Daily AI Buzz · October 27 Morning Edition: Vercel Connect & Doubao Agent Tutorial

There's a lot of new stuff happening in AI today, and several of them are about agents actually getting deployed. Today we're mainly covering these: one, Vercel Connect; two, more vulnerabilities don't equal more breaches; three, files are dead, long live URLs; four, Databricks teaching free tier; five, big fines can't rein in the giants. Then in the deep dive we'll talk about two more: six, the Doubao Agent tutorial; seven, using an AI phone as an ADHD launch buddy. Yeah, these are all pretty down-to-earth. Let's start with the Vercel one. Sure, I think this is the most practical one today. First up, just saw Vercel's changelog—they added a new command to the CLI called vercel connect discover. Right, the key is you can browse over a thousand services in Vercel Connect without leaving the terminal. You can search by name, by keyword, and filter by category. It even filters by connection type? I see it says MCP or OAuth. Yep, and you can also filter by whether Webhook triggers are supported. That's interesting. But that's not the main point—the main point is it supports format=json output. Oh, so you're saying this is made specifically for agents? Exactly. Humans look at tables, agents look at JSON—one set of commands serves both. And after discovering, you can directly run vercel connect create to build a connector—discovery to creation in one go. That's a smooth handoff. But there's a catch: the CLI needs to be version 64.1.0 or above. The version requirement is no big deal. I think the real signal here is that agents have become first-class citizens of the CLI. Noted. Now the second one: there's a video on YouTube with a counterintuitive title—more vulnerabilities don't equal more breaches. I half agree with that. The number of vulnerabilities and breach incidents aren't linearly related. When a vulnerability is disclosed, if the response is fast, it might never reach the breach stage. So the key is response speed and process, not the absolute number of vulnerabilities. Right, that's the same idea as the security scanning skill we mentioned earlier—finding it doesn't count as a win; handling it does. Okay, that's enough on this one. Third: an episode of The Vergecast titled Files Are Dead, You Only Need a URL. That's a bold title. But I agree with the direction—the way files are organized is indeed being replaced by URL-style addressing. Think about it: when an agent wants to call a resource, would it rather get a URL or a file path? Definitely a URL. URLs are naturally remotely accessible, authenticatable, and can be directly put into context. So the timing of this episode's popularity is really riding the agent wave. Exactly. Let's move on to the next one. Fourth: Databricks published a blog about using its free tier for teaching. That's a clever point. The teaching environment matches enterprise production, so students don't need to migrate after learning. It can run data engineering, SQL analysis, machine learning modeling, and collaborative notebooks. But it also reminds us that the free tier has limits on compute resources and concurrent users. Right, so the article suggests starting small and then expanding—don't roll it out to the whole grade at once. Fifth: there's a discussion on YouTube about why billions in fines still can't rein in tech giants. That's simple: fines are too small a percentage of giants' revenue to be a real constraint. So you think we need different measures? At least we can't rely on fines alone. But that's too far off-topic—let's just touch on it today. Alright, that's it for the news. Next, let's talk about a tutorial that's been really popular on Bilibili recently. You mean that step-by-step Doubao Agent tutorial. Yeah, the title is very direct: Stop Just Chatting with Doubao—Build an Agent from Scratch in One Hour, covering workflows and automation thoroughly. The fact that this kind of tutorial goes viral shows exactly one thing: people don't lack models anymore; they lack the know-how to bring them to life. That's a bit harsh. So why do you think the step-by-step format is what gets popular? Because the barrier is in the scenario, not the technology. How to string workflows together, where automation triggers—those are the real sticking points. But that's contradictory: if the tutorial has to be step-by-step, doesn't that mean the product itself isn't foolproof enough? No, I don't think the product is lacking—it's that agents are inherently tied to scenarios. If you give a generic entry point, users don't know where to start. So the tutorial is essentially filling the scenario gap. I buy that. And look at Bilibili as a platform—long videos plus hands-on practice are perfect for explaining workflows. Then the second one: I used an AI phone to make an ADHD launch buddy. This one got over 18,000 interactions—absurdly high. I had a feeling just from the title. The hardest part of ADHD isn't doing things—it's the initial launch. So you see it as something that pushes you to start? Right, and the fact that it's on an AI phone is key—phones are always with you. It's right there when you need it; you don't have to open a computer app specially. Isn't that turning the agent from a tool into a companion? Sort of. And this kind of need is very vertical—general assistants actually can't do it well. So you see, the Doubao tutorial teaches people to build agents, while this one has the agent directly切入 a specific group—completely different paths. To sum up: one side is lowering the barrier to building, the other is drilling into vertical scenarios—both happening at the same time. So the conclusion is: the opportunity in this agent wave isn't in the model itself—it's in who can solve both the scenario and the barrier at the same time. Right, and let me add: that's also why it's popular—people have moved from watching to doing. Alright, that's all for today. See you at noon.