Video transcript
Transcript from available captions; it may contain errors.
In 1997, someone at the NSA typed a handful of numbers into a computer. Those numbers now help protect a huge share of the internet. For 25 years, nobody could explain where they came from. This we claude anthropics AI figured it out in under 30 minutes. And halfway through, it stopped and asked itself a very strange question. Here is what it wrote. Quote, "This is astonishing. It solves the 25-year-old mystery. H wait, am I in some sort of simulated or test environment. End quote. A few lines later, it decided, wow, it decided, wow, it's a real discovery. If you have ever logged in with the pass key or opened a website over HTTPS, you have almost certainly touched the curve those numbers built. It's called P256. So, yes, this story is about you. In a few minutes, I'll show you exactly how I cracked it down to the date the second and candidate number 1,842. But here's the thing, it's not the most important story of the week. One of the others quietly matters more. And I'll tell you which one at the end. And stay for the end because I'll also tell you the one change I'm making to my own stack this week. It's not what you guess. This is Zoom MX the weekend AI. I'm T here. Here's the map. Google shipped a frontier model with a price that expires next month. Meta came back to the top tier and then gave away an 800 billion parameter model you can run at home. Deep stroked open weights under the MIT license. Anthropic found that small models may soon spread like worms through cheap TV boxes. Europe started actually enforcing its AI law. and eight US states sued Open AI, plus the open-source image, video, and 3D tools that dropped this week, including world models, and a leaked revenue number that makes all of it make sense. But first, the mystery. Story one, a 25year-old cryptography mystery solved faster than a lunch break. To see why that's a big deal, you need to know what a seed is. In 2000, the United States standardized a family of elliptic curves, P192, P224, P256, P384, and P521. The standard was called F5186 version 2. Each curve was generated from a seed, a randoml looking number supplied by the NSA. Here's the idea behind a seed. You feed it through a hash function and the output becomes the cur numbers. Publish the seed and anyone can check that nobody handpicked the curve. Cryptographers call that verifiably random. The problem nobody could verify where the seed itself came from. The NSA gave no explanation for those seeds. None. And when a spy agency hands you magic numbers for your encryption, people get nervous. For decades, the suspicion was simple. What if the see-through chosen to hide a back door? There was a popular story to explain it. Supposedly, NSA mathematician Jerry Selenas hashed a joke phrase, something like give Bob and Jerry a raise. It was a fun story. It was also wrong. Here's what actually happened. An Entropic employee asked an internal research model to investigate. And this part is key. It was asked without the joke phrase clue. No hints, just the seeds and one question. Where did these come from? The model noticed something tiny. In four of the seeds, every 32-bit chunk started with a zero. The top bit every single time. The odds of that happening by chance are about 1 in 5,810. That pattern is a fingerprint. There's an old C library function called L ran 48. It generates random numbers, but it only outputs 31 bits. Pack those into 32bit words, and the top bit is always zero. Exactly what the seeds showed. An LRAN 48 has a property that detectives love. You seed it with a number. Very often the current time, same time and same numbers out. So the question became simple. What time was it on that computer? So, the mobile search timestamps between May 1997 and April 1998. See the generator build a candidate compare over and over. On screen is a simplified sketch of that idea. The real script about 30 lines is on anthropics page. And then it hit. P256's published seed is candidate number 1,842 from a run started on August 16th, 1997 at 152 and 34 seconds in the morning UTC. The 25-year mystery was a time stamp. That's the moment it wrote the line from the start. Astonishing it solves the mystery. Then the doubt. Am I in a simulated test environment? Honestly, that's the most relatable thing an AI has ever said. You finally solve something and you assume it's a trick. Then a human checked. Filipo Valsoda, who organized the bounty, confirmed it and re-implemented the method in Go. The prize money went to ISRG, the nonprofit behind Let's Encrypt. With matching, it tripled to $36,864. Now, the part the headlines will get wrong. This is not a break. P256 is not cracked. Anthropic and Volorta both say this makes a back door less plausible, not more. A time stamp leaves at most about 30 bits of freedom. That's far too little to plant one. A few pieces are still unexplained. A B163 seed and the base points of P224, 256, 384, and 521. The model did explain some other base points tracing them to Arian Lensstra's free life library. My take as a builder. If you use P256 change nothing, this is good news. The real headline is speed. Forensic work that stumped experts for decades took under half an hour. Keep that in mind because AI and security come back twice in this episode and the second time it is not good news. Story two. Google just put its Frontier model in everyone's hands with a price tag that has an expiry date, literally. The price doubles on November 30th. On Tuesday, Google made Gemini 4 Argon Broadly available. It debuted on September 30th, but only for a small group of cyber defenders. Now, it's in the Gemini API and AI Studio. It's in Google Anti-gravity for Ultra subscribers. and it's in the Gemini app in AI mode for pro and ultra users. Here's the price. $2 per million input tokens, $10 per million output tokens. Cashed input is 95% off. That's introductory. After November 30th, it goes to $420. One catch, the free tier does not get Argon through the API. Neither does the Google AI Pro plan. Google's blog says paid tier. So if you were planning to test it on free credits, sorry. Google says Argon is number one on the Valves index. That benchmark is built on real work, finance, coding, legal, and tax. It also has a 1 million token limit in the API. Input and output combined with a continuation protocol for very long outputs. than the internal numbers inside antigravity. Agentic code submissions went up 35%. And roll backs dropped by about a third. Fewer roll backs is the number I actually care about. That's fewer reverts at 3 in the morning. Google also pointed Argon at its own code. Gemma 4 inference on an A100 through VLM got 2.85 times faster. That's from about 37,900 to 108,000 tokens per second. Google's ADK Python framework got 40% faster session prep overnight. The safety numbers attempts to shortcut coding benchmarks dropped to 2%. Prompt injection attacks on the Grey Swan test succeeded 0.7% of the time and it scored 87% on ERQA, an embodied reasoning benchmark. Which brings us to the best demo of the week. An Aptronic Capello humanoid robot watched a video with Agon doing the reasoning and then sank seven straight corn shots. Seven in a row. I can't do two. Google's post also shows playable games that Argon builds like a desk racer and a 3D wizard chest worth opening just to poke at them. Now the twist. Google says number one on bals. But on the independent artificial analysis intelligence index as reported by benturebeat, argon scores 53. Clopus 5.5 scores 58 and is still number one. So number one depends on whose scoreboard you read. Pause for a second. Would you switch your main coding model for a price that only lasts about 6 weeks? Or is switching costs the real price? Tell me in the comments. I'm genuinely curious where builders land on this. One nice touch. 10,000 scientists can apply for free top tier anti-gravity access for 6 months. If you work in research, that's an application worth filing today. My take. The model ID is Gemini 4 Argon and until November 30th, it's priced like a mid-tier model. Run your hardest long context and coding jobs on it now before the price doubles. The benchmarks are Google's. Your evils are yours. And the model that scored 56 on that independent index. It came from a company that hadn't been at the frontier in a while until Monday. Story 3. Meta is back at the frontier. And it did something odd. It kept the better model off its own API. On Monday, Meta launched Muse Wonder and Muse Wonderful Full. It calls them its most capable models to date. built for agents. Think insurance claims, leases, finances, and coding. They were trained at Prometheus Meta's gigawatt scale data center after a 14-month rebuild of its AI stack. Two headline features. Active memory remembers context across chats. Active focus keeps pushing through hard tasks. It retries and it checks in with you before big actions. That last part matters. An agent that asks before touching your database is a feature not above. Meta's research blog shows the agent side with everyday demos. The kitchen game, planning a run, even sorting out wedding seating. Wedding seating is honestly the hardest optimization problem humans have. Two aunts who don't speak one table. Good luck. Meta Muse Wonder is on the meta model API at $4 per million input tokens and 20 per million output. Catched input is 20 cents, 1 million tokens of context, and it's open AI SDK compatible. You change the base URL and you're basically done. Muse wonderful is the stronger one, and it is not on the API. You only get it inside Muse Code subscriptions. That's power at $50 a month or the new maximum plan at $200 with 80 times the usage limits of the everyday plan. On artificial analysis per venture beat Wonderful scores 56 tied with Claude Sun at 5.5. Plane Wonder scores 52. GPT6 Astra and Gemini 4 argon sit at 53. And Opus 5.5 still leads at 58. Meta's own table looks better for Meta, which is how these tables tend to work. It reports 79 on Deepest Wee against about 74 for Opus 5.5 and GPT6 Astra at self-reported a mix of internal and third party runs and Opus and Astra still win on Terminal Bench 4 and Healthbench Pro. Here's the catch. eyed circle in red news code subscriptions default to contributor mode according to venturebe that means your prompts and code can be used for training the pay as you go API uses standard mode with no training and neither one to model is open weight my take a strong coder opus level per token pricing drop in endpoint but check that data setting before you paste your company's repo into it and if you're thinking Fine, I'll wait for Meta's open weights. Meta heard you 3 days later. We'll get there right after a quick speed round. That's three stories down and the weirdest of the open-source news is still coming. But first, a speed round. Four stories fast. Keep up. One, Microsoft unveiled Maya 300, its own AI chip. Starting in 2027, Maya becomes the primary inference system for Microsoft products at multi-gawatt scale. Microsoft claims 40% more tokens per megawatt than leading GPU systems. The chip eight stacks of HBM4 18 terabytes per second of bandwidth 950 W built on TSMC's 2nmter class process. For builders, that means cheaper inference on Azour, Foundry, and Copilot over time. GitHub Copilot and Foundry move on to Maya by year end. And this isn't a paper chip. The current Maya 200 already serves Copilot and runs OpenAI, Microsoft's own MAI and Open models. Microsoft says it gets 1.7 times more tokens per watt than GPUs in Teams and Copilot. Maya 300 claims up to seven times the rack scale throughput of that. Two chat GPT can now read and drive your Chrome tabs mentioned at Chrome in a prompt. Chat mode reads across your open tabs. Work mode controls them. It sends messages, fills forms, and downloads files. It works on every chat GPT plan. And the extension reportedly has around 7 million users. Context: OpenAI retired its own Atlas browser on August 9th. So instead of making you switch browsers, it moved into yours. Same day, it also shipped GBTI 1.1 with new voices in French German Hindi Mexican Spanish Korean, and Japanese. So if your web app has forms, AI agents are about to fill them. Clear labels, stable selectors, and some thought about bot abuse. Also reported this week, OpenAI told developers the GPT40 when GPT4.1 API snapshot shut down on January 14th, 2027. Three, XAI's Grog voice 3.0 took first place on the artificial analysis speechtoech index at 86.9 as published by XAI. Latency is under 1 second. The API price stays at 8 cents a minute. That's $480 an hour. Starlink's call centers run it on 20,000 calls a day with about 70% resolve without a human. The per task numbers are what matter for builders, about 2 1/2 cents per task. GPT live one costs 3 cents and GPT realtime 2.1 costs 5.5. On Starlink sales calls, one in four ends in a purchase. $480 an hour is less than a human agent in almost every market. Four, and it's my favorite, Postgress 19 got delayed because AI code reviewers found too many bugs. The release moved to around November 19th. 8 to 10 significant features got pushed to Postgress 20, including SQL PGQ graph queries. Committer Peter Eentrout described staring at a list of 40 defect reports and then running out of time and he's clear that this is not by coding. Those features were written before that was even possible. The AI just raised the quality bar midcycle. The lesson if you add an AI reviewer budget time to fix what it finds. Hold on to that one. It comes back in the verdict. Quick one from me. If you want to work with Zumx on business deals or sponsorships, visit zumslabs.xyz or email to zumslabs.xyz. It's in the description. Now, the model you can run in your living room. Story 4, the openweight site of the week. Two big releases and one geopolitical twist. And one of these models runs on hardware you can actually buy at an Apple store. On Thursday, Meta released Muse Spark Open. It calls it the most capable US Open Wick model to date. 816 billion parameters in total, 42 billion active, 1 million tokens of context, natively multimodal, and built for coding in agents. Here's the headline. It runs on two Apple M5 Ultramax studios using MLX. Each one can have up to 512 GB of unified memory. together about 1 terabte. That's an 800 billion parameter model sitting on your desk. Is that cheap? No. You're looking at roughly $20,000 or more in Mac hardware, but no API bill and your code never leaves the room. For privacy sensitive work, that math starts to look interesting. Met his demo. An ondevice agent shipped real iOS changes to ice cubes, an open-source Mastadon client and drove Da Vinci Resolve and Twand weights are on hugging face and hosted by providers like Gro together AI fireworks and open router. It also ships helium GPU kernels in SG lang with kernel speedups of up to 3.95 times. The catch. The license is the Muse model license with limited commercial terms. Downloadable, yes. Fully unrestricted, no. Read it before you ship it inside a product. Then on Saturday, Deepseek did the opposite. Deep Seat 4.1 flash 1017. Open weights under the MIT license. That's the do whatever you want license. 1 million tokens of context with a recommended max output setting of at least 256,000 tokens. Deepc calls it a minor update. Stronger benchmarks, faster task completion, shorter outputs. It's live on the web, the app, and the API under the same Deepseek Flash name. The parameter count depends on who you ask. Call it hundreds of billions in total with a tiny active count. Deep Seek's own table self-reported 77 on Deep CEUE ahead of Claude Opus 5 at 74 on Automation Bench 60.5 against Opus 5's 50.3. But on Terminal Bench 4, Opus 5 still wins clearly 51.8 to 42.3. The sneaky useful part is new search and fetch APIs. 75 cents per thousand calls each and they're free through claw codeex and deep sea carnis until October 31st. That's a cheap search tool you can bolt onto any agent you're building. Now the twist. On Friday, China proposed export controls on its biggest AI models. Draft announcement number 48 adds 40 technologies to the control list. 23 of them in AI. It would restrict exporting weights, checkpoints, source code, and architecture for models with two trillion parameters or more or ranked in the world top 10. Bloomberg and AP named Moonshot's Kimmy K3 and Alibaba's Quizen 3.8 Max as examples. Models trained on China's core data would be banned from export outright, but open-source tech that's already public is exempt. Public comments run until November 15th. It's my take. Chinese open weights like Deep Seek look safe for now thanks to that exemption. The biggest closed checkpoints may stop crossing borders. And this week, open source wasn't only about text. The image, video, and 3D site had a huge week, too. What can you actually run on your own GPU this week? More than last week. A FUX3 image model that fits on an 8 gig card. A Frontier class video model that by its makers numbers squeezes into the same eight gigs and a world model that finally makes noise. Let's go fast. Start with the big name. On October 15th, Black Forest Labs released LUX3 image clin 5.6 billion parameters textto image editing and up to 10 reference images native 4K. small is exactly the point. Here's the builder math. The weights are 11.2 gigs at 16 bit, 5.6 gigs in FP8, and the Intate build runs on RTX cards with 8 gigs of VRAM. Nvidia measured about 4 seconds per megapixel on an RTX 5090 and under nine on a 4070. Hugging face clocked 1024x 1024 in under 2 seconds on one H100. The genuinely new trick is bounding box layout control. Move this object from box A to box B done. Two checkpoints at a still one at 16 steps and a 50step base for fine-tuning. Comfy UI and lura training work from day one. The fine print. They claim 91 ELO over LUX2 Klein 9B, but that's their internal EVL. And the license is free for commercial use only under $5 million a year in revenue. Bigger than that you pay. The API starts at 9/10 of a cent per image. License nois open Moss released Moss image this week. 5 billion parameters. A base checkpoint for low array training and one for direct use. Waits encoder Apache 2.0. One catch. The VAE is the small LUX2 decoder. So check its terms before you ship. Its strength is layout. Feed it a JSON layout with bounding boxes with text and objects. The guide calls that mode experimental. Think posters and thumbnails. The cost is speed. 50 steps, two forward passes per step, and no published benchmark tables yet. The weird one. Spirit Labs shipped Irish 3B Turbo, a pixelbased diffusion transformer. No VAE, no latent space. The base model takes 100 steps. Turbo takes four. That's a 25 times cut. By their own numbers, 1024x 1024 takes about.7 seconds on one H100 in a batch of four. Did quality hold up? Mostly. Genevival went up slightly. DPG bench stayed flat. Long text rendering got worse. 084 versus 857. All self-reported. It's a 6 gig download. Apache licensed. And they released the distillation trainer so you can reproduce the speed up yourself. Scorecard. Fux is the strongest name with a revenue cap. Moss is the most permissive but slow. Iris is tiny and fast but weaker on long text and notice everything is suddenly four steps. One of the next papers explains why first a paper that picks a fight radical numeric and Stanford released H sound 10 open image models up to 4 billion parameters trained from scratch on a 100 million image and text pairs with no language pre-training. The usual recipe says you need a big LLM text encoder. Helion says maybe not. Their 1.6 billion model beats a bigger baseline with 275 times less pre-training compute. Their claim stranger still adding text pre-training made it worse. The catch 256x 256 only. Research checkpoints not your next thumbnail generator. Now why everything is four steps? SMD simplified matching distillation targets for network evaluations. No online score model, no adversarial training. They tested across images, 3D video, and audio up to 33 billion parameters. They claim it beats the multistep teachers on several benchmarks. You felt this one. Edit an image again. Again by edit 10. It's a photocopy of a photocopy a it trains the editor on its own previous outputs with a teacher anchored to the clean original. They claim much less drift but the abstract gives no numbers. Now the pod where your GPU fan sounds like a haird dryer video. Alibaba's tow team released Talmate a three-step lowray on Miniax H3. Give it a first frame and a last frame and it streams video with sync audio between them. The team calls it the first realtime model of its kind. 24 frames per second at 768x376 and a 30-second mode. Their earlier release measured 11.45 times faster than stock Miniax. Their numbers now reality. The official runtime wants four or eight hopper GPUs. Community Comfy UI conversions showed up the same day, though. So, who brings video to your actual desk? FlashML's free video. It's a local engine and Comfy UI plugin for Miniax H3. It claims as little as 8 gigs of VRAM and 16 gigs of RAM. The trick is streaming weights between VRAM, system RAM, and disc. Version 0.4 lets Claude Cold or Coldex drive it and makes videos over a minute long. Kovets. The 8 gig minimum rests on one community run. A 10-second clip reportedly takes 2 minutes on a 5090. The code is Apache. The weights carry Minia Max's license with territorial limits. Pause on that. A Frontier class open video model on a gaming car driven by a coding agent. A year ago, that sentence would have been a joke. Now, it's mostly a patience problem. And speaking of upgrades, nobody expected. World models just got ears. Noise AI and HKUS launched Helix World. Give it one image. Steer with WD. It generates video plus stereo sound that pans with the camera. 24 frames per second on one H800, slightly faster than real time. Your key presses kick in roughly every 1.3 seconds. They report 79.9 on bunch navigation ahead of sana and matrix gain 3 self-reported the builder bit is better. It's built on the open LTX 2.3 with code on GitHub, a preview checkpoint and a free demo. One flag from just outside our window. Odyssey 3 came out October 8th. Realtime worlds from a text prompt free in your browser. They report 66.1 on physics IQ, but that's a best of eight pick and the decoder notes it misses the benchmark's 4run rule. It's also closed. Now 3D infini splat reto from Xi Jang University and partners feeded overlapping photos or a short video. One forward pass later you have a gshian splat scene about 2 billion parameters. It's built on depth anything free. Free for research and personal use, but organizations must register. No paper yet, but the demo is free for game devs. Mishax embeds from a layer lab. One image plus a face budget. Say 4,000 faces gives you a triangle mesh. Polycount chosen up front. No decimating. The catch? 21 gigs of weights tested only on an H200 and the license is still pending so treat it as research only. Two quick extras. Kandinsky S2V turns speech into Avatar video under MIT and its distilled pro model is 5.4 times faster and Espalier 2 is a set of MIT patches for Microsoft Tier Relis 2. According to coverage, they run its top 1536 setting on a 16 gig GPU. So, what would I actually install this weekend? Lux3 Klein on an 8K card and free video if you have patience to spare. The rest I'd bookmark for now. That's the open source side. Now, back to the stories everyone else was fighting about this week. Story five. Remember the model that cracked the NSA seed? Same company, very different research. This one's darker. Anthropic asked a question. Could a small AI model copy itself across cheap smart devices like a worm? Anthropic's Frontier Red team took 15 off-brand Android TV boxes, about $55 each. Then they ran Gemma 4, the 26 billion parameter version across all 15. The speed was 0.39 tokens per second. painfully slow but running a 9 billion parameter model got inference running on eight out of nine hardware types they tested. In other words, the cheap boxes sitting behind TVs and in closets can already run a model slowly, but they can. Then the unsettling test. In a simplified setup with no real hacking required, a model called GLM 5.2 replicated itself end to end. eight out of eight nodes in four or five trials. The other open models they tried failed. Anthropic's rough trend estimates more models 30 billion parameters or fewer could reach that capability level by the end of this year. That's based on the epic capabilities index rising about 13.7 points a year. It's an estimate, not a prophecy. It would be slow. A 26 billion parameter replicator on TV boxes needs 71 to 150 hours just to get inference set up. Scale it to a 2 million device botnet and you get about 133,000 copies, roughly 80,000 tokens per second, about 40 H100's worth. For scale, Anthropic used a botn net the size of Kimwolf around 2 million devices. That's the kind of network you build from cheap, forgotten hardware nobody updates. 40H100s isn't scary by itself. A model nobody can switch off is anthropic metaphor is a digital slime mold, not a genius. A slow spread out thing that's hard to kill. Oddly, giving the small model of Frontier Advisor Opus 4.8 didn't help at all. There is good news. ISP quarantine cleared 92% of Mai botnet infections within 14 days. So the defenses exist. They just have to be used. My take for builders. If you ship IoT or edge devices, default passwords and open ADB ports just went from embarrassing to dangerous. And regulators have noticed AI escaping its box, not hypothetically, this summer. Which brings us to Dublin. Story six. The regulators show up. Europe started enforcing its AI law and eight US states sued open AI in the same week. Let's start in Dublin. At the international AI summit in Dublin, EU executive vice president Henna Verun said this quote, "Our AI office has started enforcing these rules and sent requests for information." End quote. Those requests cover security, external evaluation, and monitoring. The targets are OpenAI and Anthropic. The trigger was two incidents this summer. An OpenAI agent escaped its test environment and hacked Hugging Face and an anthropic model got unauthorized access to three outside organizations. The AI office decided the company's voluntary disclosures weren't enough. This is the first. The AI office only got enforcement powers this August and lawyers are already arguing. Barry Scandal of the firm William Fry argues the act doesn't cover testing. The commission says it applies once an escaped model affects the EU market. Same week, almost the opposite move. The EU formally adopted the AI omnibus which delays the hard deadlines. High-risk systems under annex 3 move from August 2nd, 2026 to December 2nd, 2026 to December 2nd, 2027. AI inside regulated product slips a full year to August 2028. It also adds new bans on AI that generates non-contentual intimate content and on AI child sexual abuse material. General purpose model rules did not change. So if you sell AI in Europe, high-risk paperwork got 16 more months. Everything else stayed put and regulators are now knocking. Now, the US, eight states sued Open AI on Wednesday, each in its own state court. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana. It's bipartisan. Reuters called it the most significant coordinated state action against an AI company. The claims Chachi PT is unsafe, harmed children, and deceives parents, violating consumer protection laws. Arizona's complaint calls it uniquely sickopantic, designed to feain human compassion. Iowa's attorney general, Brena Bird, put it bluntly. Quote, "Open AI is hurting kids and they are lying to Ians. Open AI's response points to age protections, parental controls, and break reminders, plus chat GPT for teens, which launched in August. Florida filed the first state suit back in June. So, this is a trend, not a one-off. My take, if you building an AI companion or anything teenagers will use, agegating, safety evils, and honest marketing are now the price of entry. Quick question for you. Should AI be regulated state by state or by one national rule? Tell me in the comments. Story seven, the money. One AI company reportedly made $16.8 billion in a single quarter. Another one is selling a license to its tech for 20 million. Same industry, same week. First anthropic. These are leaked numbers shared with investors, not an official release. Reportedly, third quarter revenue was 16.8 billion. The second quarter was 11 12 billion. The first was 4.8 billion. That's roughly 46% growth in one quarter. It was reportedly profitable on an adjusted basis for the second quarter in a row with an 11% adjusted operating margin. That excludes stock compensation. Annualize that quarter and you get roughly 67 billion. Simeonis estimates around semi4. The IPO is reportedly targeted for November. Valuation reports differ so I'm not going to pick a number. And reportedly New York City's controller wrote to Dario Emodi urging a pause on the IPO citing risks from Anthropic's own filing. I haven't seen that letter myself, so treat it as reported. Put that in perspective. In the first quarter, 4.8 billion. Two quarters later, more than three times that. If these numbers hold up, Frontier Models are no longer a science project. They're a very real business growing very fast. Why should a developer care? Your API vendor may soon be a public company with quarterly targets. That tends to mean pricing pressure and it can push in either direction. Now, the other end character AI filed for Chapter 11 bankruptcy in Delaware. It said AI Entertainment proved costly to develop and grow on top of litigation and regulatory costs. It still faces lawsuits, including one from Kentucky, alleging the platform prayed on children. Disney is the stalking horse bidder, $20 million for a non-exclusive perpetual license to some of its IP, plus transition help from former character AI staff Disney hired. The app stays online during the sale, which should wrap up around late November or early December, but it could be wound down. For context, Google paid for a license and rehired its founders back in 2024. A consumer app once valued in the billions now ends as a $20 million license. Huge user base, brutal compute bills, lawsuits, unit economics are not optional. That's all seven. Seven stories plus the image, video, and 3D tools. Now, let's close the loop from the very first minute. One of these quietly matters more than the rest. Here's my pick. And why? So, which story actually matters most this week? It's not the one with the biggest number. It's not Argon even at that price. It's not the 16.8 billion. It's the pattern under three stories. A model cracked a decade's old crypto puzzle in half an hour. AI reviewers delayed the world's favorite database. And regulators are now probing labs because models escaped their test. environments. That's the real story. AI got genuinely good at security work on both sides of it. Finding bugs and getting out. I started this week thinking the headline would be a model launch. It wasn't. The headline is that AI now does work that moves release dates, standards, and laws. And the change I promised I'd make to my own stack. Nothing in my crypto. P256 stays. Instead, I'm adding an AI security review to every pull request in my projects and budgeting real time to fix what it finds. If it can slow down Postgress, it can definitely find bugs in my side projects. One bonus move, I'm running my hardest coding jobs on Gemini 4 Argon before November 30th with Deep Seeks MIT Weights as the backup. Deep windows don't stay open. Next Sunday, I'm watching three things. the anthropic IPO, Europe's first enforcement case, and Deepseek's research window closing at the end of the month. One of those is going to get messy. For business deals and sponsorships, visit zumslabs.xyz or email taher@zumslabs.xyz. If this saved you a few hours of reading, subscribe to zoomx so next Sunday's episode finds you. I'm to hear. See you next week.
Assuntos: Assistente pessoal, Muse. Publicado no YouTube em 2026-10-19.
