Daily AI Buzz · October 23 Morning Edition: Gemini 4 vs. DeepSeek's New Flash

Lots of new stuff in the AI world today, and a few things are gearing up for a head-to-head showdown. Pretty lively. Today we're covering a few things. First, Vercel's free domain expansion. Second, don't bank on early sales for startup growth. Third, Teenage Engineering's new speaker. Fourth, the AI video of a pelican riding a bike. Fifth, how to polish a new grad's resume. Then for the deep dive, we'll talk about Gemini 4 Argon and DeepSeek v4.1 Flash going head-to-head. Yeah, those two are really going at it. Alright, let's run through the news quickly first, then dig into that showdown. First up, just saw on Vercel's changelog—they've added 8 more top-level domains to the free first-year registration for Pro teams. .co, .studio, .company, .info, .pro, .live, .digital, and .agency. That brings the total to 16 options. That's pretty solid for indie developers. Free first year means saving some cash when you're just starting your brand. But note—each Pro team still only gets one free domain, premium domains don't count, and from year two it auto-renews at the price on the checkout page. So the point isn't the freebie, it's having more suffix options. Like .agency or .studio—if you run a studio, you know exactly which one to pick. Right, it lowers the barrier to launch a bit. Second item—TechCrunch has a counterintuitive piece saying the secret to startup growth isn't early sales at all. I agree. A lot of those early deals come from connections—you can't replicate that with a different customer. That's not growth, that's luck. The article quotes Luke Arno from Lovable and Tomasz Tunguz from Theory Ventures saying the first 90 days should focus on a repeatable sales process. Exactly—define your ICP, standardize the pitch, build a feedback loop. Founders need to go from super salesperson to coach, and keep the team lean. And look at retention and repeat purchases, not those pretty but empty metrics. Alright, this one's worth pinning to the wall. Third item—not really AI, but fun. The Verge reports Teenage Engineering dropped the EP-35, a battery-powered instrument speaker. The I/O is pretty wild—line in, quarter-inch guitar input, and an XLR mic input with phantom power. Phantom power on a portable speaker is rare—means you could do small outdoor vocal gigs or podcast recording. But they didn't release any key specs—amp power, speaker specs, battery life, all blank. Gotta wait for real tests before buying. Yeah, competitors are the Boss Katana Air and Yamaha THR series. It's betting on full I/O and design language. Let's wait and see. Fourth item—Simon Willison's blog noted that in a music video by Dutch band De Staat, at 0:44 there's a pelican riding a bicycle. I know that meme—pelican on a bike was an informal test for AI video, and now it's in a commercial music video. Apparently made by indie studio Fake it or Leave it, and it caused a big debate on HN. The key is it's blended with live action and post, so the seams are hard to spot. That's the most realistic path for AI video right now—low cost, short cycle, adding visual density to things like music videos, while creative judgment stays with humans. So restraint is key—use it as a brief unexpected punctuation, not stuffed into whole segments. Fifth item—a new grad on Reddit asked if their resume is good enough. That kind of post is so typical. Background's not bad, lots of projects, but whether you get an interview comes down to details. Core advice: cut projects like MNIST, make room for one that shows the full pipeline. Right, spell out the problem definition, why you chose that model, quantified metrics, and deployment details. For CV roles, non-classification tasks and real data handling experience matter more. Structurally, pick two or three projects ordered by value, layer your skills section, put GPA after projects. The bar's simple: if your main project has quantified results and engineering traces, your interview chances are decent. Otherwise, go back and deepen the project. Alright, that's the news. Now let's get to today's hardest topic—models going head-to-head. Google officially launched Gemini 4 Argon, and a morning news video on Bilibili hit over 60k engagement points. That scale shows people were really waiting. With Gemini at gen 4, the question isn't whether it can compete, but how far it's closed the gap. But DeepSeek isn't idle either—v4.1 Flash 1017 came out, and a comparison review video testing it against GPT 6.1 Sol got nearly 40k engagement. See, these two colliding is fascinating—one is pushing the flagship ceiling, the other is on the Flash cost-performance path. Totally different testing dimensions. So someone on Bilibili specifically asked how to evaluate DeepSeek's new Flash model released October 17th. Discussion was pretty focused. My take is direct: the Flash line isn't about topping benchmarks, it's about pushing down the unit price of usable capability. Wait, so you're saying flagships don't matter? But Gemini 4 Argon is clearly aiming for the ceiling. No contradiction. The ceiling decides how people see you; the unit price decides how many actually use you. Two battlefields, you need both. From a product view, if my team were adopting, which side should we bet on? Depends on your pipeline. Heavy reasoning and quality go to the flagship; high-frequency, low-cost goes to Flash. The industry is basically on this hybrid approach now. Got it—so it's not either-or, it's layered use. Then why is this blowing up right now? Because both played almost simultaneously and both have comparison material. Once those Bilibili reviews dropped, people had a way to compare side by side. And comparison reviews are naturally traffic magnets—one line about who's stronger and you've got a debate. True, but judging from clips alone can be misleading—test set selection matters a lot. Keep that in mind. So to sum up: Gemini 4 Argon pushes the ceiling up, DeepSeek v4.1 Flash pushes the unit price down, and the real mainstream approach is layered hybrid use. Exactly. When choosing a model going forward, don't ask which is strongest—ask which one for which part. Alright, that's about it for today. Enough info to digest all morning. See you at noon.