← Tech Brew Ride Home
Tech Brew Ride Home
Manus, The Hands Of Fate
00:20:53
More on Manus, who they are, how they got positioned to sell, and more on what Meta wants to do with them. Open AI is paying employees more than anyone in history. The TriFold phone is a bit a dud. And, of course, the weekend longreads suggestion.
Loading summary
Transcript4 lines
- [00:02]
A
Ford BlueCruise Hands Free highway driving takes the work out of being behind the wheel, allowing you to relax and reconnect while also staying in control. Enjoy the drive in BlueCruise enabled vehicles like the F150, Explorer and Mustang Mach E available feature on equipped vehicles Terms apply. Does not replace safe driving. See Ford.com BlueCruise for more details.
- [00:34]
B
Welcome to the final Tech Brew ride home of 2025. I'm Brian McCullough. Today, more on Manus, who they are, how they got positioned to sell, and more on what Meta wants to do with them. OpenAI is paying employees more than anyone in history, the Tri fold phone is a bit of a dud, and of course the weekend long read suggestion. Here's what you missed today in the world of tech. If you are looking for enterprise grade identity automation minus the enterprise grade baggage, aka having your users log on 500 times, yes, ID delivers advanced IAM automation without moving teams onto a legacy identity provider. Whether you use Google Workspace, Microsoft 365 or Okta, Yeshid integrates directly. No rebuilds or RIP and replaces are required. Yes, ID helps IT and security teams reduce risk, not just tickets. So I bet IT teams everywhere just breathe a collective sigh of relief. Every access, change, review and approval is tracked and exportable, helping security teams effortlessly demonstrate compliance with SOC2, ISO or HIPAA. IT and security teams can spot risk before it becomes a finding. Learn more@yeshid.com Techbrew that's Y E S H ID.com Techbrew we didn't talk much about Manus really, until yesterday's acquisition news, so I wanted to share with you this New York Times piece looking at how Manus distanced itself from its original Chinese roots to court U.S. investors. A source says Meta's $2.5 billion deal includes a $500 million retention pool for Manus employees. Before being approached by Meta, Manus executives and the startup's investors weighed staying independent and raising substantially more familiar with the matter said. Like many successful AI startups, Manas faced a difficult reality. Without a platform partner such as Meta, it would be challenging and costly to reach global scale, and raising additional capital could make the company too expensive for potential suitors. Meta spokesman Andy Stone said that there would be no continuing Chinese ownership interest in Manus after the transaction, and that the startup would discontinue its services and operations in China. Meta's deal apparently surprised some officials in Beijing, some of whom disliked the agreement because they considered Manus an example of China. China's AI power. People familiar with the officials thinking said they believed that the sale would give the US access to technology developed by Chinese engineers and encourage other startups to pursue a similar funding path, the people said. But Beijing appears to have few tools to influence the deal, given Manus foothold in Singapore. In Washington, the reaction was muted, a signal that Manus moves to avoid violating U.S. rules that restrict outbound investments in key technologies eased concern about its China ties. The indicators on this one seem to be all pointing, at least on the surface, in the right direction, said Chris McGuire, who worked on technology export controls in the Biden administration and is now a senior fellow at the Council on Foreign Relations. He views the deal as evidence that export and investment restrictions work and could squeeze other Chinese AI companies, pushing them to do more deals with US Partners. Manus core leaders are two young Chinese entrepreneurs, Zhao Hong and Ji Yi Chao, also known as Red and Peak. Zhao set up Manus's parent, Butterfly Effect in 2022 and launched a ChatGPT power application for browsers called Monica. The Butterfly Effect apps targeted markets outside China, mainly North America, Japan and South Korea, Zhao said in a recent podcast. Butterfly Effect had offices in Beijing and Wuhan. In October of 2024, inspired by the San Francisco based AI coding tool Cursor, Butterfly Effect started developing Manus that use several American AI models that aren't available in China. The project's name, Manus, came from the Massachusetts Institute of Technology's Latin moto mens et Manus, meaning mind in hand. Manus in March released a demo of its AI agent, which is designed to handle more complex tasks than a typical chatbot, such as producing a 100 page research report, generating a slideshow or building a website. At the time, most of its researchers and engineers were based in China. Some in China called it another deep seek moment. An invitation code that gave people early access to the tool was resold for more than $1,000 on social media and e commerce sites. Earlier this year, several local governments in China approached Manus and offered to invest in the startup, but its founders turned them down, according to people familiar with the matter. They were concerned that such connections could cause scrutiny in the west and create challenges for its global business, the people said. Manus also shelved a plan to join with Alibaba to introduce a Chinese version of the tool that had been announced in March, people familiar with the plan said. Around that time, Manus secured the funding that drew scrutiny in the US it soon moved its headquarters to Singapore and laid off some employees in China. Meta began negotiations for an acquisition in mid December, and Mark Zuckerberg wanted to reach an agreement by the end of the year, people familiar with the matter said. Some existing shareholders of the startup didn't expect the company to be bought out so quickly, the people said. Selling to Meta gives Manus access to distribution channels such as WhatsApp and Instagram, and a well funded owner able to help cover computing and infrastructure costs. Meta said it plans to continue to operate and sell Manus service and integrate it into its suite of social media products. Yes, more on that from our friend M.G. siegler, who says a bit counter to what I suggested yesterday that Meta basically just bought into the enterprise market with this Manus deal, which may eventually help it expand into cloud offerings like those of its big tech peers. Now, whether or not other businesses will want to buy enterprise offerings from Meta is another matter, but I suspect keeping the separate Manus branding with the Meta firepower and resources behind the scenes could help them here. They'll undoubtedly still face the same challenges that others trying to break into enterprise sales run into from Google on down. Selling into enterprise is just a different beast which requires different muscles. It took a Google bringing on board Thomas Kurian from Oracle for this to truly work with the necessary growing pains along the way. The tech Manus has built can obviously help Meta's other AI efforts, including across their consumer facing products. The core Manus product is well regarded by many thanks to its agentic first approach, but they also don't run on their own models and but instead use multiple models from the likes of Anthropic and Alibaba. Will that continue under Meta? Meta would undoubtedly love that, but will those other companies that compete more directly with Meta like that? Unclear. Regardless, eventually one would assume that Meta would want their own new models helping to power Manus. So all of this plays into Meta's broader AI ambitions and goals as well, and the company undoubtedly wanted to make that clear by having Wang tweet his excitement about the deal. Lest you think this is yet another sign of internal turmoil within Meta, this deal seemingly makes a lot of sense for Meta on a fronts, and it also may point to the start of a renewed push into enterprise. Again, easier said than done, but don't be shocked if this is a wedge of sorts. If they can keep Manas expanding into businesses, we should see other Meta cloud offerings follow, putting them more in line with those aforementioned big tech peers and perhaps easing some concerns Wall street has with regard to their AI spend. Meta has of course tried this push into enterprise before in ways from workplace on down, but it hasn't really worked. But buying up a hot product team and tech is what Meta does best. Just ask the ftc. End quote. Elon Musk says XAI bought a third building called Macro Harder, reportedly adjacent to Colossus 2 that will take the company's training compute to almost 2 gigawatts, quoting Bloomberg. Musk has already built one data center in Memphis known as Colossus, and is currently constructing a second nearby site dubbed Colossus 2. The new building Musk posted about is in Nearby South Haven, Mississippi and adjoins the Colossus 2 facility, the information reported earlier Tuesday, citing property records any person familiar with the project. Musk has publicly discussed plans to build the world's largest AI center for AI training and posted earlier this year that Colossus 2 will eventually have 550,000 chips from Nvidia, which would cost tens of billions of dollars. XAI has been fundraising aggressively in 2025 to finance its ambitious projects. The company was in talks to raise $20 billion in debt and equity earlier this year in part to buy Nvidia processors for Colossus 2, Bloomberg News reported. End quote. Everyone's using AI agents to automate tasks, manage workflows and even make a decision or two. But here's the AI agents and make mistakes. They delete the wrong files, make changes you didn't want, or just go completely off script. Then you're left to clean up the mess. Unless you've got Rubrik Agent Cloud Rubrik Agent Cloud is the only platform that allows you to monitor, govern and rewind AI agent actions. One platform to help you unleash more agents faster without the risk. If your business relies on AI agents, you need the ability to monitor, govern and rewind their actions. Right now, listeners get exclusive early access to Rubrik Agent cloud. Head to rubrik.com that's R U B R-I K.com.
- [10:05]
A
This episode is brought to you by State Farm. Listening to this podcast Smart move Being financially savvy Smart move. Another smart move having State Farm help you create a competitive price when you choose to bundle home and auto bundling. Just another way to save with the personal price plan. Like a good neighbor, State Farm is there. Prices are based on rating plans that vary by state. Coverage options are selected by the customer. Availability, amount of discounts and savings and eligibility vary by state.
- [10:35]
B
The Journal has seen Data suggesting that OpenAI is paying its employees more than any major tech company in history. Quote the company's stock based compensation is about one and a half million dollars per employee on average across its workforce of roughly 4,000 people. That is more than seven times higher than the stock based pay Google disclosed in 2000 before it filed for an initial public offering in 2004. The $1.5 million is about 34 times the average employee compensation of 18 other large tech companies in the year before they went public, according to a Wall Street Journal analysis of data compiled by Equilar. The analysis reviewed major tech IPOs over the last 25 years. To keep its lead in the AI race, OpenAI is doling out massive stock compensation packages to top researchers and engineers, making them some of the richest employees in Silicon Valley. The equity awards are inflating the company's heavy operating losses and diluting existing shareholders at a rapid clip. The financial data shared with investors over the Summer shows that OpenAI's stock based compensation was expected to increase by about $3 billion annually through 2030. The company recently told staff it would discontinue a policy that required employees to work at OpenAI for at least six months before their equity vests. That development could lead to further compensation increases. OpenAI's compensation as a percentage of revenue was set to reach 46% in 20, the highest of any of the 18 companies except for Rivian, which didn't generate revenue the year before its IPO. Palantir's stock based compensation equaled 33% of its revenue the year before its IPO. In 2020. Google's was 15% and Facebook's was 6%, the analysis shows. On average, each company's stock based compensation made up about 6% of revenue among tech companies the Journal analyzed in the year before their IPOs, according to the Equilar data. One more 2025 the year that was Segment for you want to know the best performing sector of the S&P 500? This year it was actually data storage stocks, with SanDisk up more than 560% to become the top performer on the index, followed by pirs Western Digital in second and Seagate in fourth. Bet you can guess why. Quoting Bloomberg the artificial intelligence trade is moving and investors seeking cutting edge ways to play it are snapping up technology picks and shovel stocks as massive cloud service providers pour billions into new data centers. Data storage companies dominated the S&P 500 index in 2025. Meanwhile, AI linked power providers and cable and fiber producers such as Amphenol, Corning, NRG Energy and GE Vernova. But we're among the top 25. What we are focused on are the picks and shovels of where that money is being spent, said Matt Solis, a portfolio manager at Tortoise Capital Advisors, which doesn't own shares in any of the hyperscalers the chips to a degree, but more so some of the names that you haven't really heard of. End quote. Quick review of that Samsung Galaxy Z Tri Fold that came out recently. Bloomberg, in its review, laments poor camera performance and some unique design flaws that make it even less polished than regular foldable phones. After a week with the device, the reviewer says it's clear that Samsung's most ambitious foldable yet is also its most compromised. The price and sheer engineering complexity immediately mark it as a niche experiment, one aimed squarely at early adopters, while a series of design missteps make it feel less refined than today's normal foldables. The Tri Fold closes like a wallet, dividing the interior display into three panels. Fold it in the wrong order and the phone scolds you with warnings and vibrations in interaction hurdle that nearly everyone I handed it to stumbled over at first. Credit where it's due, though, the hardware itself feels impressively solid when closed. It's roughly the size of a 6.5-inch phone with tight tolerances, dense construction and hinges that open and close with reassuring resistance. That thinness, however, comes at a cost. Samsung shoved all the camera hardware into a large, heavy rear module, creating a lopsided feel in the hand and an awkward wobble on flat surfaces. That imbalance undermines the core premise of a pocketable tablet, making extended video watching surprisingly uncomfortable. Despite a well suited aspect ratio. The inner 10 inch display also suffers from glare and reflections that rival devices already have learned to tame. And camera performance, especially in low light, lags behind far cheaper phones. Ergonomic quirks, from hinge bumps that interfere with gestures to a slower sight mounted fingerprint sensor further chip away at the experience. Battery life is another weak Link. A modest 5,600 milliamp hour cell struggles to power the large display, with gaming and video quickly draining it. Software optimizations, including an on device Dex mode, hint at what's possible, but many apps remain poorly adapted to the unusual form factor. In the end, they say the Galaxy Z Trifold doesn't meaningfully advance the foldable category. Instead, it highlights just how hard and expensive it is to add another crease in the name of Innov. For now, even Samsung's more conventional foldables feel closer to the future. This device promises. Only one long read for you this week, but it's a doozy. Shengdong Wang a research Engineer at Google DeepMind in London has a long essay out a reflection on AI advances in the past decade and how scaling and time horizon trends might point to a far greater capability in the decade ahead. He argues that Metter's time horizon plot, showing that the best AI can complete tasks of increasing human time length with moderate reliability became many people's go to evidence that AI progress is accelerating toward sudden world shaping change this year. But he also calls it a dangerously easy graph to misread or over extrapolate. The benchmark is heavy on coding tasks and success is often only 50 to 80% reliable. The the human baseline may be skewed by contractor speed and most importantly, it's an empirical trend without a clean theory explaining why it should continue rather than plateau. From there, the piece steps down through deeper reinforced trends that make the time horizon story feel less flimsy while still messy. He recounts repeatedly underestimating AI, first as a student impressed by AlphaGo and early generative models, later as a DeepMind researcher who saw a thousand fold compute increase unexpectedly produce a top performing embodied agent. They situate time horizons within broader scaling laws, performance predictability improving with more COMPUTE data and model size, and note epics claim that training compute has grown around four to five times per year for 15 years. Yet scaling too has survived via regime shifts and constant human ingenuity, broken and repaired by new techniques. So it's persistent but not guaranteed. Below all that sits Moore's Law, also empirical, also sustained through reinvention and Sutton's bitter lesson. General methods that scale with compute eventually beat hand design cleverness. He describes visceral compute wave moments, including newer giant models making simulated agents look suddenly smooth and capable, and likens AI's coming visibility to the pandemic. It's publicly predictable before it feels real and then nowhere and then everywhere. In the end, he emphasizes first order effects compounding compute and falling costs of intelligence as likely enormous, while second order impacts politics, labor regulation, culture, remain underdeveloped and urgently need some serious thinking. The tone is ultimately the trends are imperfect, but their stubborn persistence is exactly why we may still be wildly early when it comes to AI. There will be no shows at all tomorrow or Friday. No weekend bonus episodes. The next show of any kind on this feed will be on Monday, January 5th, where I will be coming to you from Las Vegas for ces. Thanks as always for being a listener. Thanks for sticking with me this year through the transition to techbrew and Happy New Year to you and yours. Ever spend all day fishing and catch nothing. That's what happens to hackers when Cisco Duo's on watch. Every login, every device, every user protected. Cisco Duo fishing season is over. Learn more at duo. Com.