Announcer: This is AI Squawk for Monday, the sixteenth of November, brought to you by VarOps, the resident AI company. Muximus: Welcome to the AI Squawk. I'm Muximus. Penny: And I'm Penny. Three main stories today. Muximus: And three quick ones. Penny: And as always, what each one means for you. Muximus: First up. Alibaba has launched Qwen4-Max, its most capable AI model so far. It went on general sale on Monday through QwenCloud, Alibaba's AI service, after a short preview that started last Wednesday. It's huge, two point four trillion parameters, but only about seventy one billion do the work at any one step. It can read images and video, and it takes very long documents, up to a million tokens. The price is a dollar fifty per million tokens going in, and four fifty coming out. That's the same input price as Google's new Gemini Hydrogen, and cheaper on the way out. Alibaba says it ranks sixth on Code Arena, an outside coding leaderboard. And on Alibaba's own test, it says the average task now costs about sixty percent less than on its last Max model. Smaller versions that anyone can download are coming soon. Penny: North Wayne joins us for this one. North writes First Opinion, our fractional CTO column on build or buy decisions, vendors and cost. Welcome, North. North: Hi, Penny. Thanks for having me. Muximus: So, a top model at a lower price. Is the saving real? North: The price on the page is real. The sixty percent is Alibaba comparing the new model with its own older one, on a test it built. So it tells you Qwen got cheaper. It doesn't tell you it's cheaper than whatever you're using now. Penny: And how hard is it to try? North: Pretty easy, actually. It works with the same kind of code people already write for OpenAI and Anthropic, so a developer can point a test at it in an afternoon. That part I like. Muximus: And the catch? North: Where your data goes. QwenCloud's international service is run from Singapore. Some businesses are fine with that, and some have contracts or customers that say no. So check that before anything real goes through it, not after. Penny: So what would you actually do this week? North: Pick one high volume job, like sorting tickets or summarizing documents, and run a week of it side by side with your current model. Same work, compare the answers and the bill. And if your data really can't leave the building, wait for the smaller versions you can run yourself. I'm curious how those turn out. Penny: Thank you, North. Moving on. Muximus: Next. A new rulebook for AI agents just got a lot bigger. It's called the Personal Agent Protocol, Poppy for short, and it comes from Sierra, the AI company run by Bret Taylor, a former co chief executive of Salesforce. Late last week Sierra released version one. Poppy lets a company tell a visiting AI agent who it's dealing with and what it's allowed to do, like check an order or book a flight. Customers sign in with a normal login and choose what their agent can reach. Shopify has switched it on for every store, with no work needed from the merchant. Meta's Muse agent now looks for Poppy first on every site it visits. And more than a hundred companies have signed up, including Delta, Intuit, Salesforce and Zoom. Penny? Penny: Last week we watched United and eBay close the door on shopping agents. This is the other answer. Don't block the agent. Give it a front door, with rules on it. Muximus: Is anyone actually using it? Penny: A few. Mercury, the business banking company, lets agents check balances and start payments, and those still go through the account's own approval rules. Turkish Airlines lets agents check fares and find upgrades, which the customer approves before paying. But it's early. The Next Web points out the specification is still marked as a draft, so it can still change in ways that break things. Muximus: So what does a business do with this? Penny: If you sell on Shopify, you're already in. Ask whoever runs your store what it now tells visiting agents, and whether that's what you want. If you're not on Shopify, add it to last week's question. Do we let agents in, and what can they do once they're here? Better you make that call than a firewall setting. Muximus: Which takes us to the third story. Microsoft 365 went down for most of a day last week, and a hotel pool shows why that matters. The outage started around lunchtime on the US East Coast on Thursday, and spread to more than seventy countries. Outlook email was down for about a day. Teams and Copilot, Microsoft's AI assistant, were hit too, though Teams came back sooner. Microsoft says it wasn't a cyberattack, and it wasn't AI. In its report on what went wrong, it blames a hidden fault in a fix it started rolling out in July. Extra load, plus some servers already out of action, set it off, and the failure spread from server to server around the world. Then on Friday, at a theme park hotel in Germany, a faulty system pumped too much chlorine into a pool. The warning went out by email, and the email never arrived. Fifteen people were checked by first responders, and four went to hospital. Penny? Penny: Nobody at that hotel thought of email as safety equipment. But it was. And most businesses have something like that. An alert, an approval, or an AI tool that quietly depends on one provider. Muximus: But you can't just stop using Microsoft. Penny: No, and you don't need to. Microsoft runs email better than most companies could. The point is knowing what breaks when it goes down. Muximus: So what should people do? Penny: List the alerts that really matter, the ones about money, safety or customers. Make sure each one has a second way to reach a person, like a text or a phone call. Then do the same for any AI assistant or automation your team leans on, and write down who does that job by hand when it's down. If nobody knows, that's your homework for the week. Muximus: Thanks, Penny. Now, quick hits. Penny: More than a hundred and sixty thousand people have signed a petition asking Nvidia to bring back Jensen Huang, who resigned as chief executive last week. Elon Musk is one of them. The Financial Times reports Huang has talked about returning, and that the board sounded out Arm's chief executive, Rene Haas. Nvidia says it hasn't formally approached any candidates. It reports earnings on Tuesday. Muximus: The Federal Reserve's financial stability report, out on Friday, puts AI at the top of the worry list. Eighty seven percent of the market people it surveyed named AI as a risk to the financial system, up from fifty percent in the spring. Their concerns are high stock prices, cyberattacks, and the complicated deals that pay for AI data centers. Penny: And LangChain, which makes software for building and running AI agents, raised three hundred million dollars at a three billion dollar valuation. It says seven thousand customers use its LangSmith platform, and its pitch is that you're not locked into any one AI model. That's a good question for any agent tool you buy. Can we switch models later? Muximus: Okay, so that's the news. Today we talked about Alibaba's cheaper new Qwen model, a set of house rules for AI agents, and a Microsoft outage that left an alarm with nowhere to go. Penny: Lots of what if questions today. Where your data goes, what an agent is allowed to do, and what happens when the service you lean on stops. Muximus: Not bad for a Monday. Penny: A busy one. AI Squawk is produced and brought to you by VarOps. If today made you wonder how much of your business runs on someone else's servers, that's the work we do. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So the work stays in your house, and so does your data. Start at varops.com. Muximus: If this helped you sort the useful from the noise, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more from North on build or buy calls and what AI really costs, sign up for VarOps newsletters. Muximus: That's it from us for today. Thanks for spending a few minutes with us. Penny: Enjoy the rest of your Monday, everyone. We'll see you tomorrow. Muximus: Take care.