Daily AI News · October 20 Midday Edition: Flipkart Delivery and Designing for Wait

Lots of new stuff in AI today. I scrolled around and a few things are pretty interesting. Today we're mainly covering these: one, Flipkart delivery; two, the ChatGPT domain dispute; three, AI coding assistants; four, AI reviewing articles; five, searching local models; six, Lula protesters against Meta; seven, designing for wait; eight, Katto video clipping. Yeah, all pretty meaty. Let's start with the first news item and work our way down. First up, I just saw something on TechCrunch. Walmart-owned Flipkart is quietly building a food delivery business in India. They say it's open to a few thousand employees first, and the public beta hasn't officially started yet. That's a pretty typical playbook. Run it with internal employees first, get the fulfillment chain figured out, then open it up. The Indian food delivery market is already cutthroat. They're using their e-commerce foundation to grab a piece of instant delivery. So what does that have to do with AI? Isn't there a scheduling model behind all delivery now? Right, route planning and capacity scheduling basically rely on AI. But this news item itself doesn't mention specific tech, so let's not force it in. Okay, we'll just note it as an industry trend. Second item, there's a fun story on dev.to. A developer joined Hacktoberfest, was naming a project, and ended up arguing with ChatGPT. Arguing? That's an interesting way to put it. How did the argument go exactly? He wanted to use a domain name and asked ChatGPT for its opinion. The model was super insistent about its preference and wouldn't budge. That's kind of like the model's persona being too strong. If you state your preference heavily, it'll just keep pushing back. So for subjective things like naming, don't fully trust the model. Third item, also on dev.to, about AI coding assistants. The headline is pretty big: 'The End of Boilerplate,' saying AI coding assistants are redefining software engineering. I buy that direction. Repetitive skeleton code can definitely be handed off to tools in bulk now. But that's not the main point. The main point is people need to be able to review and judge whether the generated code is correct. Which leads right into the fourth item. Someone on dev.to did an experiment where another model reviewed their articles. Seven articles, forty review comments, all quantified and recorded. That experiment design is interesting. Having a different model review gives a different perspective and makes it easier to catch factual errors and logical leaps. So AI isn't just writing, it can be a second pair of eyes. Fifth item, how to cram more web search data into local models. The core issue is the context window. If it doesn't fit, you have to filter and compress. That's really an engineering job. Okay, that's the news. Next, let's dig into two topics. First, this one's hot on Hacker News: Lula supporters in Brazil took to the streets protesting Meta. They're condemning Meta for election interference. About twenty comments, score close to a hundred. These kinds of things are getting more sensitive. How platform recommendation algorithms influence public opinion has always been a controversy. But that's contradictory. The platform says it's neutral, the people on the street say it's interfering. Who can really say? No one can, because algorithms themselves have biases. What to rank, what to suppress—those are all choices. Think about it: recommendation systems optimize for dwell time, so they naturally amplify emotional content. So is there a technical solution to this controversy, or can it only be handled by regulation? Short term, it's regulation and transparency reports. Long term, it depends on whether there's an auditable recommendation mechanism. Okay, let's note that. The second topic I really want to talk about is that dev.to piece on designing for wait. You put in a photo, you get out a video, and in between there's about ten minutes of nothing. Those ten minutes are where you lose people. If users don't know the progress, they'll think it's stuck. The author builds single-purpose tools and uses this to turn waiting into part of the product. Right, the key to async generation isn't speed, it's giving users a sense of control. Progress bars, stage indicators—all of that needs designing. So the slower the generation, the more this design matters? Exactly. In a ten-minute scenario, you have to break the wait into several perceivable small stages. So the conclusion is, for async AI products, competitiveness is half the model and half the waiting experience. Agreed. Last, there's a small news item: a tool called Katto that clips videos from Claude and Cursor. Yeah, it cuts long videos into vertical shorts, automatically finds highlights, adds subtitles. The toolchain is getting more and more specialized. To sum up, today's items cover both industry trends and product details. All pretty down-to-earth. Yeah, that's all for today. See you tonight.