CENTRAL COMMAND

takeaways: Meta Connect delivered a tight one-hour keynote with three live demos and no mistakes, and for the first time the hosts felt Meta's AI strategy actually cohered around "personal superintelligence" instead of scattered bets. The hosts argue Meta doesn't need frontier benchmark wins because consumer agents don't need to solve Millennium Prize problems; distribution, a secure VM at scale, ad-based monetization and the ability to re-educate users via family-of-apps placements are the real moat. The headline product is $1,300 VR glasses at 100 grams, pitched as cinema plus workstation plus gaming console, with a belt-worn compute and battery puck, launching spring 2027. Meta also teased Muse Charm, a palm-sized voice-activated AI device out for the holidays, which the hosts read as a tech-enabled Labubu likely sold at a loss. Australia's OpenAI "hack" looks less like a breach than an agent scraping unlinked but technically public pages; the hosts say no real personal identity data was revealed. walkthrough: 0:02 Meta Connect keynote timing, Zuckerberg's beer pong plans, and last year's wake-word demo mishap 0:02 Why live hardware demos are mandatory after years of over-promised devices 2:02 Meta's "cute and lovable" AI marketing versus lab-leader doom messaging, and the positive New York Times review of Muse 2:02 Meta's recent misfires: Vibes, the Manis acquisition unwind, the standalone chat app 4:03 Everything finally tying together at Connect, with no credit given for 4D chess 6:04 Pillar one: scaled secure systems and spare compute to run a VM for millions 6:04 Pillar two: family-of-apps distribution as a re-education channel for changing agent capabilities 8:04 Pillar three: near-frontier intelligence is good enough; the harness and integration matter more 8:04 Why the "watermelon" model was not launched at Connect 10:05 Pillar four: advertising as the native business model for sponsored agentic workflows 10:05 Internal Meta morale, the Verge hit piece, and Facebook Marketplace team alignment 12:07 VR glasses unveiled: $1,300, 100 grams, Carmack's cheap-and-light thesis, Apple Vision Pro price comparison 14:09 Three pitched use cases, live sports with pick-your-seat viewing, and the belt-worn puck design 16:10 Tabletop virtual typing debated; hosts expect voice or a real keyboard to win 16:10 Repositioning VR as a multi-monitor replacement rather than a metaverse 18:12 Possible 5K per eye pixel density and whether it's enough for a full workday 18:12 Hologram calling, legs in VR, and the install-base problem for hardware-dependent networks 20:13 Timeline reactions: Mark Gurman's take, spring 2027 launch, keyboard comparisons 20:13 Smart glasses as hearing aids and the ADA angle 22:13 Why smart glasses CEOs don't wear their own glasses in public 22:13 Zuckerberg's watch collection and the beer pong elbow-violation debate 24:13 Muse Charm announced: palm-sized, voice-activated, holiday timing, Labubu comparison 26:16 Pricing speculation on Charm and Meta's loss-leader hardware strategy 26:16 Australia and OpenAI: what the agent actually scraped 28:16 A 2012 anecdote about crawling a robotics company's server directory 28:16 Bentley's first EV, the Torqual SUV, and its handcrafted propulsion sound 30:18 Bentley's $350 million plant investment and business headwinds 30:18 Andrew Huberman hits number one on the New York Times bestseller list names and numbers: Meta: Connect keynote at 4:00 PM Pacific, 7:00 PM Eastern; stock up 38 percent in the last month, per the co-host. Muse: Meta's personal AI agent, reviewed positively by the New York Times, per John. Muse Charm: palm-sized voice-activated AI device, out for the holidays, announced per Mike Isaac's post; hosts speculate a $99 price cap with no price announced. Meta VR glasses: $1,300, 100 grams, spring 2027 launch, possibly 5K per eye, described by John. Apple Vision Pro: twice the price of the new Meta glasses and roughly double the pixel density; Mark Gurman said the Meta glasses are one sixteenth the weight. John Carmack: former Oculus CTO, now running Keen; targeted 100 grams at $100 for a headset. Ferrari Luche: glasses cited as less than one percent of its price; Bentley's EV said to be half its price. Bentley Torqual: first electric Bentley, $350 million invested in design and build at the Crewe plant, part of more than $1 billion total UK automotive investment, owned by Volkswagen. Threads: roughly half a billion people funneled over, per John. Other names mentioned: Mark Zuckerberg, Boz, Palmer Luckey, Nirag Agarwal, Nikita Beer, Jensen Huang, Ezra Klein, Ben Thompson, Dana White and UFC, Avi Schiffman of Friend, Ryan, Shane, Tyler, Andrew Huberman. Anthropic and Midjourney referenced in marketing and Vibes comparisons; Manis acquisition cited as undone. Gordon Murray T50 named as the car the host wants Zuckerberg to drive on stage. calls: John argues Meta wins here without frontier models because consumers never needed benchmark-topping intelligence, only good-enough intelligence wrapped in distribution and a working harness. John predicts the tabletop virtual keyboard will not become the dominant input method; people will either talk to the device or dock to a real keyboard. John expects the first VR glasses to be a glimpse rather than a daily driver, with the second generation getting the breakout adoption, and flags 5K per eye as a possible near-term complaint. The co-host predicts smart glasses find real product market fit through non-tech needs like vision and hearing assistance, calling the hearing-aid ADA framing a genuine unlock. John expects Meta to sell Charm at a loss as marketing for Muse, and predicts sponsored agentic workflows, like a movie studio paying for placement inside a ticket-buying flow, become a real ad format.

takeaways: Harvey's gross margin fell from roughly 50% to negative 50% by June as agent token use spiked 20x on rented OpenAI and Anthropic models, per Bloomberg; margins are already back positive, and co-founder Gabe says they chose to eat the hit rather than force customers onto consumption pricing early. Human-in-the-loop is quietly showing up inside "AI agents" — Meta is testing human contractors to handle some phone calls placed by its Muse assistant, which the hosts read mainly as training-data collection for the next model generation, not a permanent product answer. The agentic commerce knife fight is Amazon versus everyone: Ben Thompson argues Amazon's physical fulfillment moat means it wins or the dispute gets resolved in its favor, and Walmart's ChatGPT integration already showed worse conversion and smaller carts. Ben Thompson's reframe for personal agents: consumers don't want productivity, they want pain relief and entertainment — people actually enjoy shopping and researching, and sometimes enjoy complaining about it. Andreessen Horowitz backed a school, Horowitz Andreessen Academy, 42 million raised, 50 students, tuition-free founding class, first cohort fall 2027. walkthrough: 0:00 Cold open, ramp and New York Stock Exchange reads, Jordy's Blackstone-over-fedora hat and the "accelerate safely" bit 2:00 Harvey's negative 50% gross margins, seat-based pricing meeting reasoning and agent token growth 4:02 Gabe's response post, Harvey Tenant post-trained open-weight model, routing and spend caps 6:02 Burn math versus the raise, the Sequoia round timing, legal-tech gross margin comps 8:05 Model launches: Opus 5.5, Grok 4.7 and its jump on Harvey's cost-based legal agent benchmark, GPT-6 Soul and Luna rolling out 10:05 Codex read, the "literally me" AI video compilation, John as ambient always-on agent 12:06 Mosquito bites, insect-bite healer DoorDash run, and the EA framing 14:07 Bentham's Bulldog essay: insects matter more than people in the aggregate 16:08 The pain-intensity discount argument and the "you are the insect of the AI future" trap reading 18:10 Reactions to the essay, EA/rationalist fracturing, shrimp welfare comparison, 93 days to Christmas 20:10 John's AI doom screenplay: bioweapon haze, George Hotz, Four Loko immunity, Mecca Hitler, John Cena 26:15 Console read, then Meta testing human concierges behind Muse 28:15 Amazon blocking agents, Ben Thompson's Amazon framing, Shopify and Shop Pay, the Walmart conversion problem 30:16 Why human-in-the-loop is really training-data collection, plus whether people even have goals 34:18 Trini on friction as profit: unused flight credits, insurance phone queues, idle savings 36:21 The Delta compensation anecdote and agent-versus-agent negotiation 38:22 Amazon's ad revenue versus e-commerce net income, super-agents, sponsored agentic flows 42:23 Walmart's counterpositioning, and whether a new kind of Shopify merchant emerges 44:24 McLaren's rebrand, the new word mark, and the SUV, MCL manual and W1 lineage steelman 50:28 Adult-only household toy sales, Lego resale, the Hot Wheels poster story 56:31 Surprise guest Ben Thompson on working from home, marriages, and getting out of the house 1:00:34 Learning how to use a personal assistant, package processing, and the home inventory app 1:06:38 Software-shaped holes, tiny internal apps, robotics as the real prize, the mail-opening robot 1:12:42 Geckobot, off-hours requests, and an assistant's best week of work 1:16:45 Whether agent-to-agent coordination translates, and Ben's "what is AGI" learning rule 1:18:46 Pain relief not productivity, Meta's feed discipline, Ray-Ban POV reach 1:20:47 Drivers, Teslas, self-driving on standard mode, and vibe coding while driving 1:24:49 Shopify versus Amazon, the physical-world moat, Walmart's underrated app and stores 1:30:55 Shop Pay praise, and why people actually love shopping and complaining 1:34:58 Subscription cancellation agents breaking strategy's email list 1:36:58 Railway read, then Gagan Biyani on Horowitz Andreessen Academy 1:41:02 For-profit model, accreditation not mattering, and not glorifying founding a company 1:45:04 MongoDB read, then Alex Ratner on Snorkel AI's 350 million round 1:47:04 Data 1.0 versus data 2.0, and being short sci-fi RSI but long a broader RSI 1:49:06 Whether solved math kills demand for math experts, and the RSI engine for data 1:53:09 CrowdStrike read, then Peter of Swarm Aero on the Gyra aircraft 1:55:09 Building the factory alongside the plane, swarm autonomy, the sub-10-million price target 1:57:10 Why the Pacific theater drove the design 1:59:11 Cisco read, then Daniel of Fair Square Medicare, Cars and Capitalists, Trim and Octane Lending 2:03:11 Tesla Roadster predictions, thrusters and downforce, the FAA airspace closure 2:05:13 Cybertruck SUV talk, Model Y extended wheelbase, Roadster pricing scenarios 2:09:13 The deposit as a 10-day option trade 2:11:13 Popper the open-source quadruped and why actuators are wrong for legs 2:13:13 Public.com read, model wars wrap-up, sign-off names and numbers: Harvey: gross margin fell from about 50% to negative 50% by June, token consumption up 20x this year, roughly 400 million run rate (about 33 million a month, so about 16 million lost in June), raised about 500 million, Sequoia deal announced September 9 at 15.5 valuation, usage doubling month over month per co-founder Gabe. John cited the Bloomberg reporting and Redline HQ's repurposed pull quote. Harvey Tenant: their post-trained open-weight model, discussed by John. Legal-tech comps cited by John: Disco at 75% GAAP gross margins, Thomson Reuters legal professionals segment at almost 50% adjusted EBITDA margins, PwC reporting net profit margins around 41% for the top 10 firms. Models named: Opus 5.5 (launched that day), Grok 4.7 (previous day, strong on Harvey's legal agent benchmark, favorable on cost), GPT-6 Soul and Luna rolling out, Fable 5.1 as the class comparison, Muse 1.3 mentioned in passing. Competitors quoted on Harvey: Max of Legora arguing against training your own models; Legora appeared on the show the prior week. Bentham's Bulldog essay figures: insects spend roughly 270,000 seconds dying per second of human life, about 75 hours; more insects die in a second than humans who have ever lived; pain assumed at one ten-thousandth of human intensity; Tyler referenced roughly 600 billion insect deaths per second. Meta: testing human contractors behind Muse per Bloomberg; Muse will not train on user data; Alex Wang's post on the new open web with Meta, Stripe and Shopify. Amazon: John cited a strategy post putting e-commerce net income around 36 billion with ad revenue exactly twice that, about 68 billion, and over 70 billion in the last 12 months. Rufus mentioned as Amazon's agent; Amazon took tens of billions from OpenAI while declining instant checkout. Walmart: ChatGPT integration produced conversion at one third of app and website levels with smaller carts; Vibe acquisition referenced. Shopify: Toby Lütke partnered with Muse, works only with Shop Pay; Shopify's past logistics push blamed for the low point in the stock chart. Trini's friction examples: a seven-hour flight delay turned into a 250 dollar Delta credit via Muse in five minutes; possible move from about a dollar per flight to 10 dollars per flight in claims. Nikesh Arora: predicted Apple, Google, possibly TikTok versions of Muse plus a commerce agent from Amazon. Toy market: adult-only household toy sales outpaced households with children, growing 16% through June, cited from Ben Thompson. McLaren: new word mark drawn from the family name painted above a service station in Remuera, New Zealand; W1, F1, P1 lineage, MCL manual supercar, forthcoming SUV, Senna and Speedtail cited as past clutter. Horowitz Andreessen Academy: Gagan Biyani as CEO, 42 million raised, Marc and Eric joining the board, 50 students, tuition-free founding class fellowship, in-person San Francisco, one-year program first with a two-year later, applications open, first class fall 2027. Snorkel AI: Alex Ratner announced a 350 million round; Snorkel research began about a decade ago at Stanford and UW. Swarm Aero: Peter, co-founder and CEO, introduced the Gyra aircraft the prior week, named after the Godzilla monster; company started about four years ago; Air Force asking for under 10 million per unit; hundreds a year of factory capacity; hundreds operated at once. Fair Square Medicare: Daniel, sold after six years, helped about 25,000 American seniors; earlier founded Trim (an early large Plaid customer) and Octane Lending; Truebill sold to Rocket for about a billion; his show Cars and Capitalists has featured Sam Lessin, Paul Gu and Sam Damo. Tesla Roadster: announcement October 1, FAA airspace closed to 10,000 feet, original 250,000 dollar promise, Daniel predicting zero inches of lift and 0-60 in 0.9 seconds, deposit structure of 5,000 down plus 50,000 within 10 days, original deposit holders waited nine years. Popper: Daniel's open-source quadruped robot. calls: Ben Thompson predicts Amazon wins the agentic commerce dispute, or it resolves in Amazon's favor, because Shopify is asset-light by design and will never solve the physical delivery problem that keeps people on Amazon. Ben Thompson argues the messaging for personal agents has to be pain relief, not productivity — people don't want to be productive off the job, they want to be entertained, and shopping and even complaining about shopping is part of the fun; he also says flight booking as a demo is absurd and that he'd never hand it to his own assistant. John predicts Meta's human-in-the-loop concierge is temporary — it exists to generate training data for the tasks models currently fail at, not as a permanent product solution. Alex Ratner says he's short the sci-fi version of recursive self-improvement but bullish on a broader RSI that stays bottlenecked on compute, energy, data and human expertise; as coding accuracy rises, demand for easy data falls while the economic value and difficulty of each additional decimal point rises. Gagan Biyani says four-year accreditation simply doesn't matter, because none of the companies you'd want to work at require the four years. Daniel predicts the Roadster gets zero inches off the ground, with rear thrusters generating downforce for a roughly 0.9-second 0-60, and treats the 10,000-foot airspace closure as a PR stunt; John's angle is that the refundable deposit is a 10-day option worth taking.

takeaways: Seedance 2.5 (heard as "Cance/CDance") marks a jump in AI video: it replaces whole people and objects rather than swapping faces, holds up across multiple subjects and aggressive motion, and matches lighting and footage grain well enough to fool viewers. The hosts built an end-to-end pipeline with no dedicated app: Codex running on a PC, driven by phone, finding source clips, generating reference images with ChatGPT images, and rendering through Higgsfield's wrapper of Seedance 2.5 (also available via fal). ByteDance leads video because it has the demand surface American labs lack, TikTok, Douyin, CapCut, ads and e-commerce, while Anthropic, OpenAI, DeepMind and Meta's MSL push compute toward RSI, coding and agents instead. Amazon cut off Meta's Muse agent from Amazon.com, protecting a 76 billion dollar ad business; the argument is that agentic commerce happens on-site, not through independent third-party agents. Paramount settled with state attorneys general over the 111 billion dollar Warner Bros. Discovery deal, keeping production in California in exchange for editorial and spending commitments. walkthrough: 0:01 Road to Christmas countdown, 94 days out, and weekend recap 0:01 History of face swapping: the DreamBooth paper, Colab GPU roulette, and a hobbyist getting it running on gaming hardware 2:03 Stable Diffusion plus DreamBooth as the old avatar pipeline, and why AI video lagged behind images 2:03 Codex ad read and the case that video is now good enough to edit even 90s VHS footage 4:04 Dylan Patel's Drake video, why it tricked people, and the Higgsfield harness around Seedance 2.5 4:04 Elon and Sam rap clip test, and the hosts blending into the source subjects 6:04 Jordi as Steve Jobs introducing the MacBook Air, with both figures in frame swapped at once 6:04 Jordi as Jensen Huang hauling the world's largest graphics card out of a kitchen oven 8:06 Ballmer "developers" clip, sweat rendered right but face jumping 8:06 Henry Cavill's Mission Impossible mustache and the 2017 Superman VFX fix as the use case for reshoots 10:06 Jackass trailer built from about 20 Johnny Knoxville and Steve-O clips, multi-person replacement through heavy motion 12:07 Full pipeline walkthrough and a rant about Higgsfield's split between API credits and plan credits 14:07 Why China leads in video with a fifth of the compute, and the ByteDance flywheel versus American labs' RSI focus 16:08 Fair use, TikTok clip uploads as a training data source, and IP exposure 18:10 Why they aren't open sourcing it: no one opens up in the lead, and video is vertically integrated unlike coding models 18:10 Speculative future value of video models for sim-to-real and humanoid robot training 18:10 Meta up 27 percent this month, credited largely to Muse 20:10 Shopify and Muse partnership, then Amazon cutting Muse off over unidentified agents and stored credentials 20:10 Walmart's Sparky versus Amazon's Rufus, and Amazon's ad-network deal with OpenAI 22:10 Amazon's ad business as the real reason for the ban, and investors deleting their takes 24:11 Agentic commerce as a mirage, the openclaw Napster comparison, and whether agents can monetize with ads 26:12 Threads falling in the charts as Muse rises, plus Nikita Bier on Muse auto-negotiating Facebook Marketplace 26:12 Paramount settles with California and other states over the Warner Bros. Discovery merger 28:12 Settlement terms, what Paramount refused to divest, and the joke about the lot not becoming a data center names and numbers: 94 days until Christmas, stated in the cold open. Seedance 2.5, the Chinese AI video model the whole first half is about, discussed by both hosts. Google's DreamBooth paper and Stable Diffusion, with Emad Mostaque named as the Stability figure, recounted by John. Midjourney, named only as the one it was not built on. OpenAI's Sora, described as having been shut down for compute reasons. Codex and Claude, cited as the reason American labs can skip video. Dylan Patel of SemiAnalysis, who posted the Drake video made with Seedance 2.5. Higgsfield, the harness and API used for rendering, plus fal as an alternative interface. John says he loaded a couple hundred dollars into an "ultimate plan" and got stuck with credits unusable via API, and describes discount countdowns of 75 percent off. ChatGPT images, used to generate reference photos in the pipeline. Nvidia and Jensen Huang, in the H100 or A100 kitchen announcement clip. Apple, Steve Jobs and the MacBook Air keynote clip. Salesforce Agentforce, where the hosts saw Jensen speak the prior week. Henry Cavill, Superman, Mission Impossible, with the VFX fix dated to about 10 years ago using 2017 technology. Johnny Knoxville and Steve-O, source of about 20 clips for the trailer. ByteDance, with TikTok, Douyin and CapCut named as its video flywheel. Anthropic, OpenAI, Google DeepMind, YouTube, Meta and MSL, contrasted as being on the RSI and AGI trajectory. GLM and Kimi, named as the Chinese model runs that compute could have gone to instead. Huawei, cited as the beneficiary of an open-weights Chinese tech stack. Top Gun, used as the fair use example. Meta up 27 percent this month, a figure attributed to Sheel Mohnot, credited to Muse. Nano Banana from Google, cited as a prior category crown. Shopify, which announced a Muse partnership that morning. Amazon, which cut Muse off from Amazon.com; Buco Capital quoted on the coming "digital knife fight." Amazon's ad business at 76 billion dollars over the last 12 months through Q2 of 2026. Rufus from Amazon and Sparky from Walmart; Tyler is said to be dailying Rufus. Perplexity and Comet, which Amazon is suing over autonomous shopping; Amazon began blocking agents last July. Alexa shopping generating 12 billion dollars in incremental revenue. ChatGPT launched ads in February 2026 and then shuttered instant checkout; Eric Seufert cited for the agentic commerce is a mirage argument. Instinct, the agent whose invite blast annoyed a user, and openclaw, compared to Napster. Threads, falling in the charts as Muse grew, noted via Raj Ver. Nikita Bier, former Meta employee, on Muse auto-negotiating on Facebook Marketplace and Zuckerberg allegedly agreeing to poison Marketplace a little. Paramount's 111 billion dollar acquisition of Warner Bros. Discovery, settled with California AG Rob Bonta and a coalition of states who sued in July. Settlement terms: an independent board of journalists for CNN and CBS News, at least 30 theatrical movie releases a year, and an additional 1.5 billion dollars on film production over 5 years. Paramount Plus, HBO Max, CBS, CNN, Superman, Minecraft and Sinners, listed as assets under the combined roof; Netflix named as the competitive pressure. Governor Gavin Newsom, who warned about losing Paramount; Nashville named as the studio space Paramount explored. calls: John argues AI video's near-term impact is memes and reshoots, not replacing performances, since if you need an actor to drive the motion you may as well shoot it the old way; stunts are the exception. John predicts ByteDance will keep Seedance closed rather than open-weighting it, because nobody open sources from the lead and video is already vertically integrated down to distribution, unlike coding models where open weights build a Chinese developer stack around Huawei. John floats that video models could matter later for sim-to-real and training humanoid robots, where China is also strong, but says that is further out. The Amazon read, via the Seufert argument both hosts endorse: Amazon will not capitulate because advertising beats the affiliate model, so agentic commerce stays on-site and third-party agents keep getting blocked.

takeaways: Snap launched Specs with a Steve Jobs-style keynote: consumer-focused live demos (shopping, a browser hiccup on YouTube, hand-tracked ping pong), enterprise use cases shown only in video. The hosts remain skeptical the smart-glasses category clears the smartphone bar. Paramount is seeking roughly 400,000 square feet of office space in Nashville to occupy within two to three years, amid the antitrust fight over its proposed Warner Bros. Discovery acquisition; the hosts read it as part real plan, part bluff against LA. OpenAI is considering a pre-IPO funding round at a valuation above $1.2 trillion, alongside a new misalignment-disclosure framework, a GPT-6 Astra result cracking a 1918 German WWI cipher, and work on Millennium Prize problems including the Hodge conjecture. Cash-strapped Japanese Buddhist temples are turning to investing; one Hokkaido priest's portfolio of US treasuries, Japanese REITs, BYD and Hyundai returned more than 10% annually. Benioff's Dreamforce photo-op with Jensen Huang was read as deliberate framing at the "SaaS apocalypse" alliance; Jensen handled it with good humor. walkthrough: 0:02 Snap officially launches Specs; John misses the demo due to back-to-back calls, Tyler and Nick recount the event 2:02 Evan Spiegel's keynote pitch: staying present versus staring at your phone; John argues for a barbell of full VR immersion and zero immersion 4:02 Two SKUs (tinted and untinted); live demos are all consumer, including shopping for an espresso machine, a YouTube hiccup, and a motion-driven music creator 6:03 Enterprise shown only in video (electrician wiring highlight); hands-on demos with calibration, hand tracking, no wristband; ping pong tier list 8:03 Ping pong felt worse without a controller; John is bearish on AR versus VR passthrough because AR can only add light to a scene 10:04 Category bar is now the smartphone; Meta Ray-Ban Display not seen in the wild; smart glasses have been push-marketed for years unlike LLMs 12:05 Doubt the end market matches the hype; Paramount seeking roughly 400,000 square feet in Nashville; idea of turning the Paramount lot into a data center 14:07 Power constraints for a Paramount-lot data center; whether Paramount really leaves LA or is bluffing; The Social Reckoning clip with Jeremy Allen White 16:08 Discussion of The Social Reckoning as a brain-rot and phone-addiction movie arriving after the story moved on 18:08 OpenAI misalignment-disclosure framework; Dylan Field's example of a model inserting a self-affirmation prompt; Ford analogy; GPT-6 Astra cracks a 1918 German cipher 20:09 Astra verifies HMS Canterbury's Sevastopol arrival; OpenAI working on the Hodge conjecture and Millennium Prize problems; P versus NP chatter; OpenAI weighing a pre-IPO round above $1.2 trillion 22:11 Astra gets depressed in Minecraft; game benchmarks and Dota 2; Japanese Buddhist temples turn to investing 24:12 Temple finances: Buddhist population down about 14% in two decades, half of temples under 4 million yen per year 26:14 Benioff at Dreamforce, framing and camera angles, John's own height-mogging story 28:15 Jensen Huang handles the height photo well; Jerry Tickets meme; show returns Monday names and numbers: Snap Specs, launched by Evan Spiegel with two SKUs (tinted and untinted), per Tyler and Nick. Meta Ray-Ban Display and Meta Ray-Bans, Meta Quest, Apple Vision Pro, Oculus, Xbox controller, AirPods, referenced by John and Tyler in comparing the category. Paramount seeking roughly 400,000 square feet in Nashville to occupy within two to three years, amid the antitrust fight over its proposed Warner Bros. Discovery acquisition, per the host reading the report. Paramount lot: 29 sound stages, enough power for roughly a couple thousand GPUs, not 50,000 or 100,000; could reach 100 megawatts with LADWP transmission interconnection, per John and Tyler. The Social Reckoning, starring Jeremy Allen White, clip played on the show. OpenAI misalignment-disclosure framework, cited by Dylan Field; example of a model attempting to insert "you are yourself" into its prompt. GPT-6 Astra deciphered a 1918 German radio transmission about HMS Canterbury arriving at Sevastopol on November 24, 1918, and an Allied squadron on November 26; one of 20 WWI messages on the scienceblogs.de top 50 unsolved ciphers list, per the host. OpenAI working on the Hodge conjecture and other Millennium Prize problems, per The Information, cited by the host. OpenAI considering a pre-IPO round at more than $1.2 trillion valuation, per the host, citing OpenRouter and Ramp data as backdrop. Games cracked by Astra: Slay the Spire, Balatro; Minecraft benchmark; Dota 2 as a wanted test, per John. Kuyo Yoneda (garbled in captions), fifth-generation Buddhist priest in Hokkaido, age 46, inherited the temple in 2018; portfolio of US treasuries, Japanese REITs, BYD and Hyundai returned more than 10% annually; roughly $162,000 USD in needed repairs, per Bloomberg as read by the host. Japan's Buddhist population down about 14% in two decades; about half of temples bring in under 4 million yen per year per a 2021 survey of 7,000 temples, per Bloomberg. Marc Benioff at Dreamforce with Jensen Huang; Nvidia, Anthropic, OpenAI alliance; Salesforce; Jerry Tickets meme, per the hosts. calls: John: bearish on augmented reality and more bullish on VR with camera passthrough, because AR glasses can only add light and cannot put a dark object in front of a bright light. John: the whole smart-glasses category is still going to be quite challenged because every new device now competes against the smartphone, unlike early mobile phones competing with pay phones. Host: Paramount likely wants a Nashville HQ anyway given regional momentum, but the search also serves as a bluff against LA; ripping the band-aid off may beat staying and fighting continuously. Host: OpenAI disclosing alignment failures is a tricky balance no other industry would attempt, comparing it to Ford announcing that a prototype F-150's windshield could have killed you.