Fable 5.1. Now this whole open weight conversation is really interesting to me because the whole idea here is with these openweight models, you could kind of do what you want with them. Theoretically, you can download them, you can run them locally, you can do fine-tunes on them so that they're more optimized for your specific use case. And the idea is you're not needing to send them off to like the big companies like OpenAI, like Google, like Anthropic where they can train on your data and potentially store things you've said and keep it in their memory and things like that. So there's like privacy and security and, you know, data retention issues that sort of get mitigated when you have really good openweight models. The thing is this one's a 2.8 trillion total parameter model. That all needs to get stored in memory. Even with 104 billion active parameters, it still really kind of needs to store the whole thing. So to put that into perspective, like the best of the best gaming graphic cards, like your Nvidia 5090 might have like 32 GB of RAM in order to run a 2.8 trillion parameter model, you'd need about 44 Nvidia 5090s. If you had a maxed out Mac Studio with 512 GB of RAM, you'd need at least three to run this model. And those things run like 10 to 15 grand each. So yes, while the dream of open weights is to be able to have like the smartest models running on your own home machines, it's still not really feasible at the sort of size of these models. So, you know, realistically for the normal person, you're not going to be running this model on a home computer. Now, I do believe these like larger openweight models still matter. They just don't matter for running on like your home PC or laptop. things like hospitals and banks and governments can run servers that it controls. So data never goes to anybody else's data centers. So like the privacy argument still kind of hold up if you're capable of having systems on premises that can run these kinds of models. But I do think it's important to talk about these models just to show that they are catching up. And it's probably only a matter of time before we get smaller and smaller, more efficient models that are really smart but start to fit on devices that we actually own at home. I mean, with memory cost the way they are, it might be a while before we are able to afford the devices at home, but eventually these models will get more efficient and we'll have home devices that are capable of storing more memory. All right, so the next thing I want to share is probably the coolest thing I saw this week and that is Genie 4 out of Google DeepMind. Now, this is another one that is in early access only. You can join a wait list and they're letting people in slowly, but I did manage to get early access, so we'll actually be able to play with this one. But real quickly, what Genie 4 is is a world model where you can prompt any scene and any character into that scene, and then you can move around in that scene, move the character around in that scene. And in this newer version, you can actually prompt changes. So you can tell it to put a dinosaur over there or, you know, make a cat falling from the sky here. So as you're moving through this world, you can prompt new things into the world. This one also adds sound. Previous iterations did not have sound inside of these world models. This one, well, you'll see it does. Now, they also claim it got a lot more realistic, which in some of my early testing did show some scenes looked a lot more realistic. Other scenes still left a little to be desired. Uh, but for the most part, it does look a lot better than the previous iteration. But I'm not going to just talk about this one. Let's go ahead and jump in and play with it. >> That's what she said. >> So, let's go ahead and create a world. We have the option to like roll a dice and it will pick a world for us. Doesn't seem to randomly pick a character for us. For the environment, I'm going to do a colorful Japanese city at night in the rain with cyberpunk neon signs. I want to see some like reflections and stuff for our character. I'm going to do a ninja. Let's see what it generates for us. You can create it with a sketch or without a sketch. I'm just going to roll the dice and jump in. So, I can control with WD, I can use my arrow keys to move around. And here's our scene. You can actually see in the bottom left when I press a key or in the bottom right when I press one of the arrow keys to like move around the scene. So, right now I'm moving backwards. If I move forwards, you can see I get closer to our ninja here. And then over on the right, you can it gives me some ideas of things I can prompt into existence, right? So, I can do a police drone. If I click this, you'll notice at some point a police drone's just going to fly into the shot. Now, I can also type a prompt, which I find kind of annoying cuz then I can no longer uh control the camera. But if I wanted to, I can be like, "Godzilla stomps through." Submit that. And now we have Godzilla busting through the wall and stomping into the scene. Uh, let's have Godzilla swat the drone. Let's see if it'll do it. There we go. Godzilla. Well, tried to swat the drone and it froze on me. Something went wrong. Please try again or send feedback. So, you know, it's still very very beta testy. This time I'm going to select first person down here. Let's do a cartoon Candyland world and a character. Let's do a bumblebee. So, here's our first person view from a bumblebee that I can actually control in this Candyland world. So, as I push forward, he moves forward. As I push to the right, spins the camera to the right here. And I can just move around in this candyland world. Now, there are some issues that I've noticed. Like, if I turn all the way left, notice how there's a candy cane on the right right now. And there's some flowers. Notice what those look like. If I do like a full 360 spin here and then try to come back around, is that stuff still going to look the same or is it going to be completely changed? Okay. So, some of it's still there, but I seem to have moved further away and now there's a cupcake in the scene. So, there's some like coherence issues you'll notice in this for sure where if you move around the scene and look around a little bit, stuff will change. We got our uh cookies waving at us back there. If I rotate around from the other direction, I mean, there's still some cookies. I think they changed positions a little bit. Let's do a uh lollipop signal switch. Oh, there's our signal switch. Don't really see the lollipop, but we got a signal switch. And notice how there's now music. There's sound effects as the train comes through. And uh yeah, I mean it's it's pretty cool. You also only get about two minutes. So, you can see our little timer up here is counting up. Once it hits two minutes, it pretty much just freezes and boots you out. So, it's not the longest thing in the world. I'm not totally sure who this is for. I can see probably the biggest demographic might be filmmakers that want to like rough out a scene before they shoot it. They can generate a world. They can generate their character in the world. They can sort of flesh out what it might look like if this thing enters the scene. They could flesh out what it would look like if, you know, Godzilla crashes through a building or something like that and kind of get a rough idea of what that scene will look like before maybe going and shooting it in real life. I don't think the quality is really good enough to use for like the actual video footage. But maybe if you're just kind of planning through things. I can also see game designers wanting to use something like this where they can test an idea in minutes and just see if it works without having to like mock it all up in Unreal Engine or something like that first, mock it up in this and if it looks good and you get a better idea of what you're going for, then go build it in Unreal Engine or Unity or whatever. I can also see use cases where like robotics companies can drop in a robot into one of these worlds and use it for like training purposes. Like that's one of the big use cases of the world models that Nvidia makes is that they actually train robots inside of world models before putting them into the real world so that the robots know how to interact with things once they actually get into the real world. But I don't really see this as like a game where you're going to get in and sit down and play with it for hours cuz a you only get 2 minutes, b there's some coherency issues still and c it's really just kind of like a demo test kind of thing. It's like a showoff of the capabilities and where all of this is headed. But I imagine, you know, give it a year or so and we'll be able to create 20 minute videos that are completely coherent and you can make things pop into the screen. Maybe you can voice prompt what you want to see. Like, hey, on the right side of the screen, make Godzilla uh pop in, right? Like, I can see that is where they're headed with all of this. And in this final story that I'm going to talk about, OpenAI did something that on the surface feels pretty insane. Now, I'm not like a mathematician, and this was really hard for me to understand. So, I'm going to do my best to explain what's going on here in a way that finally sort of clicked for me. So, on Monday, OpenAI said one of its unreleased internal models disproves something called the exponential time hypothesis. This is a 25-year-old assumption in computer science. So, imagine a row of light switches. Let's say you've got like 10 different light switches. And now imagine you have a list of rules like if switch three is on, switch seven is off. If switch 9 is on, switch six is off. And let's just say you have a ton of rules like this, all with different sort of criteria. Your job is to find a setting on these light switches that does not break any of these rules. So with 10 switches, there's about a thousand different combinations. With 30 switches, you have about a billion different combinations. Every switch you add doubles it. And by a few hundred switches, there are more combinations than there are atoms in the universe. Now, computers solve puzzles like this constantly for things like scheduling and chip design and testing software. And for 25 years, computer scientists have assumed that for the hardest versions of this, every method eventually just gets bigger and bigger like that. That's an oversimplified version of the assumption called the exponential time hypothesis. Now, nobody's ever actually proved this hypothesis, but people assumed it was true because nobody could find a way around it. And as a result, hundreds of computer science papers rest on this model. But again, according to OpenAI, its model found the disproof on October 16th. So now imagine all of these hundreds of computer science papers that rest on this theory. basically can now all potentially be refuted. Now the proof was verified in lean and what that means is like you can sort of think of lean as like this spell checker for math. A computer's going to verify every step that follows from the one before it. So like a human could do a bunch of stuff and you know let's say it was a 40step process and they screwed up something on step seven. Everything after that's going to be messed up, right? But lean when you run it through this it won't miss a step. It's not going to let a step slide. So by having it verified through lean, it's saying that like we backtracked this whole thing and it all checks out essentially. So as far as like why this all actually matters, AI has been mostly great at things humans already know, right? Like uh passing exams or writing code or doing things we already know how to do. This would actually be AI answering a question that experts haven't been able to so far. So, if this actually holds up, it's like way bigger than any benchmark score we've seen so far. Again, a lot of this is going to be an oversimplification. If you want to dive deeper, take this article that OpenAI put out about working with academic communities to build scientific understanding, the exponential time hypothesis. And uh, you know, read through it yourself and maybe even throw it into something like chat GPT and get it to explain it back because I don't know if I did a good job and everybody sort of learns in different ways. But I did my best based on my understanding because this feels really big. It feels like we're getting to a point where AI is starting to solve problems that humans haven't solved yet, which is what we all really want it to do in the end anyway, right? Like helping us write emails and, you know, check off things on our to-do list is cool, but solving real world problems is where like AI's real potential still stands and where most of these big tech companies say that's where we're trying to get to. But that's what I got for you this week. A little bit different of an episode. I cut out some of the rapid fire. I cut out some of the other random stuff and decided, let's narrow in on just the most important things and cut out all the fluff and all of the other like random product announcements from random companies that are baking AI into their products and things like that and talk about the stuff that feels like it really mattered this week. So, let me know if you liked this version. I think it's important to sift through all the fluff, all the hype, all of the announcements. There's hundreds of announcements every week. To be honest, I list them all for you. So, if you want to know all of it and you want to get the same stuff that I would have shared in a rapid fire, head over to futuretools.io/news and I literally list every bit of news that comes out through the week here. So, you could kind of get the exact same thing I used to do in a rapidfire by just looking at this news page and skimming these real quick and you'll get updated on every single piece of news that came out. But I think for a lot of these news videos going forward, I'm going to reserve it for the things that actually feel like big needlemoving moments. Hopefully you found that valuable. You feel more looped in on what's going on in the world of AI. I'll drink from the fire hose all week and then tell you exactly what you need to know. Thanks so much for hanging out with me, tuning in for this one, and nerding out. Thanks again. Hopefully I'll see you in the next one. Bye-bye.
Kling 4.0 is finally out, but is it good enough to beat SeaDance 2.5, one of the best AI video models right now? To test this, I put them head-to-head on the exact same six prompts and scored every single round out of five. In this video, I'll show you every result side-by-side with a running score and what each one costs. And by the end, you'll know exactly which one to use for the videos you make. The first round is realism, and I'm testing it with the simplest scene I could think of, a guy making breakfast. That's intentional. With a simple and relatable scene, we can easily notice all the details. Everything in this video happens inside OpenArt, an all-in-one platform with the best AI image and video models in one place. Thanks to OpenArt for sponsoring this video. To follow along, you can find a link to OpenArt in the description below. Use code YURI10 at checkout for 10% off your OpenArt plan. And when you sign up through my link, you'll also get my OpenArt VIP package completely for free. It's a collection of tools I built that includes the AI script writer to write your scripts, the consistent character maker to keep your character looking the same in every shot, the scene to shots planner to split scenes into shots and write the prompts, and the style matcher to copy an image's look, plus whatever I build next, all for free. The only way in is to sign up through my link in the description below. OpenArt costs exactly the same either way, so you might as well get the VIP package on top of it. To make sure I can actually get you the VIP package, though, clear your browser's OpenArt cookies and site data before clicking my link. You can do this in your browser settings by going to OpenArt and deleting it through here. If you don't, your purchase could be attributed to someone else, and I won't be able to unlock the package for you. When you first log in, go to tools in the menu in the middle. I'll click on video, then I'll click on text to video, and from the model selector, I'll start with SeaDance 2.5. I'm keeping auto polish off. Otherwise, OpenArt rewrites your prompt for you, and the two models wouldn't be getting the exact same text. The audio stays on with one output, and in the output settings, I'll go with 16x9, 1080p, and the full 30 seconds since both of these can now make a 30-second video in one go. I'm scoring every video out of five on three things: how real it looks, how closely it follows the prompt, and the one thing that round is built to test. The overall score is the average of those three. The price stays out of the scores until the end. Otherwise, the cheaper model would pick up extra points in every single round. Then, I'll paste in the first prompt. It asks for a man in his 40s making coffee in a moka pot, eggs and toast, filmed like a quiet documentary. You can find this exact prompt and the whole workflow in the description, completely free. SeaDance comes out well. The only comment I have is that the man cracks only one egg, yet two come out. But, the textures and the ambient sounds make the whole scene feel real, and that's what this round is testing. That's a four across the board for SeaDance. Then, I'll switch the model over to Kling 4.0 with the same prompt. Kling doesn't have an audio switch at all. It adds the sound on its own. Switching models can also change your output settings. Open them again and make sure it's still 16x9, 1080p, and 30 seconds before you generate. Kling's version is definitely lacking compared to SeaDance. He puts the moka pot down next to the burner instead of on it. Then, a few seconds later, it's on the burner with no lid, and the toast pops up without ever being toasted. Here, we also have the same problem of him cracking one egg, yet two come out. By the time he sits down, the eggs are gone from his plate, too. Even if you don't notice these details, the whole video has that plastic look that instantly makes it look generated. The one thing it gets right is the sound, with the egg sizzling the second it hits the oil. That's a two for Kling on all three, and two out of five for the round. The first round goes to SeaDance, four to [music] two. And if you're making anything with everyday actions in it, that's a big gap. But, nobody says a single word in this scene. And once people start talking, it's a lot easier to tell if a video is made with AI. For the second round, I'm testing how good these models are with dialogue, using three friends sitting around a campfire teasing each other. When someone's talking, their mouth is the first thing you look at. If it doesn't match the words even a little, the whole scene looks fake. I'm keeping the same settings as before, but this time, I'll start with Kling 4.0. Then, I'll paste in the second prompt. It asks for three friends around a campfire at sunset joking about how long it took to light the fire until one of them burns his marshmallow. The prompt also says only one person talks at a time and their lips only move when they're speaking. And that makes it easy to spot when a model gets it wrong. >> 40 minutes. 40 minutes for one fire. >> [laughter] >> Because someone started with the wet wood. >> It wasn't wet. It was a little damp. >> Well, it's burning now. Who wants the first one? >> Me. I earned it. >> [laughter] >> Not that close. >> [laughter] >> Kling's version has the same plastic look again. The lip sync itself isn't the best, but it works. But, the worst part is the way they talk. It just feels really robotic and fake with zero emotion. Small details like the marshmallow burning in an unnatural way also drag this down more. So, Kling gets a two for realism, a three for following the prompt, and a three for the lip sync, which comes out to 2.7. Next, I'll switch back to SeeDrones 2.5 and run the same prompt. >> 40 minutes. 40 minutes for one fire. >> Because someone started with the wet wood. >> It wasn't wet. It was a little damp. >> Well, it's burning now. Who wants the first one? >> Me. I earned it. Not that close. >> [laughter] >> This one actually sounds like three friends talking. The voices are natural and the lips match the words a lot better than Kling's. There are a couple of mistakes. Nora switches sides between two of the shots, and when the marshmallow catches fire, their reactions look a bit forced. But overall, the conversation sounds real, which is the whole point of this round, and SeeDrones gets a four on all three. That's four to 2.7, and SeeDrones is now two rounds ahead. But both of these scenes were trying to look like real footage, and the next one doesn't need to look real at all. For the third round, I'm switching to animation with the same kind of everyday story, but made as a full 3D animated film. Since nothing here has to look real, I'm scoring style instead of realism this time. This time, I'll go back to SeeDrones 2.5 first with everything else the same. Then, I'll paste in the third prompt. It asks for a little boy at a market who insists on carrying a huge watermelon by himself until he trips and it rolls away. >> You sure you don't want help? >> I got it, Grandpa. >> Together then. >> SeaDance's version is decent, but it's not nearly as colorful as you'd expect from an animated movie like this. The expressions are nice though, and you can really see the kid struggling to hold the watermelon. The story is where it goes wrong. He steps over the stone just fine and then falls anyway. Then the watermelon just stops before the grandpa stops it and the pacing is a little off. At the end they don't carry it together, they just walk while the watermelon drags along next to them on its own. In an animation test, the colors are the first thing you notice. SeaDance gets a three for style, a two for following the prompt, and a three for the animation, which is 2.7. Next, I'll switch over to Kling 4.0 with the same prompt. >> You sure you don't want help? >> I got it, grandpa. >> Together then. >> [laughter] >> This is the first round where Kling is better. The colors pop and they look like an actual animated kids movie and the voices fit the characters really well. His fall still isn't realistic, but it's a lot more cartoonish and for animation that works. When the watermelon rolls off, it heads straight for a group of pigeons and it's the curb that stops it instead of grandpa, but it's still more believable than what SeaDance did. That's a five for style, a four for following the prompt, and a five for the animation, making it 4.7 for Kling. Kling wins this one 4.7 to 2.7, bringing it to two rounds to one. And since it's also a lot cheaper, for animated videos Kling is the easy pick. But a cartoon can get away with movement that isn't real. So next, I'm testing what happens when it has to be. For the fourth round, I'm testing motion and physics with a moving day in Brooklyn. As the viewers, we know exactly how elements should move and interact, and anything that's unnatural instantly stands out. Kling goes first again with the same settings as always. Then, I'll paste in the fourth prompt. It asks for two friends carrying a heavy sofa down the stone steps in front of their building while their friend follows with a box of fruit. The sofa is supposed to nearly slip on the way down, and when the bottom of the box splits, she has to stay on her feet while dozens of apples and oranges bounce down the steps. >> [sighs] [laughter] [laughter] >> Kling's version is honestly pretty bad. The way they hold the sofa looks fake, and they never seem close to dropping it. Then, instead of the box breaking, she just bends down and drops the fruit, and it looks like she's doing it on purpose. The fruit also lands way up on the steps and then suddenly jumps to the middle. In a roundabout physics, almost nothing here moves the way it would in real life. Kling gets a two for realism, a two for following the prompt, and a one for the physics, which is 1.7. Next, I'll run the same prompt on C-Dance 2.5. >> [laughter] >> C-Dance does better, but it struggles with this one, too. The moment the sofa almost slips and they catch it actually looks real, but then they put it down like they're throwing it. The way the girl carries the box looks natural, but the fruit falling out looks fake again. Way more fruit comes out than that box could ever hold. Finally, when the guys collect the actual fruit, a lot of it seems to disappear into their clothes. Since physics is what this round is testing, [music] that's only a two there with a three for realism and a three for following the prompt, making it 2.7 for SeaDance. SeaDance takes it 2.7 to 1.7, bringing it to three rounds to one. But with that many things falling and bouncing, neither model is reliable yet. But every character up to now has been for a one-off generation. Let's test how consistency holds across multiple locations. For the fifth round, I'm following one woman through her whole day on the bus, at the office, and back at home. If her face or clothes change even a little between those places, she stops looking like the same person. For this one, I need a photo of her first. I'll go back to the navigation bar and click on image. From the model selector, I'll pick GPT image 2.5 with the sunburst mode. Then I'll set it to 3x4, 4K, and high quality with auto polish off again. I'll paste in a prompt for a full body photo of a woman with wavy auburn hair, freckles, and an orange jacket. I gave her really specific details on purpose, and that makes it easy to see if the videos keep her the same. What comes back is a full body shot of her against a plain gray wall with her whole outfit visible. Back in video with SeaDance 2.5 selected, I'll upload that photo under add visual references. OpenArt names it at image one, and the prompt uses that same tag. And once it's linked, you'll see a small thumbnail of the photo right inside the prompt. That's how you know the prompt is actually using the photo. Then I'll paste in the video prompt. It follows her through one ordinary day with the light changing in every place. >> In our killer >> You know what you should done? >> Take that off. >> [snorts] >> This is SeaDance's best result up to now. She looks like the same person in every scene, even when the light changes from the cool morning on the bus to the warm lamp at home. The close-ups look real, too. The only mistake is when she gets home and takes off her shoes. While she pushes them back, two shoes turn into three. But, that's small compared to keeping her the same across three completely different places. That's a four for realism and a five for both following the prompt and consistency, which is 4.7. Next, I'll switch over to Kling 4.0. And even though the model list doesn't show a reference tag next to it, you can still add the photo the same way. >> [sighs] >> Kling also keeps her the same in every scene, but there's a lot less movement than in SeaDance's. And when she walks into her apartment, her office is still behind her. So, Kling gets a three for realism, a two for following the prompt, and a five for consistency, which is 3.3. So, SeaDance wins this one 4.7 to 3.3, which makes it four rounds to one. Both keep the character the same, but SeaDance makes it look more real. For the last test, instead of checking one thing at a time, I'll combine everything and see each model's true potential. The last round is a busy Italian restaurant where a waiter drops a full tray of plates and wine in front of everyone. It has a crowd, a moving camera, things breaking, and people talking. If a model is weak at any of those, it'll show up here. I'll start with Kling 4.0. Paste in the last prompt and generate. >> Tom, that was table six. >> I know. Sorry, chef. Get them up. I'll make them again. >> The overall result is honestly better than I expected, but it still misses the details. The waiter walks out with two trays instead of one, and the kitchen door opens by itself. The camera following him looks good though, and so does the moment the chef comes out. But then, a chair appears out of nowhere, the kid pushes it into the waiter, and the chair disappears again. The drop doesn't look real at all. Most of the plates don't even break, and there's wine on the floor, but no broken glass. The lip sync and dialogue is pretty good though, and it at least attempted to hold everything together. For around the test everything at once, Kling gets a two for realism, a two for following the prompt, and a two for putting it all together, which is two. Then, I'll switch back to SeaDance 2.5 for the last generation of the whole test. >> TOM, THAT WAS TABLE SIX. >> I KNOW. SORRY, CHEF. Get them up. I'll make them again. >> SeaDance handles the details a lot better, but there's a crucial mistake. The camera follows the waiter more naturally, and when the tray falls, it actually looks heavy, with the plates breaking into the pieces on the floor. But nothing ever bumps into to waiter, and he just drops the tray on his own, which is not at all what I asked. Right after the chef leaves, the waiter looks straight at the camera and smiles, like he dropped it on purpose. The story goes into a completely different direction. That's a three for realism, a 1.5 for following the prompt, and a four for putting it all together, which is 2.8. SeaDance wins the last round, but it was a lot closer than expected with a 2.8 to 2, and finishes the six rounds five to one with an average of 3.5 against Kling's 2.7. But on price, Kling is way ahead. 1/32 video at 1080p costs 9,625 credits on SeaDance 2.5, while Kling 4.0 is 2,700 right now with the 50% discount, or 5,400 without it. Even at full price, Kling costs a little over half of what SeaDance does, and with the discount, it's less than a third. If you're making anything professional that needs to be realistic, especially people talking or the same character across different scenes, SeaDance is worth the extra credits. But for animated videos, Kling actually looked better. And when you want to create videos in bulk, or your budget matters more than getting the most realistic result, Kling is the one to use. The best part is that since both of them are in OpenArt, you don't have to pick just one. You can run your realistic scenes on SeaDance, and your animated ones on Kling, all on the same credits. Use my link in the description to sign up to OpenArt and start making your own AI videos today. Thanks for watching, and I'll see you in the next one.
I built an entire app for an ad agency just by chatting with Claude. My team can move campaigns across a board and clients can log in to approve their [music] own work. Today, I'll show you how to do all of this with a real database and user accounts without writing a single line [music] of code. Let's get into it. The tool doing all my app structuring under Claude here is Softr, and they're sponsoring this video. If you want to follow along, the link in the description gets you 200 AI credits. I asked Claude for a campaign board for an agency, and a couple of minutes later, I have a good-looking page. It has a column for each stage and cards with client names and budgets. At first glance, it looks like a finished tool. Now, try to actually use it. I refresh the page, and every card I moved jumps back because the campaigns are written straight into the file. >> [music] >> There's no login, so I can't send this to a client without showing them everybody's budget. And if two people on the team open it, each of them gets their own copy. This is how a lot of internal tools get made right now. Someone in marketing or ops builds a page like this with AI, and the team starts relying on it. Then IT finds out, and they rewrite the whole thing so it's secure. The page itself was never the problem. What it's missing is a database and real user accounts, and we'll add both from the same Claude chat. Claude talks to outside services through MCP, the model context protocol. It's Anthropic's standard for connectors, so it works the same way as hooking Claude up to your calendar or your docs. In Claude, I open customize, go to connectors, click add, and choose add custom connector. I paste in the Softr MCP URL and sign in. That's a normal OAuth login, the same one you get when an app asks to connect to your Google account. Then I check what Claude can do now. The list has Softr tools for databases and workflows. Another set handles pages and user groups, and there's one for vibe coded blocks. From here on, when I ask for a table or a page, Claude builds it inside the app instead of explaining how I could. I start with the data because everything else sits on top of it. I name every table and every field in the prompt, including the exact stages a campaign goes through. If I skip the stages, Claude invents its own and I spend 10 minutes renaming them. Claude works through it one table at a time. First comes clients, then campaigns with a link back to the client. After that, it builds tasks and fills everything with sample records. When it finishes, I open the data tab in the app to see what it actually did. The links between the tables work and the stage field is a drop-down with my six stages in order. I also open one client and see its campaigns listed right there. So, the link goes both ways. Claude can read this data, too, so I can ask it questions right in the chat. I ask which campaigns are overbudget and it compares spent against budget. It gives me a list that shows how far over each campaign is. Normally, an account lead answers that by opening a spreadsheet and sorting by hand. The team needs to see the pipeline, so next comes a Kanban board. This one goes through vibe coding. Claude writes a custom block, basically a small React component, and publishes it as a page in the app. >> [music] >> Unlike my page from the start, this block reads and writes the campaigns table. >> [music] >> Now, I drag a campaign from internal review to client review. Then I open that campaign in the data tab and the stage says client review. Next, a small change from the chat. I open the block in the editor first so you can see what Claude actually made. The source tab points at my campaigns table and the code tab holds the React code if you ever want to read it. Now, I ask Claude to make the cards compact and add a budget badge. It edits this same block instead of building a new one and after a reload, the board has the new [music] cards. Every version gets saved. So, if I don't like a change, I pick the earlier one in the version menu and hit restore. Now, I'm building the dashboard for the agency owner. It isn't always obvious how to write this prompt, and the first result is rarely perfect. [music] If I just typed build a dashboard, I'd get a few generic counters. The AI won't read your mind. It does exactly what you ask. So, I brief it the way I'd brief a developer on my team. That gives me a solid first version. Then I add detail one request at a time. I group the active campaigns by stage and put a progress bar on each one based on how many of its tasks are done. After that, I flag anything past its deadline or launching this week. While you iterate, you'll catch things that look broken. Here, the dashboard shows 40 overdue tasks out of 40. The block itself isn't miscounting anything here. Claude gave every sample task a due date in the past. [music] For the demo, I ask it to spread the dates over the next 3 weeks, and the count drops to something believable. None of this means starting over. Claude can read what's already in the app through the connector, so every fix is one more message on top of what works. That's why I never open a fresh chat when something looks off. Sometimes the first try comes out right. That mostly depends on how well you know your own process. If you can describe it clearly, the AI can build it. Right now, I'm the only person who can use this app. The team needs logins, and so do the clients, but they shouldn't see the same things. So, I write the whole access plan into one prompt group by group. Claude creates the three groups and adds the client portal page. Then it sets who gets to see each page. The dashboard only shows up for admin. On the portal, a client gets their own campaigns waiting for review, plus an approve button. Then I preview the app as each group, starting with team member. I get the board and the tasks, but there's no dashboard in the menu. As a client, I only get one campaign card with an approve button, and the budget field isn't on it. Switching to admin brings back the full app, dashboard included. Hiding a menu item isn't enough for me. So, as the client, I also try to open another client's campaign directly, and the app won't show it. Then, I click approve on my own campaign and go back to the data tab. That campaign has moved to approved. Here, the user accounts and the data live in the same app. The board Claude road follows the same visibility rules as every other block on the page. So, I didn't have to build a separate permission system on top of it. Once a client approves, someone on my team has to get the launch ready. I don't want that to depend on somebody remembering it, so I ask Claude for a workflow. Claude builds the workflow, and I open it in Softr to look at what it made. It's a short chain on the workflow canvas. The trigger is the campaign stage changing to approved. The action creates a new task and assign it to the campaign owner. This automation works for my team, since the task lands on their side of the app. Small campaigns don't really need a separate launch task though, so I narrow it down from the chat. Claude edits the same workflow, and now there's a filter step sitting between the trigger and the action. It stops the run if the budget is under 10,000. >> [music] >> To test it, I approve a big campaign from the client account. A prepare launch task shows up in the tasks table assigned to the right owner. Then, I approve a small one, and no task appears. Everything so far ran inside the editor, so now I publish the app. I get a live link that I could send to a real client today, and I didn't have to set up hosting anywhere. If you want to use your own address later, Softr lets you connect a custom domain, too. Before I trust it, I open that link in a private window and sign up with a brand new client account. I want to see the app the way a client would on day one. The new account lands on the portal with just their own campaign. There's one more thing this MCP connection can do. Claude can also connect to the live app as a regular user. So, I connect it with that client's login, the same way I connected the builder at the start. Claude comes back with one campaign, the same card the client sees in the portal. Then, I ask it to approve that campaign, and the campaign moves to approved. The client account is allowed to approve, so Claude can do it, too. If I ask it about another client's campaign, it has nothing to show. For a busy client, that means they can check on their campaigns from a chat they already use. The same works on the agency side. If I connect Claude with a team account, I can ask which tasks are overdue this week. It answers from the board my team actually works in instead of a spreadsheet someone exported on Monday. The building all happened in Claude, but the data and the access rules ended up in an app my team can open and check. If someone in ops builds their next tool this way, IT gets something to review instead of something to rewrite. Most of the work went into the prompts, and they got better every time I added detail. If you want to build your own version, the link in the description gets you 200 AI credits in softer. And tell me in the comments what you'd build first, an internal tool for your team or a portal for your clients and partners. I'll see you [music] in the next one.
I am building my own dictation app because I talk to my computer way more than I type on it. I just got tired of paying a subscription every month for something that I do every single day. But I'm not starting from scratch. Gladia, the company that makes the speech engine I'm using, open source their own dictation app and it's called Gladia Flow. And I can hold down a shortcut and it dictates into whatever app I'm in. And whatever you say, it will paste it right here. So we'll fork it and then we will run it. And then we will make it our own and add more features to it. And the nicest thing about it is it's free. Check out the Gladia API for free to get €50 of credits. And the free credits give you about 70 to 75 hours of speech for free before you even have to pay a scent.
Another week in AI and a bunch of stuff came out, but honestly, most of it not that relevant. I personally am more interested in what AI can do for me, which often is helping with the boring stuff rather than another benchmark chart. I'm referring to Google's carbon model that beat many benchmarks, but nobody has access to it. Nobody could try it. Nobody knows if it's actually any good. So, this is my Friday video and I always talk about what's new in AI, what came out and I want to show you a bunch of features and connectors that I think you'll actually be using versus just talking about some numbers that, you know, maybe sometimes in the future could be useful. But honestly, at this point, model improvements, I don't care about as much. It's kind of hard to talk me out of using Opus 5.5 even though the benchmarks might look good. So, that being said, let's look at some of the new features starting with Chat GPT. So, they're doing a few things. They're iterating on their little dots, the always on assistance with a computer that proactively texts you. It's a really neat feature, although I did run into some troubles as I started to use it more and more. One of them is if you set up the rules, which are available under personalization. If you didn't know this, you can go there and then you have these custom rules that you can add. Even though you set up some custom rules, it doesn't always follow it. For example, I tried to outsource all of my meeting booking in Slack to my dot. Didn't always follow it. But again, what's new and useful here? Well, they added a feature where you can actually connect it to your phone and you can start using it for messages on your phone. So, we tested this across the office and we found that many people don't have it. It seems to start rolling out on the pro accounts and on my account I actually do have it, but it turns out that me connecting my international phone number has not worked. So, I don't have a US number yet and I don't want to connect this to somebody else's phone. But, you know, it's here. It's a new way to interact with it. fusing dots. This might come in handy. Additionally, I want to point out that they've been iterating on the call feature, and it's way better now. Hey, Eigor. Good to be back with you. What's on your mind today? Hey, what's up? Can you hear me? Yeah, loud and clear. Nice. Now you can just get things done as I talk to you, right? While on the go. Exactly. Yeah. Just talk through what you're So, it fires instantly. There's no dialup time. Remember in the beginning, it sometimes like was calling the dot for 40 seconds before it connected. and also the responses are snappier. I find this to be an amazing way to interact with your AI assistant, especially in the morning. Maybe you're driving, you just hit the phone call, or you're walking somewhere. I just wanted to point those things out to you along with some changes to chat GPT plans. So, remember when a month ago they essentially said that they're having the usage you get with your chat subscriptions? Well, they just backpaddled on that in a big way because they're actually doubling your usage on certain subscriptions. I mean, check this out. Tibo here from OpenAI who is becoming the communicator for these kind of things put it out there that if you're in the 100 200 or $500 plan you're actually going to be getting double the usage effective today instead of cutting them in half. So in other words the higher up plans are getting more value while in more and more ways the $20 plans are starting to feel limited. You don't have dots. You don't have this big usage limits. But to be fair with the value and the intelligence that AI can provide it's I mean $20 is an incredible steal. and we're probably entering into this era where if you want to use it for serious tasks, it'll probably have to be on the $100 plan. We'll keep an eye on that though. Additionally, there were rumors of them removing the chat work feature and Tibo confirmed that, hey, they're working on it, but it's not ready just yet. So, similar to Claude, so we'll probably see a change in chat where they remove this work functionality that many people find confusing like I'm doing work in chat already. When should I switch to work? The answer is if you're doing longer running tasks or creating things or doing bigger things or you need skills, but it should just route automatically and we'll get that eventually. All right, back to the simple wins that I was talking about. Another one is you can now connect iCloud apps like your calendar and your email account. This wasn't possible up until now. If you had a Mac and just use the native email address or the native calendar with your Apple email, you couldn't connect up. you had to either migrate to a Gmail account or do what many people did, which is auto forward all the emails from that inbox to another inbox that you could connect. Well, no more. You can go to customize plugins and in here iCloud mail, iCloud calendar available as of today. And at this point, I realize many people watching this might be like, okay, um, cool, good features, but I'm still looking for ways to use this AI. sort of always a lingering question for many people and even the people who are deep into this are always looking for more ways to use it. I found this post on a Grockbot subreddit from a small business owner that listed every single thing he's doing in his home life and then in his work. Look at that. This is a list of like 30 use cases. The simplest thing you could do is copy paste all of this into your AI. If it has sufficient context on you, it will be able to parse out and tell you what of this matters. But honestly, I just find reading this fascinating. Look, here's all the family use cases where it manages the school emails for his children so he never misses anything even within an attachment. Then you have the classic, you know, meal and sports use cases. But look at some of these specific ones. It writes every vendor purchase order in QuickBooks. Tags each line to the job, emails the vendor. For me, I do a similar thing where I reconcile the invoices in my email inbox with our company card that where it needs the invoices for every transaction. and just have it pull it from the email and put it into ramp. There's a plethora of things you could do, but most of them are specific to the business or the goals or pain points that you have. And I realize with AI education, there is an exploratory process where you just need to engulf yourself in this and look at how different people are using it to find how you're going to be using it. This post does a really great job at it and you can do most of this inside of your dot. Now, if you do want to use Grockbot, which she uses it in this case, let me just briefly intro you to that if you're not familiar. It's basically the chat GPT dot but split out into multiple chats but it's not chats it's separate bots meaning each bot has its own computer and they can communicate and work together. Really the framework of thinking about it as hiring AI employees is the best mental model for this. If I would want to hire a person to help me with my invoices well I could just add a bot and then work with this stuff. But what makes it really attractive is kind of the features around it that are available. And a bunch of new stuff came out over the last week. So, let me quickly tell you about it. If something here sounds interesting, you might want to add it to your Grockbot setup or even just try it on a free plan. So, similar to Chach, they added the ability to use your phone's apps. And first, I couldn't find it. I wanted to link to my Apple contacts, but then I realized I need to go into the mobile app. And if I'm in here, you can see you can connect the various Apple apps. Not all of them, but many of them. Can remote control your home, see your entire calendar, too. And kind of the core pattern with each one of these is always the same. It's you set something up with an idea of what you could do. You get inspired by a video like this and you know you kind of just set it up and it's okay. But then you start giving it feedback on the things that it didn't do cleanly and it improves. It adds rules to how you do things and over time it accumulates rules and maybe you even build certain skills to do specific things that you can then again train over time. So, it's always this loop of like you share something, it does it, and then you criticize the output, and the next time you do something, it's just better, and you just keep running that loop. All right. Another feature that you could use in your work with Grockots, they added pages. Similar to Chhat PT, they have these spaces, and inside the spaces, you can create pages. Ooh, they actually rolled out slides now. I just noticed this wasn't available like a week or two ago. They didn't even tweet about this. Look at that. So, they have their own copy of Google Slides right in here. Anyway, I was going to talk about these pages where if you go here, it's basically like a mix of a word doc and a notion page. It's a markdown file which is a text file that AI loves. And now Grockbot has the same thing. So, for example, I have this podcast tracker set up where regularly it scans for new podcasts that I've done from my calendar and then pulls together the links and the transcripts and then I can share that with the team. Well, I set that up inside of a bot that only looks for these podcasts and it has a page to track some of the recent ones. Oh, by the way, if you want to check out one of these, we'll put a link in the description. My conversation with Omar recently was really good. You can find it on his channel, The Passionate Few on YouTube. Arguably, the packaging is a bit aggressive for my taste, but it's a really good conversation about some of the timeless principles surrounding AI and where all of this leads. Two more things I wanted to share with you, and this goes for kind of all AIS. One of them is one password. They've been integrating into more and more apps. And with some of the recent security breaches around PHP, if you heard about that last week, there was a whole security scare. Managing your passwords properly has become more and more important. And I'm just realizing I'm starting to sound like a sponsored segment, but I just genuinely think you should have a password manager. Whether it's one password or another one, it's just one password is the one I've been using since I don't know, years. And it's just one of those digital things that truly is a game changer. I mean I have thousands of login here and it neatly connects to many of these tools. So for example, what I can do is I can set up a separate password vault just for my AI agent. You can see this one shared with Grockbot. But then my Grockbot has access to those login securely stored in one password. And the security of that is actually superior to you even having your own passwords because each password is unique and the whole thing is encrypted. So even if it's hacked, they cannot get the information out of it unless they have your master password. So you really need to remember one password. And if you do that and if that's a complicated enough password, this thing is more secure than reusing some semiscure password across multiple services. Again, just a recommendation, but this thing now works inside of Grockbot. And let's see, does chat have one password connector? I don't really use chat that way, but okay, I guess they only have developer tools, so that's different. But yeah, independent of AI usage, I just highly recommend for your digital life to have a password manager. And if you do use Grockbot, you could connect it in there and give it a separate vault with some login to certain sites which you can then access on its web browser because remember each bot has its own computer, a browser on it. If you have one password connected, the login to those tools too. Okay, final tip here that I discovered this week and you know this is unfortunately only for iPhone users and you know versions of this have been around for a while but mapping this action button if you have it on the iPhones from the past 2 years. It's this one about the volume buttons. You can make it do specific things. And while in the past I've often had it on starting a new chat and chat GPT, now I remapped it to access my Grock bots because if you have a look here in the action button settings, there's a option where you can put it on start voice chat. So whatever I'm doing, I could just immediately talk to my orchestrator bot, my main bot, and with the press of a button get things done like order food or reply to emails or get my daily digest. It's very similar to talking to your chat GPT dot, but the dot doesn't have the shortcut yet. Rockbot is a more mature product at this point in time. Oh, and actually, I'm just getting notice that there's a side button on Samsung. Can I see the side button? So, you know, they have it, too. And apparently, it's programmable. So, I guess you could do the same thing on a Samsung, too. Nice. If you watch this channel regularly, my feeling is you would probably get a lot of value from it if you're not using it already. And again all of the context is transferable which talking about that I actually thought of one more thing I want to show you. If you go to settings general and then here under import you can say start import and they have this prompt that allows you to run that inside of the AI of your choice like chat GPT or cloud and you could just bring all the memories and your context you have tied to the account into Grockbot. I'm telling you that because you could run the same thing within Grockbots to then migrate back to your dot or claude do it or whatever you want to use in the future. Like the context is transferable. All right, that's the interesting stuff I found this week. I hope this was interesting or useful. And I want to end with a recommendation. I created a little video on this reverse engineer anything repo where you can rebuild any app literally in one prompt. It's insane. I'll link that video on the screen here. Go check it out. And other than that, my name is Eigor and I hope you have a wonderful week. News you [music] can use. Breaking down the stories you actually care about. [music] AI news you can use. Eagle's [music] here to help you figure it
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=Bfo_GZWfRVo! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
If you keep playing on this knee, you may not walk properly again. Don't tie your life to the street. I want to take care of you. I'll take care of myself. Management wants to replace you. We both know the team needs me. Hi, I'm Adil. You guys often ask me how to turn your own characters and clothes into a commercial. It's three simple steps: assets, scenes, and editing. This story jumps between three heroes in one ad. So I built it on my four-beat script: setup, problem, choice, and resolution. Every scene I generate gets one of those jobs. That decides how I split the story into blocks, and later, how I cut them together. Take the goalkeeper. His fall during training is the problem, and the save in the final match is the resolution. So they're two separate blocks. And in the edit, the save sits right next to the other two heroes' resolutions. Every prompt and asset, including the scenes I skip, is in the blog post below. And at the end, I'll show you what this actually cost, including the takes that didn't work. Now, before I generate anything, I turn the story into a shot list. That tells me which characters, clothes, and locations I actually need. For every shot, I describe it to Claude, Claude writes the prompt, and I generate it in Cinema Studio with my assets. Everything lives in one project, a folder for each hero. Once an asset is done, I save it in Higgsfield as an element with an @ name, and I give Claude the same image with the same name. That way, when I paste Claude's prompt back into Higgsfield, the right element attaches itself. Every character gets the same clean sheet. A gray background, body views with the head cropped out, and a big close-up of the face. That's what keeps him looking the same in every scene. Goalkeeper first. I'm asking Claude for an ordinary working guy: dark curls, a warm face, real skin that holds up in a close-up. And he's a goalkeeper, so even in a regular shirt, he has to look like an athlete. Now I'm taking Claude's prompt into Soul Cinema and generating four options at once. For people, that's my go-to image model. It keeps real skin textures, so faces don't turn into plastic AI slop. Now, the first one smiles a little too much for me. The second one feels too stern. He looks like he's about to give me a parking ticket. The third is closer, but the smile still feels posed. But the fourth one, I like this guy. He looks like a real person, not a model. Let's go with him. And let me know in the comments which one you would go with. Now, before I dress him, we need a logo. Simple enough to spot on a shirt or a shoe, nothing fancy. So I'm switching to GPT Image. It's the best at clean, flat graphics. I'm generating four options. The loops feel interesting, but the double outline on someone's chest is just noise. But the C, yeah, it's one wide open shape. So even small on a shirt, you can tell what it is. I'm saving it as an asset so the model doesn't invent a new logo for every outfit. Now for the kit. We'll do it once in detail, then it's the exact same move for everyone else. Claude gets his character sheet and the outfit sheet and writes the edit prompt. I run it in Seedream inside Cinema Studio. Now, Seedream is great at edits. He changes the clothes and keeps his face exactly the same. And there he is. Same face, but now in the green kit. I'm checking the crest, panels, number and logo against the outfit sheet. Now, a quick note: these clothes are generated. For your own product, just shoot it front and back. That's how the model gets the cut, the print, and the logo exactly right. Alright, he also needs a version with a knee brace because he cuts it off later. That means a clean prop sheet of the brace, then the kit with it on. And here's the brace version. Same kit and face, with a brace on the right knee. I'll bring in everyone else's looks when a scene needs them, and you can grab all those prompts from the link in the description. The assets are ready, now the scenes. And to write those prompts, Claude needs to know how I direct, so I give it my own directing skill. A set of instructions you load once, and Claude follows them every time. This one's called Directing Workflow. It turns a scene into a shot plan. It's linked below. Loading it takes a few seconds: Customize, Skills, upload the file, and switch it on. Then I just mention it in my request. From now on, I work like this. I split every story into blocks. One block is one prompt, and one prompt can hold several scenes. I run it a few times, pull the best moments from different takes, and fix whatever breaks. First up: an injured goalkeeper with one last big match, and a knee that might not hold. I build his story from five blocks. The clinic, his home, and the street are the setup, training is the problem, the tunnel and changing room are his choice, and the match is the resolution. Block one is the setup that establishes the main character and his reality. Two scenes in one prompt. At the clinic, the doctor tells him how bad his knee is, and at home, he pulls the brace on. So I'm bringing in all the assets for the scene: location assets for the clinic and the flat, his character sheets and his everyday clothes, his home outfit, and of course the brace. Then I describe what I want in the first scene to Claude. That's all I type. The skill does the rest. It breaks the scene into shots, decides how the camera moves, and even tells the model how he should look, lost and empty. Then I paste the prompt into Cinema Studio and pick Seedance 2.5. I'm generating two takes. Okay, in both, he's slumped on the table, practically lying down. I wanted him standing against it, and the prompt says standing, but the shots are so close that the model never sees his whole body. The home half is fine in both takes, the acting works, so I'm only fixing the clinic. For that, I'm adding a wide shot at the start, so the model sees him standing at the table before we move in. That makes four shots. And there we go. He's upright in both, and he stays that way. Now I'm taking the best of the start and the end, the clinic from one take, his home from another one. Let's watch the clinic and his home cut together. The damage to the ligament is serious. If you keep playing on this knee, you may not walk properly again. Before he gets back to training, I wanted one more scene to make his setup feel real. He walks past a mural of himself under a bridge, just as someone sprays Retire, old man across it. That's block two, one scene. The mural is part of the location. I'm starting with an empty street under a bridge, then a Seedream edit paints the mural onto the pier and hangs an old lamp right above it, so the light on the mural matches the street. In the request, the timing does the work. A car passes, the mural shows up behind it, and the vandal is just finishing up the tag. This one worked on the first try. I'm taking the walk and the reveal from one take, and his face from the other. Next comes the problem, the obstacle in his way. He tries to come back to training, but the knee says no. First, though, he needs somewhere to train. I build a field at a three-quarter angle, so the model gets the depth of the pitch and a clear line to run. For lighting, it's blue hour. Cool, but bright enough to follow the action. Claude's prompt goes into Cinema Studio, Image, Soul Cinema, and before I hit generate, I pick the color. Cinema Studio has built-in HEX color presets, and you choose them up front. First test: Twilight Fable. Let's make a batch of four images. Way too purple. In combination with my other shots, it would look like it's from another film, and I'm losing half the pitch in the shadows. I want it dark, but I still need to read the space. Same prompt, but I'm switching the HEX preset to The Morning After Rain and running four more. Oh yeah, The Morning After Rain works: colder sky, no purple cast, more neutral shadows. This is the one I'm keeping. The field and the look are locked. Block 3: for this beat I want a training sequence, a warm-up, then the drill, both in one prompt. He sprints to the cone, turns, and his knee gives out. That turn is the part I need to nail. Two training sheets, and the brace stays on his right knee in both. I start with the warm-up. First try, the simple way: his character sheet, the field, and some words. Mmm, nope. Okay, he jogs and that's it. No rhythm, no knee lift. It looks like a guy who's late for the bus. This is where I got stuck. I burned a bunch of credits trying to describe this warm-up in words, and it never moved the way I wanted. So I made it easy: I filmed the movement myself. Watch the right knee. It comes up high on every other stride, and that's much easier to show in a clip than explain with a natural warm-up. And here's how the clip gets in. In Claude, I'm attaching the clip to my message, together with his character sheet and the field, and I'm telling Claude what this clip is for. Only the movement, the rhythm, and the knee coming up on every other stride. Claude writes me a full prompt that takes the motion from the clip and everything else from my assets. Then in Cinema Studio, I'm picking Seedance 2.5, adding the clip next to the images as a video asset, and pasting in the prompt. There. So the clip is called @Video 1. Here's the warm-up I went with, and there's the knee lift I couldn't get with words. Now to the second part of the block, the fall. I want it to read as the knee giving out, not a trip. So I spell out the order: he sprints to the cone, plants his right foot, starts to turn, and the knee gives way. He drops onto his left side, rolls onto his back, and grabs his right knee. Then, away from the bench, the coach just stands there with his arms folded and watches. Two more things in this prompt. For the acting, the Directing Workflow skill scripts his face beat by beat: lips pressed, jaw clenched, breathing hard through his nose. And for the sweaty look, just one line: sweat at the hairline. This one's 26 seconds and nine shots, so I'm not going to render it in full quality straight away. I'm making four 480p drafts in Seedance: cheaper, faster, and they're good enough to judge the movement and the framing. Okay, the knee lift from my clip made it in, and the turn leads into the fall. So the training scene comes from three takes: the sprint and the fall from one, the turn at the cone from another, and the coach from a third. With drafts, you only finalize the takes that go into the edit in 1080p. The warm-up comes from another take, so I stitch it together with the drill in the edit. Take a look. Now, his knee gives way right as he turns. My own clip gave me the movement here, so if you need a specific move, film it yourself first. It's cheaper than running take after take and hoping the model gets it. On to the choice: block 4, two scenes again. In the tunnel, the coach tells him management wants to replace him. That outside pressure pushes him. Then in the changing room, he cuts off the knee brace. That's his choice. For the tunnel conversation, they need space from the walls, so he can finish his line and walk out. In the changing room, it's a close-up. He cuts the brace and the camera pushes in on his face. Both scenes go into one request. Two takes this time. He's hugging the wall, but I asked for both men in the middle. The changing room already works: he cuts the brace and pulls it off. That leaves the tunnel. I doubt more words would fix it, so I'll just draw it to save some time and credits. Don't judge my drawing. Honestly, this is my best work. So blue is the keeper, yellow is the coach. That just tells the model where they stand and which way they face. Now, with the prompt from Claude and the added sketch, I run it again. And this time, both men are right in the middle of the tunnel. So now I can mix the shots from different batches: the tunnel from this one, him cutting the brace from the second, and the final look from the last one. Here's the mix. Management wants to replace you. We both know the team needs me. Okay, on to the final part. The resolution comes in block 5: his comeback after the fall. The walkout and the match share one prompt. We start behind him, walking out of the tunnel, Pedrinho 1 on his back. Then the play: the shot, the save, the counterattack, and his reaction. First, the bricks. Anything recurring more than once gets its own asset: his team, the opponents, the stadium, the ball, because with pure text-to-video, the kits and the players would change every take. So I give them their own sheets. The crowd is already in the stadium asset so the model doesn't invent people from scratch. And he's in his match kit so no knee brace this time. Here's the request. It's quite a detailed one; that's a lot of story for 22 seconds. Let's see what we got. The walkout is spot on: the silhouette, Pedrinho 1 on the back, stewards at the tunnel. The kits hold every take, too: blue for home, and red and black for the opponents. Now the save. He flies top corner, hugs the ball, grass sprays... I mean, I could watch that on a loop. So the match comes from four takes of this prompt. The walkout from one, the face, the shot and the save from another, and the long kick and his reaction from two more. I end on his reaction. He looks up, smiles and raises one finger. His smile lets all that tension go. One hero down, two to go. Same four-beat script for both of them, so let's pick up the pace. Our next hero is a ballet dancer who wants to do hip-hop, and her family doesn't get it. Her story takes five blocks with the same four steps. It starts with the match cut and the dinner. The locker room is her problem, practice and the argument with her dad are her choice, and it all ends at the dance battle. From here on out, I'm not walking through every asset step by step. You know the workflow; I'm gonna stick to what's new in each scene. Her setup is block one, and it opens on a match cut. She closes her eyes in the ballet studio, then opens them on the bridge. Her face stays in the same spot in the frame, and only her clothes and surroundings change. Then we cut to dinner. She sits next to her dad, he keeps eating, her mom glances at her once, and nobody says a word. I want you to feel how uncomfortable she is. The assets go in as usual: her three looks, both parents, and all three locations. I asked Claude for both scenes together. Here's what I got. The match cut works here, but the dinner scene, it's too wide. I'm watching the whole table here, not her. And I also want to put both parents on the other side of the table to highlight the emotional distance between them. So from this take, I'm keeping the match cut. So it's just the push-in and the mother's position I'm changing. Start on her empty chair, push in as she enters, and land right on her face, keeping her dad at the edge. Perfect. Once the camera knows where to go, I get the exact cut I want. I'm keeping two takes, one for the match cut and one for the dinner. In the edit, her setup looks just like this. As for her problem, her parents are already against her dancing, but I wanted to push it further. So block 2 is a single scene: the locker room, where she reads the negative comments under her dancing videos. The trick is keeping her dance clip on the phone screen with the locker room around her. Here's the move. First, I generated the vertical clip of her dancing, then I uploaded it to Claude, along with the phone asset and everything else we need for the scene, and write the following request. Here's the first take. The dance stays inside the display and the comments are readable. That's what she reacts to. I keep the close-up on the screen, cut to her face, and trim that long pullback at the end. I run it a few more times and build the locker room scene from five passes, one take from each. The girl leaving, the phone, the comments, her lying down, and the wide shot from the other one. Put together, it plays like this. Her choice comes in two halves. In the first, she keeps practicing despite the comments. Block 3 covers that: the bridge and the studio, an outfit for each, in one prompt. And now, she has to really dance. First test: black streetwear on the bridge, dancing toward the camera. I'm keeping the request simple. She moves like a robot, and then she just starts jogging. I asked to mix walking with dancing. What I wanted is one move flowing into the next: shift the weight, turn, drop low, and keep going. On that first try, I didn't have a dance skill, so I fed Claude videos of my friends dancing to break down the footwork, weight shifts, and arms, and turned that into a Dance Direction skill. It loads the same way as the first one. Now, instead of just typing make her dance, I can prompt actual steps: two side steps, a quarter turn, a low drop. Those are easy to check. Eight shots in all, four for each location. With the skill, I tried twice. Here they are side by side, without the skill and with it. Now that's dancing. On the bridge and in the studio, she moves through the steps instead of breaking into a run. That gives me the practice. She keeps dancing no matter what they wrote. Her story needs two more parts. She stands up to her dad and then steps into the dance battle. The argument comes first. One scene and it finishes the choice part. Her dad follows her down the hall and she answers firmly before closing the door. So I'm describing that to Claude and this is what comes back. Don't tie your life to the street. I'll take care of myself. It's clear who's speaking, she lands her answer, and the door shuts after it. I just trim the walk and keep the take. I'm leaving out a few scenes, like her grabbing the sneakers. You'll find every one of them, with the prompts and assets, at the link below. The battle is her resolution, block 5, with three scenes: her footwork, the celebration, and her dad watching from the crowd. She needs to own the center of the circle, with the judge and opponent framing the edge. For the first take, I even tried having the opponent step in while his crew pulls him back. I'm attaching all the assets for the scene: the venue that already brings the crowd, her opponent's and the judge's character sheets, and her final look. And obviously, my request to Claude. Eyes on the opponent? Bro, it's not even his turn. Actually, I don't think I like him coming into the circle. My eyes go to him when I want to be watching her feet. That interruption was on me. I thought it would add some tension, so I'm taking his entrance out. He stays behind the line, and I keep the celebration and her dad's reaction. One change to Claude: the center stays clear, her feet are visible, and he never crosses the line. I run the new version twice. The first low shot comes from one take, and the other two shots come from the second one. The crowd lifts her, but her father is there too. His reaction wraps up the family story. Without it, the argument would just end at the closed door. Last up, the schoolboy. A quiet kid gets called a nerd. Again, it's five blocks, and by now you know the four steps: school and home, the note, the jersey walk, and the tournament. His setup shows two sides of his life, so that's two blocks, one scene each. Block one is school, where he keeps to himself. Block two is home: he's at his desk, and his mom is at the door. Same face and glasses, two outfits: the navy hoodie for school, the grey sweatshirt for home. His room gets the same trick as the field: a three-quarter angle from the corner. In the classroom, I lock that angle on him in the prompt instead. For the classroom, I run two takes. Take two is the one: the reverse comes in low and close, and he's alone in the frame. On to block two, his home. He's bored at his desk, staring out the window, and that's when his mom comes to the door. I run two takes and keep the second one. The window light falls warm on his face, and his mom is right there in the doorway behind him. And I'm putting them back to back. Next, something has to hurt him. The problem is block 3. One scene in his locker and it all hangs on one prop. Someone's left him a note saying NERD. One word, and you know how the other kids see him. Like any prop, the note has its own asset. Text is where the model slips, so the word always starts from this one image. Since it's just an image, I run eight takes and keep the last one. The word NERD comes out clean, and the caricature has his glasses and a gamepad. The locker scene took two takes. The second one's better. You can already see the note on his locker as he walks up, so the close-up lands. Here's the finished scene. I'm also skipping two of his scenes, the gym and his first gaming room. Block four is his choice. Again, one scene, how he answers that note. He walks back through the school in his esports jersey. Before, NERD was a label they stuck on him. Now, Nerdy23 is his gamer tag on his back, and he picked it. I want the scene to start behind him, so we can read Nerdy23. Then the camera tracks in front of him as he walks and we end on the other students noticing him. Two takes of this one. The walk works, but check his back: Nerdy23 turns into all caps. So now the jersey is yelling. The prompt's already specifying the name correctly, so I'm not going to rewrite it. I'll just check each take against the outfit sheet. Okay, this one keeps the exact lettering. Here's the walk with that jersey. We read it, then cut to his confident face. In block five, the resolution is just one scene: the tournament. It's a bigger room with stage lights and his team in matching jerseys. I bring in the stage and one sheet with all four teammates, because again, without it, Seedance kept generating random people in the background. Here's one of my first takes. Without that sheet, he looks right, but the guys around him are random kids in black hoodies, not his team. I'm going to type in my request, then I take that prompt into Cinema Studio, two batches, and hit generate. Okay, from the first take, I'll get the wide shot, his face, and the arena. And from the second one, I'm taking the headset, the hands, and the smile. Here's the result. There's that smile. At school, he was alone. Now, he's part of a team. And that's it for the schoolboy. All three stories are done, so step three is cutting them into one film. This is where that four-beat script from the start really helps. I can cut between the heroes when they hit the same kind of moment in the story. That order is crucial. If a story doesn't work on its own, intercutting won't save it. It just hides the flaws for a few seconds. When you do mix them, every single cut needs a clear reason. Match the beats. A setup matches a setup: in the first 17 seconds, we meet all three. The ballet studio, the doctor, the classroom. A problem matches a problem: the goalkeeper falls, cut to the ball hitting the boy in the gym. A choice matches a choice: he cuts off the brace, the boy puts on his jersey, she picks up the sneakers. Three decisions back to back. A resolution matches a resolution: the stadium, the tournament, the battle, and at the peak, his fingers on the keyboard cut to her feet on the floor. Here's one from the resolution: the ball in the goalkeeper's gloves, then the schoolboy's hand on the mouse at the tournament. I also use the details I generated: the knee brace, the jersey, the shoes. I cut on those close-ups, so the clothes and gear stay part of it. All right, before we finish, let's watch the full ad one more time. If you keep playing on this knee, you may not walk properly again. Don't tie your life to the street. I want to take care of you. I'll take care of myself. Management wants to replace you. We both know the team needs me. And here's the most asked question: how much did it cost? This project cost about 8,500 credits, counting the takes I didn't use. That's a bit pricey, but for a commercial like this, it's fine. A real shoot with three casts, a stadium, and a dance battle would cost hundreds of thousands of dollars. So if you make a commercial like this for a brand, you can put the credit costs into the production budget. The whole project took 72 hours. On a real set, this would take months from start to finish. What I take into the next one is approaching the story as a real director, so when you generate something, you already have a vision of how you're going to make the story work. If you get stuck, jump into our Discord server to get help and share what you're making. This whole project started from a viewer comment, so drop your idea for the next commercial below. If one hits a thousand likes, I'll bring it to life and break down the entire process. And as always, if you found this video helpful, hit the like button, subscribe, and I'll see you guys in the next one.
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=9Ky5QTpVjAw! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
So frequently on this show, you will hear me ramble off a bunch of numbers that are supposed to tell you something significant about a new model that has just come out. And more often than not, those benchmark numbers are in some ways related to artificial analysis, the company that has become the leading provider of the overarching benchmark that people use to judge AI models. And as AI models become more significant, the benchmarks that tell us how they perform relative to one another consequently get more important. Today we are looking at a major update to artificial analysis's intelligence index, as well as a new leader for the benchmark, which comes, believe it or not, from Google. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors KPMG, C data, Granola, and Blitsy. To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. On aidailybrief.ai, you can also find links to everything else going on in the ecosystem, free trainings, paid trainings like our new Champions program at Superintelligent, or of course the new How We Got Here podcast, and the first episode out this week called How Tech Lost America. Again, you can find all of those links from aidailybrief.ai. Welcome back to the AI Daily Brief Headlines edition, all the daily AI news you need in around 5 minutes. We kick off today with a story that could have easily been the main, and honestly probably would have been if more people had had their chance to get their hands on this model. Google has unveiled their next frontier model, Gemini 4 Carbon, just 2 weeks after the release of Gemini 4 Argon. As I mentioned on the show a number of times since Argon arrived, having a new base model means Google can iterate quickly with reinforcement learning rather than needing to complete a lengthy pre-training run to put out a new model. And this is definitely a quick pace, one of the shortest gaps we've seen between versions of ultra large models. CEO Sundar Pichai even suggested the pace would increase from here, commenting on Wednesday night's earnings call, "This rapid progress underscores our incredible model momentum. We are at the point where AI is helping us increase our training and launch velocity. I would expect the next 12 months to see continued acceleration because of RSI. So, Google is iterating at a record pace, but I'm sure you are sitting there asking yourself the same question that everyone is, which is is the model any good? In particular, is it a significant improvement on the underwhelming release of Argon? Certainly, when it comes to the benchmarks, the answer is a definitive yes. In coding, Carbon gained eight points on Terminal Bench 4.0 and six points on Frontier Suite. Cybersecurity saw a 12-point increase on Exploit Gym and a three-point increase on CWE Bench. For computer use, Carbon picked up seven points on OS World and eight points on Agent's Last Exam. On six of those eight key benchmarks, Gemini 4 Carbon is now the top of the leaderboard, ahead of Astra and Opus 5.5. Gemini 4 Carbon is now the frontrunner on the new Artificial Analysis Index V5, which we will be discussing in the main part of the episode. It scored 59 points, representing a nine-point jump on the new scale, and placing it ahead of Opus 5.5 in first place. On the AA Coding Index, Carbon landed in second place, tied with Sonnet 5.5 and three points behind Opus 5.5. Now, these are huge jumps for Google. When Argon was released two weeks ago, some felt that Google had missed their window. The model might have been impressive a month or two earlier, with benchmark scores topping Fable, but the release of Opus 5.5 and GPT-61 Soul had pushed the frontier sufficiently that it left Google in the second tier alongside Meta. This release, if people's experience with it matches the benchmarks, puts Google genuinely back in the race for the frontier. Cost is basically unchanged compared to Argon, with Artificial Analysis measuring cost per task at $7.76. For reference, that's 23% cheaper than Opus 5.5 and 20% more expensive than Astra, so right in the middle of the road for frontier models. Now, here we come to the less great part. We don't have many first-hand testimonials, as Google unveiled the model before it's ready for general release. They're following the administration's pre-release model testing guidelines, and said they're taking what they call a deliberate approach to the rollout. The model will be made available to a subset of cyber defenders beginning this week with access extended to trusted partners in their cyber defense program. As for the rest of us, Google wrote, "We're continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product services as soon as possible." One of the few groups talking about their experience with the new model was Arena, who had access for testing purposes. They put out a video review full of render and game design tests that were generally slammed for not being as pretty as Opus 5.5. Now, as I have said before, I am not particularly convinced that 3D design is the right yardstick for LLMs, no matter how much social media users like to focus on that, but certainly at least when it comes to that particular use case, Carbon is not at the same level as its rivals. Arena's Peter Gostev did give a few more details, however, posting, "I have used Carbon a bit before the release, and I can tell you it feels a lot better than Argon. This is a real deal from Google this time. I hope they keep up with the rapid pace of releases from now on." And that does seem to be the plan. Responding to Teo tax's commenting that Fable 5.5 isn't here yet, though, DeepMind researcher Zachary Nado wrote, "Gemini redacted isn't here yet, though." Chubby on X commented, "Okay, now things get crazy. They have another unreleased Gemini model that's even better than Carbon, and Carbon has already reached number one in the AI intelligence index today." Gemini post-training researcher Shao Zhang added that Google AI is out of their funk and operating like a well-oiled machine. Less than a year ago, he wrote, "Launching Gemini 3 Pro was a big bang event where the whole team celebrated number one benchmark numbers. Today, Gemini 4 Carbon hits number one on the AI index, yet the feeling inside is distinctly different. We've been dogfooding continuously for weeks, and we've almost adapted to the new intelligence, taking ambitious code rewrites for granted, leaving loops overnight that launched dozens of experiments to repo, root cause, and fix mysterious bugs, and using swarms to inspect and fix RL data. Launches just feel like periodic checkpoints now." And despite taking 10 months off the early adopter AI conversation, you can see that Google still has lots of people's benefit of the doubt. Crusader Ventures Hesam writes, "Why is Google suddenly dropping models like they invented transformers, own half the internet's data, and sit on a notion of TPUs. Oh, wait. Latent Spaces Twix writes, "Gemini number one on AAV5. This is making me open up anti-gravity for the first time in a long time. If Google actually pulls this off, it is officially comeback season with Meta, Grok, and now GDM being cool again." For my part, I think we will suspend judgment until we get our hands on the actual model, but it looks promising, and that in and of itself is exciting. Now, speaking of Google, the other story was in their and Microsoft's earnings reports. Both companies delivered another incredibly strong quarter in cloud growth, which is where, of course, they contain their AI revenue. Microsoft said their Azure division grew at 48% over the past year to reach 33 billion. This is the first time Microsoft has broken out Azure revenue as a separate line item, which seems like a very good decision with growth this strong. Azure drove overall revenue growth of 19% to reach 92.3 billion for the quarter. Each of those figures was a slight beat over analyst estimates. Microsoft also disclosed that OpenAI contributed 10 billion to Azure revenue, around a third of the total business. Microsoft's stock was up as much as 4.8% to reach a new all-time high in after-hours markets. Google's cloud division, meanwhile, doubled revenue compared to year-ago figures, reaching 30.7 billion for the quarter, and meaning they've almost entirely closed the gap with Microsoft. This was well above Wall Street's high expectations. However, as David Wagner of Apttus Capital Advisors put it, "The problem is everything else." Search revenue growth came in at 12%, a massive slowdown from 17% in the previous quarter, and 19% the quarter before that. YouTube revenue also added a third continuous growth slowdown to reach 9.2%. Even TPU sales was a bit of a question mark, with executives dodging a question about profit margins for the chips during the earnings call. Overall revenue growth still came in at 25%, roughly in line with expectations, but to some analysts, it looks like Google might have reached the moment where they're forced to cannibalize their own search business because it's no longer a reliable growth driver. In fact, CEO Sundar Pichai acknowledged this dynamic during the earnings call, commenting, "It's worth noting that the strong growth in AI mode queries had an impact because of the lower monetization we're seeing in AI mode. In addition, Google missed on bottom-line earnings largely due to a markdown on their SpaceX stock. Tech analyst Dan Ives, ever the optimist, tried to strike a positive tone writing, "This is the cloud print we have been waiting for, and in our view, it matters more than a lighter search number the bears will lean on tonight." And lean on it they did, sending the stock plummeting as much as 10% in after-hours markets. Much of the coverage focused on the cloud race between the two companies, which are now neck and neck, but the pace of the data center buildout might be even more consequential. Both companies held CapEx forecasts steady for this year and next, but actual CapEx spend for the quarter came in under expectations. Google, by way of example, is guiding a 50% increase in CapEx next year to 300 billion, but they're currently undershooting this year's numbers. Oracle and Meta are in a similar position. So, it looks like the AI buildout may have reached peak velocity. That's not at all surprising given physical constraints around components, power, and labor, let alone community support running on empty. Now, we are quite early in this trend, and Google said that they would have more in-depth on CapEx during the next quarterly earnings. But what makes this interesting is that many AI bears suggested that a CapEx slowdown would be the catalyst that pricks the AI bubble. What we're seeing, however, doesn't exactly line up with that theory. The logic was that demand would falter and force the hyperscalers to slow down. Instead, it looks like the hyperscalers, if they had their druthers, would like to continue accelerating their data center buildout, but are running into more fundamental and physical constraints. Tech analyst Luke Longo thought the slowdown was more about finance, but are running into more fundamental and physical constraints. Lastly today, more antitrust scrutiny for Nvidia as the DOJ queries their acquisition of Hugging Face. Reuters reports that the Department of Justice has made inquiries about how the acquisition would impact competition in the AI industry. In particular, the DOJ has asked whether ownership of Hugging Face would allow Nvidia to control a key distribution point for AI models. When they made the $13 billion acquisition in September, Nvidia maintained that Hugging Face would remain free to offer a platform and partner with a wide range of model developers, chip manufacturers, and cloud providers. They said the deal was expected to close in the first half of 2027 pending regulatory approval. These inquiries form part of what is now revealed to be a multi-year antitrust investigation into Nvidia. In September, the New York Times reported that the DOJ had opened an investigation in Nvidia's licensing and talent acquisition deal with chip-making startup Groq. That deal was structured as an acquire deal, which has been a common practice in the AI industry dating back to Google's 2024 deal with Character AI. Because these deals are not technically acquisitions, they don't trigger automatic antitrust reviews and can therefore theoretically close much faster. The assumption had been that the DOJ was using an investigation into Nvidia as a way to crack down on this practice more broadly. However, sources told Reuters this week that the investigation is much broader and seems to predate the Groq deal. Acquisitions of data center optimization companies Scale AI MD and Run AI are also within scope for their impact on the competitive landscape in AI infrastructure. And both of those deals were traditional acquisitions, not acquire deals. Furthermore, sources said that recent questioning from the DOJ covered Nvidia's venture investments, sales practices, and whether the chip designer has exerted pressure on its customers. In other words, rather than just a narrow investigation about circumventing antitrust rules, it now seems the DOJ is embarking on a full top-to-bottom antitrust investigation of Nvidia. Obviously, no charges have been filed at this stage and the investigation could conclude with no action, so there's no implication that Nvidia has done anything improper. Still, compared to what most people thought this was, it appears to be a much more concerning investigation for the dominant chip maker. Something to keep an eye on as we go forward for sure, but for now, that is going to do it for today's AI Daily Brief. Next up, the main episode. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers. This episode is brought to you by CData. Token spend is climbing, security is holding agents back, and companies rolling out AI initiatives might not know how it's truly being used. The root cause, AI is cut off from the systems and context your business runs on. That's why CData built Connect AI Gateway, one control point between your AI and your systems. Connect AI Gateway is not just another static model router. Prompts get intelligently routed to the right model with exactly the tools and data it needs, governed down to the record. The result, accurate answers at a fraction of the token cost with a context engine that learns your business with every interaction. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more. Give it a try and start your free trial at cdata.ai. That's C as in connectdata.ai. You're in back-to-back meetings all day. You're trying to stay present, but you're also worried you'll forget the decision, the action item, the important next step. This is a scenario that we absolutely live here at the AI Daily Brief. Given how fast things move and change and how quickly we're evolving everything that we do. That's where Granola comes in. Granola is an AI-powered notepad for meetings. You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends. No bots joining your calls, no distractions, just a clean notepad that helps you focus. During or after the call, you can chat with your notes. Ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes, which are pre-made prompts. Once you try it on a first meeting, it is very hard to go without. Head to granola.ai/aidaily and use code AI daily to get 100% off for 3 months for new users. That's granola.ai/aidaily and code AI daily to get 100% off for new users for 3 months. Let's use deep code-based understanding unlocks the thing every roadmap owner cares about, shipping new features. Here's the truth about building inside a massive enterprise codebase. Writing code was never the bottleneck. Context is. Which system does this touch? Which contracts can't break? Which standards apply? Blitzcy already knows because it reverse engineered your entire codebase into a dynamic knowledge graph before future work began. With that complete picture, Blitzcy builds features end to end. Architecture, APIs, UI, and tests all validated against your existing systems. One Blitzcy customer built an AI native application from scratch with 100% autonomous completion, saving over 2700 engineering hours. Features that respect your codebase instead of fighting it. Stop letting your backlog grow faster than your team. Accelerate your roadmap at blitzcy.com. That's b l i t z y dot com. Welcome back to the AI Daily Brief. Today we are talking about benchmarks. Specifically, a new set of benchmarks and how they're shaking up our sense of the AI field. Now, the eagle-eared among you may wonder exactly what I think about benchmarks, given that on the one hand, with every new model release, I implore you to only give them so much attention, and yet at the same time still rattle off a whole bunch of them as though they matter. In short, I think that benchmarks are inherently limited, but also important. Without them, we don't have good tools other than strictly vibes to understand the difference between different models, and yet in almost no circumstances do I think we should take the performance of one model on any one benchmark or even in the aggregate as the gospel. Done well, they are a useful guide to helping us understand where different models fit into our overall stack, but as we will see today, there is more and more effort to addressing some of the inherent limitations of benchmarks and trying to make them better. Whenever there is a new model, you'll hear me talk about the scores that it gets on the independent test administered by Artificial Analysis. Well, that company has just overhauled the way that they test model, giving us a brand new yardstick for frontier AI with version five of their intelligence index. The index takes around a dozen key benchmarks and distills them into a single score. And while the index is now entrenched in how we understand the AI race, leaning so hard on that single score is a relatively new phenomenon. Prior to the era of agentic coding, when LLMs were in many cases still glorified Google or writing assistants, vibes reigned supreme. Ranking on the chatbot arena leaderboard was the pinnacle test of a new model because it told developers what users wanted. But as research and coding became more important use cases, we got more objective benchmarks giving a clearer indication of how each model would perform on real work. These evaluations have evolved over time. First by adding harder questions as benchmarks became saturated. In other words, as all the models crowded the very top score of a benchmark, the creators of said benchmark would go add in a bunch of newer harder questions creating an iterative version that would hopefully provide more granularity on the difference between different models until that model got saturated as well. As agentic computer use became a fundamental part of model performance, the evaluations got an even more complete rethink. Through each of these changes, Artificial Analysis also evolved switching out benchmarks and rebalancing their weightings as different model abilities became comparatively more important. This hasn't always been a smooth process. There have been moments where the Intelligence Index looked extremely saturated and stopped giving us a good indication of which models were best. There were moments when benchmark maxing allowed developers to game the system and rank much higher than real-world testing would support. In fact, this has been a fairly consistent story when it comes to Chinese open-weight models. In early September, Artificial Analysis pushed an early update with version 4.2 pulling forward many of the changes they had planned for V5. That decision was triggered by the release of GPT-6 Astra which focused heavily on improving computer use. The problem was that V4 one of the Intelligence Index still had relatively light weightings for agentic benchmarks leading GPT-6 Astra to have a slightly disappointing score. And more than just disappointing, one that didn't really comport with people's lived experience of the model. After the update, Astra was still ranked slightly behind Fable 5.1, but the index had become a closer reflection on how people were actually using the new crop of models. And this has been the story of the Intelligence Index, a constant cat and mouse game of updating the mix of benchmarks to match the rapidly evolving real-world use cases for AI. Introducing version 5 of the index, Artificial Analysis highlighted that for the first time, the majority of the benchmarks are built on private data sets. The idea is to make benchmark maxing more difficult as model developers theoretically can't train their models to the test. And that by the way is what we mean by benchmark maxing, when developers include those sort of evaluation tests and their answers in the training data set leading to an artificially inflated score as the benchmark ceases to measure what the model can do and instead reflects something a memory of the correct answers. For version 5, AA added two new benchmarks. Frontier Code 2.0, which is the brand new version of Cognition's private benchmark of real-world coding tasks. This test features scoring based on whether a repository maintainer would merge the code. They also added Terminal Bench Science, which is an agentic scientific research benchmark. Humanity's Last Exam, once pitched as a near-impossible knowledge search task, has been upgraded to HLE Diamond, the even more difficult version. Similar saturation updates have been made to AA Omniscience, which measures knowledge and hallucinations, as well as Crux PT, a physics test. GDP val AA, the measure of agentic work in real-world settings, is out with the index now relying on AA Briefcase, Automation Bench AA, and GDP.pdf to measure white-collar work along more granular lines. Also removed are SciCode, a scientific research test, and AA LCR, which tests long-context reasoning. Now, rather than just a list of names and acronyms, these adjustments say a lot about how AI models have changed since before the index was released in January. At that time, long-context reasoning was one of the most critical features used for working on large code bases. Many coding models still had limited token context windows and suffered from fatal issues with context draw. Since then, improvements to compaction and memory have allowed us to take a lot of that long-context performance for granted. Agentic work is the other huge change that's evident from the redesign of the index. The coding benchmarks included in the index now only measure agentic performance as that has been cemented as the way developers use modern AI, and the science benchmarks are now testing real-world agentic research work rather than pre-existing scientific knowledge. For the time, Artificial Analysis has chosen to equally weight the index across four categories: knowledge work, coding, science in general, which encompasses both general knowledge as well as hallucination rate. In some ways it feels like that 25% waiting for science is a forward-looking bet based on how Anthropic and Open AI have been trading science and math headlines over recent weeks. So, what did this change in the scores? In version five, Opus 5.5 was still in the lead with a score of 58, at least until Google dropped Gemini 4 Carbon on Wednesday. GPT-6 Astra jumped into then second place with a score of 57, and Astra had previously been in sixth place, showing there's a pretty big jump based on the index giving more weight to computer use in scientific tasks. GPT-5.1 Soul takes third place up from ninth, followed by Sonnet 5.5 which slipped from second to fourth. GPT-6.1 Terra was in fifth up from 10th, suggesting the model is a little better than many assumed earlier in the week. In just a moment we'll get into what I think that reflects. Gemini 4 Argon gained a rank to tie Meta Wonderful in sixth place, which is a pretty notable drop for Wonder, which was previously in third, although many felt that that ranking didn't line up with real-world experience. Fable 5.1 is looking a little long in the tooth in eighth place, sliding from fifth. And in ninth place we have one of the most notable drops in the new index, with Kim K 3.1 Pro falling all the way from fourth. In fact, Artificial Analysis noted that the gap between open weight models and US proprietary models had opened up massively after the index rebalancing. They wrote, "The open weights gap widens from three points to 12. Kim K 3.1 Pro is the top open weights model at 46. Among models with released weights, DeepSeek V4.1 Flash and GLM 5.3 lead at 32, 26 points behind the leader, and below Claude 3 Haiku 5.5 with 36 points." Now, in addition to the new gulf between US and Chinese models, there was also a clear drop-off between the handful of leading labs and everyone else. Anthropic, Open AI, Google, and Meta all have models with scores of 50 and above. Fable 5.1 and Kim K 3.1 Pro are just outside that cut-off. However, there's then a 10-point gap between Kim K 3.1 and Claude 3 Haiku 5.5, which leads the rest of the pack. Basically, by adding more agentic tasks to the index, there's now a clear separation between frontier models and the rest. Artificial Analysis noted that costs vary wildly among the top 10 models that arguably make up the frontier. That's largely because the top 10 is a mix of large models, token intensive models, and a few smaller, more efficient models. At each end of the spectrum, we have GPT-61 Terra as the cheapest at 96 cents per task, and Fable 5.1 is the most expensive at $13.04 per task. For a more apples-to-apples comparison based on model size, Opus and Sonnet both cost around $10 per task compared to 624 for GPT-6 Astra, and only a buck 27 for GPT-61. Muse, Wonder, Grok 4.7, and Gemini 4 Carbon are all slightly more expensive than GPT-6 Astra in the high $7 range. Now, most people found this to be a welcome update. OpenAI head of strategic futures Dean Ball wrote, "This is the first time an Artificial Analysis ranking has even roughly approximated my own subjective ranking of models." Haseeb Qureshi wrote, "Artificial Analysis re-ranking their intelligence score seems to match the vibes a lot better and clear out a lot of bench marking. Shows a deeper gap between Anthropic, OpenAI, and everyone else. Also shows a bigger gap between closed and open weights." Ethan Mollick wrote that the analysis, quote, "matches the general vibes that we haven't seen a Fable class open model." And AI content creator Theo drove that point home, writing, "Terrible day to be an open weights models are just as good as Opus 5.5 guy." Now, there were a few people who spoke up for the open weights models. Brookings Institute fellow Kyle Chan pointed out that even before this, Artificial Analysis had some private benchmarks. It's just that the share went from 45% to 60%. Using Kimi 3.1 Pro score as an example, he suggested that while the headline news is the gap with the top US model widening from three points to 12, that it's more important to look at the specifics. Science, he points out, is a major weakness, with Kimi scoring just 34% on Terminal Bench Science versus 63% for GPT-6 Astra. But, he writes, "Kimi is still extremely strong at coding agents," pointing out that it actually leads all models on Terminal Bench 4.0, scoring 67%. Kyle's conclusion, "Is this evidence of benchmark gaming? Not necessarily. Some of Kimi's biggest deficits are on non-private benchmarks, while it actually beats Claude on the private automation bench double A evaluation. The bigger story, he says, Chinese models can match or beat US frontier models on certain agentic tasks, but a significant gap remains in advanced scientific reasoning and research workflows. And ultimately, even if in general people think that this is a positive update, no one is going to agree on everything. In particular, I think one of the biggest debates will become what the relevant waiting should actually be. Yes, OpenAI and Anthropic are clearly emphasizing scientific discovery, but relative to the average individual use case, does that matter as much as agentic coding as it does in the waiting of the new V5 index? Now, that might be an argument for a different waiting, or it might be an argument that instead of paying attention to the headliner number, we'd be better suited to focus on the individual benchmarks for the categories of use cases that we care about. One critique that I do think is worthy of consideration is the fact that the leaderboards are run based on max settings. Sri. Dat.World writes, "No one runs max in real life. Where's the medium/high/extra high leaderboard?" And to be clear, it's not like Artificial Analysis ignores that, especially in their analysis that involves cost, the effort setting does become part of the calculus. But I do think that on a practical level, people might be better served by the ability to organize the index not just by model, but also by effort. Now, one last note, alongside the new intelligence index, Artificial Analysis introduced a new progress index. The index places all models released since November 2022 on a capability scale to see how each lab's leading model improves over time. It's designed less to be about the current batch of models and more to give a long-term view of model progression. Still, the new measure gave people easy access to a bunch of data to play with. Usama al-Ghareeb, for example, looked at how the switch to reasoning models impacted the major labs, commenting, "OpenAI really does have the best reinforcement learning, or their models just suck without reasoning." And by charting the data in a different way, they estimated that the gap between the US and China is now 3.3 months and closing. Summing up, like I said at the beginning, I think benchmarks are important and inherently limited, but I love that more effort is going into improving them, and I think even the debates that follow are also extremely value accretive when it comes to how we understand how different models compare to one another. For now, however, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always, and until next time, peace.
Yesterday, Google DeepMind released something called Genie 4. You type a sentence like a motorcycle ride along the coast or a diver exploring a sunken cave, and within seconds you're inside that world. You can walk around it, look around it, and change it while you're in it. It even makes its own sound. Nobody built that world. There are no 3D models, no game designers, no physics engine. An AI is imagining every single frame live as you move. And when most people see this, they think, "Cool, a new AI video game toy." But I think that's completely missing the point. Because the real reason Google is building this has almost nothing to do with games. It's about giving robots something they've never had before, an imagination, a way to predict what's about to happen in the real world before they act. And if Google gets this right, then in a few years, Google could quietly end up with the most advanced robotics business on the planet. Oh, and it also makes one of the strangest ideas in philosophy, that we might be living in a simulation, feel a lot less crazy. Here's Demis Hassabis, the head of Google DeepMind, explaining what's actually going on inside these models. >> This model is reverse engineering intuitive physics. So, you know, it's watched many millions of videos and YouTube videos and other things about the world, and just from that it's kind of reverse engineered how a lot of the world works. It's not perfect yet, but it can generate um a consistent minute or two of um interaction as you as the user uh in many, many different worlds. >> makes a clip. You type a prompt, you wait, and you get a video. You can watch it, but you can't do anything with it. Genie 4 is different. [music] It makes a world. You press forward, and it imagines what you'd see if you walked forward. You turn left, and it imagines what's to your left. It's predicting the next moment over and over, many times a second, based on what you're doing. And the way it learned to do that is fascinating. Nobody sat down and taught it the rules of the world. Nobody told it that water splashes, that shadows move when the sun moves, or that if you drive forward, the road gets closer. It learned all of that by watching an enormous amount of video. Think about a person who has watched so many films, so many travel videos, and so many nature documentaries that they can close their eyes and picture walking down almost any street in the world. That's roughly what Genie is doing, except it's not just picturing it, it's showing it to you live [music] and letting you steer. Google CEO Sundar Pichai shared a clip of a motorcycle ride along the beach. Everything you see there is being generated in real time. The road, the waves, the light. Justine Moore, who got early access, shared a diver exploring an underwater cave full of treasure and said it felt like a massive step forward. Ethan Mollick, a professor who writes a lot about AI, typed in a neo-gothic brutalist city full of robots and horses, and then asked it to add dragons, and then a friendly otter. All in real time. Ethan also made a Victorian tea party with a mouse in a little jacket, then added a porcupine, made the mouse grow wings, and had it fly over to the teacup. Ethan's verdict was that it's glitchy, but also impossibly advanced compared to a couple of years ago. And one of the people building it, Tejas Kulkarni, a research director at Google DeepMind, even showed Genie 4 simulating the 3D design software Blender in real time. Not running Blender, imagining what Blender would look like as you use it. Here's what's new in Genie 4 compared to the last version. It's more realistic. Characters move more and splash more like they should. You can type instructions in the middle of a session to change the weather, move the camera, or add people and objects. It can keep a scene consistent for up to around 2 minutes, which is about double what the last version could do. And for the first time, it creates sound on the fly. Music, background noise, and even characters you can talk to. Now, it's not perfect. Google says the longer you stay in a world, the more details start to drift. Objects can change shape or disappear. Text in the world is often unreadable. But, that's not really the point. The point [music] is what it can already do and how quickly it got here, which leads me to point number two, which is how fast this has moved. The first Genie came out in early 2024. It made blurry two-dimensional worlds, a bit like an old platform game, at about one frame per second. That's basically a slideshow. Genie 2 came out at the end of 2024. It could do 3D worlds, but only for 10 to 20 seconds at low quality. Then, in August 2025, Genie 3 arrived, and it could run in real time at 24 frames per second in sharp video for about a minute. In January 2026, Google let paying subscribers try it through something called Project Genie. And now, Genie 4 with sound, better realism, and double the length. So, in about 2 and 1/2 years, it's gone from a blurry slideshow of a 2D platform game to photorealistic interactive worlds with sound that you can steer with your voice. And the reason it keeps getting better so quickly is the same reason chatbots got better so quickly. More data, bigger models, and more computing power. Google's own researchers have compared where world models are today to where chatbots were around 2021, just before ChatGPT changed everything. If that comparison is even roughly right, the really big moment for world models hasn't even happened yet. That's the speed you need to keep in mind for the rest of this video. Because whatever this looks like today, it's going to look very different in 2 years. Which leads me to point number three, which is when you'll actually get to use it. Right now, Genie 4 is in what Google calls a research preview. That means only a small group of trusted testers have it. People in film, gaming, education, journalism, and engineering. Google says that in the coming weeks, it will roll out to Google AI Ultra subscribers through Project Genie. Ultra is Google's most expensive AI plan, and in the US, it costs around $250 a month. It's also coming to some early access developers through Google's AI tools for businesses. So, for most people, the honest answer is not yet. If you're willing to pay for the top plan, maybe in a few weeks. For everyone else, my guess is that you'll see it first in other people's videos and in apps built on top of it. Expect your social media feeds to fill up with people driving through made-up cities, walking around inside famous paintings, and recreating their childhood homes from a single photo. That's usually how these things go. The creators get it first, and everyone else sees it through them. And over the next year or two, as it gets cheaper to run, I'd expect it to slowly trickle down to cheaper plans, the same way every other AI tool has. When ChatGPT first came out, the best models were expensive and limited. Now, very capable AI is free on your phone. The same thing will probably happen here. But there's a big reason it's going to take a while, and it's the next point. Which leads me to point number four, which is that this thing is probably incredibly expensive to run. When you watch Netflix, the video was made once, and Netflix just sends you a copy. Millions of people can watch the same file. But Genie can't do that. Every person's world is different, and it changes depending on what they do. So, the AI has to imagine every single frame live, just for you. About 24 times every second. Using a giant AI model running on Google's own specialized AI chips inside Google's data centers. Think about what that means. One person playing for 1 minute is over 1,000 AI-generated images. Each one made instantly. Each one matching what came before and each one reacting to what you just pressed. Now, imagine a million people doing that at the same time. And we've already seen the evidence of how expensive this is. When Google launched Project Genie earlier this year, sessions were limited to 60 seconds and reports said that was because of compute costs. Google's researchers have also said that letting it run longer would add a lot of cost for relatively little benefit. That's why it's only on the most expensive plan and why sessions are short. The good news is that this always gets cheaper. Every year, AI chips get faster and researchers find clever ways to make models smaller and more efficient. Waymo, for example, already made a smaller and cheaper version of its world model so it could run more simulations for less money. So, the cost per minute of a Genie world will almost certainly keep falling. At some point, a version of this could even run on a games console or a powerful laptop. But, we're not there yet. For now, every world you walk around in is being dreamed up by a server farm. Google hasn't said exactly how much each session costs them, but it's safe to say that right now this is one of the most compute-hungry consumer AI products ever made. And that's actually a big advantage for Google because most AI companies have to rent their chips from Nvidia. Google designs its own. So, if anyone can afford to run something like this at scale, it's Google. Which leads me to point number five, which is why this could change video games. Because the key word with Genie is real time. Most AI video tools take minutes to make a few seconds of footage. Genie responds instantly to what you do. And anything that responds instantly to what you do is basically a game. Today, making a big video game can take hundreds of people and 5 years. Every tree, every building, every character is designed by hand. Every rule of physics is programmed by hand. Demis Hassabis actually started out making video games. So, this one is personal. Now >> And for me, it's it's it's very close to my heart, this project, but it's also quite mind-blowing because in the '90s in my early career, I used to write uh video games and AI for video games and graphics engines. And I remember how hard it was to do this by hand, program all the polygons and the physics engines. Um and it's amazing to just see this do it effortlessly. All of the reflections on the water and the the way materials flow um and and and objects behave. And it's just doing that all out of the box. >> Imagine where this goes. A game that builds itself around you as you play. You could say, "I want to explore an ancient Roman city at night." And it just appears. Characters you can actually talk to who answer you in real time instead of repeating the same three lines. Worlds that never end because the AI just keeps imagining what's around the next corner. A kid could describe a game and play it 5 seconds later. And Google isn't the only one chasing this. A company called Runway released its own playable world model with sound a couple of months ago. And the money flowing into this area is enormous. Reports say AMD is buying a world model company called World Labs for around $8 billion. And Amazon is reportedly in talks to buy another one called Decart for around $7 billion. When companies start spending that kind of money, it tells you they think this is going to be a very big deal. I don't think this replaces normal game engines anytime soon. Right now, worlds only stay consistent for a couple of minutes. And games need rules that don't randomly change. But for brainstorming, prototyping, and brand new kinds of games that couldn't exist before, this is huge. And remember the speed from point two. Two and a half years ago, this was a slideshow. Which leads me to point number six, which is the simulation hypothesis. So, the simulation hypothesis is the idea that our entire reality might be a computer simulation. The most famous version comes from a philosopher called Nick Bostrom back in 2003. And the simplest version of the argument goes like this. If a civilization ever gets powerful enough to create perfectly realistic simulated worlds, it would probably create lots of them. And if there are billions of simulated worlds and only one real one, then the odds are you're probably in one of the simulated ones. Bostrom's actual argument is a bit more careful. Bostrom says one of three things must be true. Either civilizations almost always die out before they get that powerful, or they get that powerful but choose not to run simulations, or we are almost certainly living in one. And we don't know which. Elon Musk explained his version of this on the Joe Rogan podcast, and it's all about how fast games get better. >> Now, if you entertain that, well, the argument for the simulation I think is quite strong. Because if you assume any improvements at all over time, any improvement, 1%, 0.1%, just extend the time frame, make it a thousand years, a million years. The universe is 13.8 billion years old. What would a civilization, if you count it, if you're very generous, civilization is maybe seven or 8,000 years old, if you count it from the first writing. This is nothing. This is nothing. Um so if you assume any rate of improvement at all, then games will be indistinguishable from reality. Or civilization will end. One of those two things will occur. >> Think about that argument and then look at Genie 4. 50 years ago, the most advanced video game was Pong, two lines and a dot. Today, an AI can watch videos of the world and then imagine a whole new world with light, water, sound, and physics that reacts to you in real time. Nobody programmed the rules. The AI worked them out by watching. And each version is dramatically better than the one before. That's why Genie makes the simulation idea feel more real. It's not proof of anything. A world model that can keep a scene together for 2 minutes is a very long way from simulating a whole universe. But it shows that a computer can generate a world just by learning the patterns of how things behave. And if we can already do this in 2026, it's not hard to imagine what we might be able to do in 100 years or 1,000. Interestingly, Demis Hassabis has a more nuanced take. Talking to Lex Fridman, Demis said the idea of billions of throwaway simulations isn't really convincing. But Demis does think the universe might be best understood as information. >> And matter is actually just, you know, we're we're just out the way our bodies and all the molecules in our body arranged as information. So, I think information may be the most fundamental way to describe the universe and therefore, you could say we're in some sort of simulation because of that. >> [music] >> Um but I don't I do I'm not I'm not really subscriber to the idea that um you know, they these are sort of throwaway billions of simulations around. I think this is actually very critical and possibly unique this simulation. >> And here's something that makes this even stranger. You already run a world model every night. When you dream, your brain builds a whole world with people, places, light and sound and none of it is coming from your eyes. It feels completely real while you're in it. Your brain is generating it frame by frame from everything it has learned. That's not so different from what Genie is doing. Which raises a slightly uncomfortable question. If a brain can generate a convincing world and now a computer can too, how confident are we about which world is the real one? Whatever you believe, it's a strange feeling to watch an AI build a believable world out of nothing and then look around at your own. Which leads me to point number seven, which is the real reason Google is building this and why most people are missing it. Here's something you probably don't think about. Your brain is a prediction machine. When someone throws you a ball, you don't wait to see where it lands. Your brain predicts where it's going and moves your hand there. When you walk down a busy street, you're constantly predicting what other people will do. That person is about to step out. That car is going to turn. Scientists who study the brain have been saying for years that a lot of what we call seeing is really our brain predicting what comes next. Robots are terrible at this. Most robots today can't really imagine what will happen if they take an action. They just react. And that's one of the biggest reasons robots are still clumsy in the real world. A world model like Genie is basically that missing piece. It's an imagination. Before a robot does something, it can run a quick simulation in its head. If I push this cup, what happens? If I step here, will I slip? If I cross the road now, what will that person do? Here's Shlomi Fruchter, one of the people who led the Genie project at Google DeepMind, explaining exactly this back when Genie 3 came out. >> Another thing is that we're excited about is actually using those simulations for [music] planning. So, if you have a robot or you have some again an agent that wants to accomplish a goal, they can maybe do some rollouts in the simulation, figure out my what might happen. For example, if they want to go across the road and they they the the agent can use the model to predict a few options. Maybe there is like a few scenarios. Maybe a person going to cross its path. Maybe something else going to happen. And then using those rollouts, it can decide what's the next action it should take. And then this is used for planning. >> It's the bit people are missing. Genie isn't just a game. It's a way for a robot to see into the future, a few seconds at a time. Imagine a home robot clearing your dinner table. There's a full glass of water near the edge, a plate with a knife balanced on it, and your cat about to jump on the table. A robot without a world model just grabs things one at a time and hopes for the best. A robot with a world model can run through a few quick imagined futures. If I grab the plate first, the knife might slide off. If I move my arm here, I'll knock the glass. The cat is about to jump, so wait a second. Then, it picks the best plan. [music] That's the difference between a robot that breaks your stuff and a robot you'd actually trust in your home. And it's also a training ground. Teaching a robot in the real world is slow, expensive, and dangerous. If a robot falls down the stairs, it breaks. If a self-driving car makes a mistake, someone could get hurt. But in a world model, you can create millions of situations, including rare and dangerous ones, and let the robot practice safely. A kitchen fire, a child running into the road, a slippery floor. And this is already happening. Waymo, Google's self-driving car company, built a world model on top of Genie 3 to simulate rare situations for its robo-taxis. Things like strange weather or something unexpected in the road. And in the Genie 4 announcement, Google showed it working with its Gemini robotics research, where Gemini breaks a task down into steps, and Genie 4 simulates them in real time. One demo even showed a humanoid robot from Apptronik, a robotics company Google works with, playing checkers. Demis has said for a while that this understanding of the physical world is exactly what's missing from robots today. >> And and and so we've been very interested in this because for an AI to be truly general, to build AGI, we feel that the AGI system needs to understand the world around us. And the physical world around us, not just the abstract world of languages or mathematics. And of course, that's what's critical for robotics to work. It's probably what's missing from it today. >> Which leads me to point number eight, which is why I think Google could end up with the most advanced robotics business in the world. Think about the ingredients you need to build a truly smart robot. First, you need a brain that understands language and can plan. Google has Gemini. Second, you need something that understands physics and can predict what happens next. Google has Genie. Third, you need a huge amount of video of the real world to learn from. Google owns YouTube, the biggest collection of video on Earth. Fourth, you need massive amounts of computing power cheaply. Google makes its own AI chips. Fifth, you need real robots in the real world collecting experience. Google has Waymo, which already has self-driving cars carrying real passengers in multiple US cities. And sixth, you need the people. Google DeepMind is one of the best AI research labs in the world. Now, to be fair, Google isn't the only company with big robot ambitions. Tesla is building its Optimus humanoid robot and has a huge amount of driving video from its cars, though it recently pushed its Optimus plans back to late 2027. Nvidia has its own world models called Cosmos, built to help train robots, and it sells the chips that almost everyone uses. Companies like Figure and Apptronik are building humanoid robots, and Meta, OpenAI, and others are all working on this, too. But almost no other company has all six. Some have great robots, but weaker AI. Some have great AI, but no robots, no video library, and no chips of their own. Google has the whole stack. And we've actually seen this movie before. Google started working on self-driving cars around 2009. For years, people joked that it was a science project that would never make money. Then, quietly, Waymo became the first company to run fully driverless taxis for the public at scale, while everyone else was still testing. I think there's a real chance robotics goes the same way. Years of quiet research that people dismiss as demos and toys, and then one day, Google is just ahead. And Google's robot AI is already moving fast. Google DeepMind has been releasing Gemini robotics models that let robots understand instructions, use their hands more precisely, control their whole body, and even work together. Each one of those robots gets better when its brain gets better. And Genie is a big part of making that brain understand the physical world. So, when you see a mouse in a teacup or a dragon flying over a city and think, "It's just a game." remember this. The same technology that imagines those worlds could be what lets a robot load your dishwasher, help in a hospital, or work in a warehouse without breaking things. And Demis thinks the big moment for robots isn't that far away. >> I think in the next couple of years there'll be a sort of real wow moment with robotics, but I think the algorithms need a bit more development. The general purpose models that these these these robotics models are built on still need to be better and more reliable and and better understanding the world around it. I think that will come in the next couple of years. And then also on the on the hardware side, the key is I think eventually we will have millions of robots helping helping helping society and increasing productivity. >> Which leads me to point number nine, which is what this means for you. In the short term, you'll mostly see Genie in entertainment. New kinds of games, interactive stories, and creators building worlds in minutes instead of months. Filmmakers are already using it to quickly sketch out scenes before they shoot them. A director can walk around a set that doesn't exist yet, try different camera angles, change the time of day, and try 10 versions of a scene in an afternoon before spending a single dollar on a real set. One of Google's engineers called it rapid ideation. It won't replace filming, but it could change how films get planned. Teachers could use it to let students walk around ancient Rome or the inside of a cell. Google even suggested you might one day step inside a historical painting and talk to the people in it. And if you want to try it yourself, keep an eye on Project Genie over the next few weeks. Even if you don't pay for it, a lot of the most impressive things people make with it will end up online and it's worth watching because it's one of the clearest signs of where AI is heading. There are downsides, too. If you work in game design, 3D art, or visual effects, your job is going to change. Not overnight, but it will. And as these worlds get more realistic, it'll get harder to tell what's real footage and what's generated. Google says everything Genie makes carries an invisible watermark called SynthID, which is a good start, but it won't solve that problem on its own. In the longer term, and this is the big one, world models could be what finally makes robots useful in everyday life. Robots that can predict what's going to happen are robots that can work safely around people, in your home, in hospitals, in care homes, in shops. That's a much bigger deal than any video game. Which leads me to my honest take on all of this. Here's what I think. Genie 4 looks like a toy, and it's being shown off like a toy. Motorcycles, dragons, and mice in teacups. And honestly, it's an amazing toy, but I think that's the least important thing about it. It's still early. It's expensive to run. It only stays consistent for a couple of minutes, and the details drift. Most people won't get to use it for a while. But look at the speed. In about 2 and 1/2 years, it went from a blurry slideshow to real-time worlds with sound. If it keeps improving at anything like that rate, the version we get in 2 or 3 years will be hard to tell apart from reality, at least for short periods of time. And the real prize isn't games. It's giving machines an imagination, a way to predict the real world before they act in it. Google has the brain, the imagination, the videos, the chips, the cars on the road, and the researchers. [music] That's why I think most people are sleeping on this. In a few years, the company everyone currently thinks of as the search engine could be the company that builds the smartest robots in the world. So, what do you think? Is Genie 4 just a cool toy, or is it the start of something much bigger? And does it make you think about the simulation idea a little differently? Let me know in the comments. I read them. And if you want to stay ahead of what's happening in AI explained in plain English, make sure you subscribe. I'll see you in the next one. >> Mhm.
So what is Docker and what's the point of it? Say you clone a teammate's project. Before you can run it, you need the right version of Node, a Postgres database, and the same settings they have. And if one thing is off, it works on their machine and breaks on yours. Docker is a tool that lets you package and run your apps in something called containers. A container is basically a box that holds your app along with everything it needs to run like the runtime, the libraries, and the config. As long as a machine has Docker installed, your app runs the same way. Every container comes from an image. The image is the readonly blueprint, and a container is a running instance of it. You can start as many containers from one image as you want. Sort of like a class and objects in programming. To make an image, you write something called a Docker file, which is just a text file with instructions. This one says, "Start from a node image, copy in your code, and install the dependencies." The last line is the command that starts the app. Then docker build turns that into an image and docker run starts a container from it. And you don't have to build everything yourself. Docker hub is basically GitHub for images. For that project, you don't install Postgres at all. You just run the official Postgres image. Now a lot of people think a container is just a virtual machine, but it's not. A VM runs a whole operating system of its own. Containers all share one kernel, which is the core of the operating system. So, they're a lot lighter and they start up fast. I don't think you need Docker for everything. If you're building something small on your own machine, that's totally fine. But once there's a database, a team, or a server involved, it makes life a lot easier. So, at its core, Docker packages your app with everything it needs to run the same way anywhere. So, hopefully this helped you guys out.
I did not think this would happen in my lifetime. Let me try to explain. For many decades now, computer scientists have been thinking up devilishly difficult problems and found that trying to solve some of them takes incredibly long. Some of them would take a quadrillion years to solve by brute force. But we humans have invented so many amazing algorithms. We can write programs that simulate light rays to imagine never before seen scenes. I love this. so much and we can write physics simulations to create really lifelike or imaginary worlds. Absolutely ingenious human developed techniques and you are saying that some problems just cannot be solved efficiently. Well, that is exactly what the exponential time hypothesis says. It says that some problems are cursed and might always remain so. Here is a cursed problem. The cat is sleeping or the dog is barking or the robot is dancing. When is this true? Well, just make the cat sleep. Problem solved. Easy peasy. What is Caro even talking about? Get out of here. No, no, no, no. Wait, let's add another line. The cat is awake or the dog is silent or the robot is dancing now. The first one wants the cat to sleep. The second one wants the cat to be awake. So, we have to look for help with a different variable. Let's look at the dog. Oh my, same problem. The dog is barking in the first, silent in the second. But wait, the robot is luckily dancing on both. So, that can satisfy both lines. Hooray. But the point here is that one line was very simple, but two lines needed a bit of computation. Still simple, of course. And now you'll see the crazy part. We call this the three satisfiability problem. And if you increase the problem size with more animals and lines, what happens? Well, just get a processor that goes through a billion combinations per second and you are done instantly, right? Not quite. This computer program would take so long it might never finish. Well, it could take one quadrillion years to finish if we check every combination. Well, then this is so many combinations. So, better write an algorithm that does it a smarter way and do it a lot quicker. And here comes the crazier part. We couldn't. Nobody had ever been able to do that in the general case. Yes, we can be quicker than just trying all the combinations. We have smart algorithms but none of them could defeat the exponential barrier. This is called the exponential time hypothesis. So it's over right? Do we have no hope left to solve this? Dear fellow scholars, this is two minute papers with Dr. Koa Eher. Well, all this dril means that this problem is cursed. Cursed how? Well, cursed in a sense that as the problem size grows even just a tiny bit, it gets much harder. In the worst case, exponentially harder. Ew. Now, we can't prove it, but so many people believe it to be true that many papers have been written assuming that this is true and building on that. And then something insane just happened. Hold on to your papers, fellow scholars, because the age of AI has led us into a breakthrough. OpenAI published a research paper where they say they proved that this hypothesis is not true. They say that it is possible to defeat the exponential barrier. I'm going to be honest, I did not think this would happen within my lifetime and it just did. if the proof gets properly verified. It looks increasingly convincing, but the expert review is not 100% done yet. But if it holds up, then one of the most important assumptions of theoretical computer science has just fallen. That is insanity. Now, I will note that this is not to be confused by the P question mark equals NP problem. That has not been solved. Now, you could say, "Okay, Caro, but who cares about this silly satisfiability problem? Just bring your dogs and cats elsewhere." Well, I say you have to consider carp reductions. Before you ask, no, this does not mean shrinking a fish. It means something completely different. Imagine you have to deliver life-saving medicine to hundreds of hospitals and you are a bit low on gas. You need the most efficient route. Or imagine you have to assign radio frequencies to thousands of transmitters so they can talk without too much interference. Or place the fewest number of sensors needed to monitor an entire system or select the most valuable scientific equipment to load it onto a small spacecraft. All of these are very different problems except that they are not. All of these problems can be molded to look like our three satisfiability problem with the cats and dogs. You take your roads and transform them into a dog and cat question. That is carp reduction. Wait, and does that mean that? Oh yes. [laughter] Yes, it does. If you can write an algorithm that can do that one problem faster, it could work for all of these problems and then even more. Not always, but it's often the case. And that is why this work is so incredibly surprising and valuable. Plus, it gets better. The paper reveals not just the proof that such an algorithm exists. The paper actually gives us the algorithm. And it gets even better. It not only finds the solution, but also tells us exactly how many solutions are there. This is a historic moment in our lives. What a time to be alive. And just to demonstrate how incredibly efficiently scientists can communicate with each other, all of this summed up in short. A subexponential bound has been claimed for three set refuting the exponential time hypothesis. Very concise and super awesome. If you enjoy this, consider subscribing and hitting the bell. We have a lot more papers coming and if everything goes well, an interview with Spencer Huang as well. Very fun. I use Lambda to reproduce AI research papers often in minutes. It's also great to train your own models or fine-tune an existing one. Run inference or text to image or video. Easy peasy. Running a Deep Seek chatbot or agent. Super fast, super reliable. Lambda gives you powerful Nvidia GPUs to run your own experiments. I test ideas from the papers I cover and moments later, results. Love it. Seriously, try it out now at lambda.ai/papers. Ei peepers.
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=vhqq4U3mnNQ! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
What happens if I lower this weight into the water? Is the scale going to go up, down, or stay the same? >> Oh god. >> Oh. >> Um, >> no. It will get lighter. >> I think it might get a little bit heavier. >> I think it might add like a little weight. >> It's not going to touch the bottom. It's just going to I'm just going to hold it in there. >> You're not going to touch the bottom. >> No, just hold it. >> Ah, that then it will not move at all. >> Then there's no change. I think >> I don't think it's [music] going to change. >> Let's see it. Let's see it. What? >> Oh, well, it's increased by 13 g. >> Why? >> When you lower the weights into the water, the water pushes up on it. That's buoyancy. But [music] that push works both ways. The weights push back down on the water with exactly [music] the same force. And that downward push travels through the water and into the scale. If I take the weights out of the water and just weigh them themselves, you can see they're 100 g. So, the weight of the beaker is not increasing by the weight that we've added. It's increasing by 13 gram, which is the weight of the water we've displaced. >> So, that's why you're reading 13 g extra. It's the weight of the water you've displaced. >> Wow. It's crazy. That's crazy.
This is my childhood game cupboard. This is my childhood living room. And this is my childhood favorite park. And I'm here today to offer $100 to anyone who beats me in any of them. >> Is this for real? >> Yeah. >> Dirty dog. >> Uh-huh. >> I won. A >> next. >> I won. I won. Man, >> the game's over. >> I got 400. Next, >> I feel like you owe me $100. Do you have $100? >> And the competition was unrelenting. Round after round, my strategies were proving unstoppable, which wasn't a surprise to me because for the last 6 months, I played hundreds of games, ran millions of simulations, and analyzed a bunch of data to discover the secret strategies to win every single time. Almost every single time. >> THAT WAS THE WIN. AND ALL THE TRICKS yourself right
So, you know that thing [music] that you keep on worrying about? To be honest, it's not going to matter in a few years. It's certainly not going to matter in a few decades. And certainly when you're on your deathbed, it's [music] not even going to vaguely register. So, it's not to say you shouldn't do it, but like you can totally do it. Just stop worrying about it so much. Just approach it with a little bit more lightness, a little bit more ease. Try and enjoy the process cuz it's not going to matter anyway in a few decades. [music] So, why waste all of that effort worrying about it?