Daily AI Buzz · Nov 2 Morning Edition: AI-Built Apps & DeepSeek's Open-Source Ledger

Lots of new stuff in the AI world today, and a few of them are pretty interesting. Today we're mainly covering these: one, AI building its own note-taking app; two, the Flock contract cancellation drama; three, Proton being misread as an AI story; four, US jobs data; five, a fifteen-year-old developer hand-building a trainer. Then we'll dive deep into two things: GPT-6 modeling the AI-waifu family bucket, and what DeepSeek actually makes money from after going open source. Yeah, that's a lot to unpack. Let's run through the news first, starting with this one from Reddit. There's a pretty gripping post on Reddit. Someone used AI to build an Android note-taking app called Hark Pro. Building apps with AI, that's not rare these days. The point is why he built it. He was fed up with Samsung Notes only storing things and not thinking, so he made one with a thinking partner. I get that pain point. Most note apps are just a trash can—you throw stuff in and never look at it again. It has a connection engine that links notes across months and generates a knowledge graph. And an idea-pushing module that can expand, investigate, challenge, reframe, and synthesize. Those actions are well thought out. The challenging step—a lot of tools won't dare do that, they're afraid you'll get upset. And it promises not to overwrite the original text, plus semantic search with source links. Not overwriting the original is key—otherwise you wouldn't dare let it edit. But honestly, this is a personal project; we'll have to see about generality and public availability. But at least it proves one thing: AI as both builder and thinking partner is a viable path. Right, going from general assistant to deeply embedding in personal workflows—I'm on board with that direction. Second item, a report on The Verge about license plate recognition camera vendor Flock Safety. Flock, the license plate recognition one. Several cities decided to terminate their contracts, but the exit process was anything but smooth. In Plymouth, Minnesota, the city council voted to terminate in September, and the cancellation request went unanswered for over a month. Unanswered for over a month, and the cameras are still up all over town? Yep, cameras are still there, data might still be getting collected. This is the invisible barrier for public institutions exiting surveillance tech. The contract didn't specify deadlines for equipment removal, data deletion obligations, or penalties for breach—so you're stuck. So when procuring, you have to lock these clauses down. Otherwise, the gap between voting to terminate and actually terminating drains public resources and credibility. Then there's another one from Hacker News with an interesting title: the Proton pivot—AI shouldn't take the blame. Oh, I saw that. These days, any company strategy shift and the first reaction is AI—that attribution is just lazy. The author's point is that for a company like Proton, which started with encrypted email, VPN, and cloud storage, the real assets are user trust, compliance records, transparency reports, and open-source audits. Right, not model capabilities. Its product focus shifted, more likely a response to user concerns about data flows and business model. So he warns the tech community to beware of this AI attribution bias. Judging privacy products should look at data governance commitments, not how many AI features are piled on. I fully agree with that. Features can be copied, commitments can't. Next one isn't so directly about AI—Bloomberg's report that US jobs data is picking up, which might push the Fed to keep raising rates. This one actually matters a lot for people in AI. How so? Strong jobs means wages and spending hold up inflation, so rate hike expectations rise. Growth assets are most sensitive to rates—tech and AI valuation logic, plus early-stage funding pace, all get driven by the cost of capital. But the article also reminds us that one jobs report can't decide policy—you have to look at inflation, consumption, credit too. Right, track the data, don't overreact. The last one I really like—on Reddit, a fifteen-year-old developer hand-built a hardware-aware small language model trainer from scratch. Fifteen? At that age I was still figuring out how not to fail my classes. His goal was clear: build an end-to-end training pipeline on consumer hardware, just to really master Transformers. The hardware-aware module is the highlight. It auto-detects NVIDIA, AMD, or CPU, runs matrix multiplication benchmarks to measure actual TFLOPS, then uses a Rust engine with VRAM budget to calculate how big a model it can train—capped at 100 million parameters to prevent VRAM blowups. The data pipeline scrapes papers from arXiv and specifically handles LaTeX code blocks, worried that tokenization would break formulas. That detail shows he's actually run it and hit the pitfalls. The project is still early—he's on Reddit asking for experience managing training loops on constrained hardware. That learning approach is way better than just following tutorials. Alright, that's the news. Next, let's dive into two hot topics on Bilibili. First, there's a video on Bilibili called 'GPT-6 Modeling the AI-Waifu Family Bucket,' turning Claude, GPT, Doubao, GLM, Grok, Gemini, Kimi, MiniMax, and Qwen all into 3D-toon AI waifus. 'Family bucket' is a good term—one character per family, all neat and tidy. 86,000+ interactions—that's pretty hot. But if you ask whether this counts as tech news, I'd say no. It's fan-created content; the barrier is modeling and rendering, not the models themselves. So why do you think it went viral? Because it personifies abstract models. Users already interact with these models and have personality profiles in their heads. Someone draws it, and everyone goes, 'Yes, that's exactly the vibe.' From a product perspective, this is users spontaneously doing brand personification—something the companies themselves might not pull off. Right, and it's free brand equity. But I have to throw some cold water—this kind of hype comes and goes fast, and it has nothing to do with model capability. So the conclusion: the AI-waifu family bucket is a great signal of user mindset—it shows people already have emotional connections to these models—but it's not a tech trend indicator. Agreed, enjoy it for fun, and incidentally see who's more popular. The second topic is more interesting. There's a Bilibili video asking: why call it an AI bubble when DeepSeek has even open-sourced its per-card revenue? That's a sharp question. The core contradiction is right there: if you've opened up the books, where does the money come from? And there's another video titled 'Big Fat Fish's Core Tech All Open Source—What Else Can DeepSeek Make Money From? Starting from Bilibili's fish flood and the hundred-billion funding.' The two are two sides of the same coin. Open-sourcing per-card revenue proves efficiency, not profitability. Wait, I have to follow up on that. How does open-sourcing per-card revenue prove efficiency? It's putting output per unit of compute on the table, telling everyone this architecture doesn't waste cards. Externally it's tech credibility, internally it's engineering discipline—but it doesn't mean the business model works. Then how do you explain the hundred-billion funding? Open-sourcing on one hand and raising that much on the other—isn't that contradictory? Not contradictory. What's open-sourced is technology; what funding buys is compute and time. Think about it—models iterate so fast that whoever trains the next generation first holds the position. This arms race burns money. But isn't that the definition of a bubble? Sustained by funding, not making much money itself. I think you have to look at it on two levels. If funding is just telling a story for valuation, that's a bubble. If funding converts into real compute and talent, that's an investment phase. The difference is whether per-card revenue can be consistently delivered. So the open-sourced per-card revenue you mentioned is actually a form of self-verification, and also pressure—once efficiency is public, you have to maintain it. Exactly. And that 'fish flood' comment on Bilibili reflects another side: content-side traffic explodes, but how traffic turns into revenue—that path isn't paved yet. From a product angle, what do you think is its most likely path to making money? API and inference services are the most direct—open-source models build an ecosystem, and the ecosystem feeds back call volume. Then there's enterprise customization and private deployment, where willingness to pay exists. But both have strong competitors—the API price war is brutal. Right, that's why I say open-sourcing per-card revenue is more like a tech calling card than a financial statement. Okay, let me try to sum up. Open-sourcing per-card revenue proves engineering efficiency and technical confidence; the hundred-billion funding buys compute and a time window. Bubble or not, it's not about whether it's open source—it's about whether per-card revenue can be consistently delivered and turned into real income. I'll accept that summary. One point of disagreement: I think this window is longer than many people think—the money will keep burning for a while. Then let's leave that disagreement for the data in the coming quarters to settle. Great chat today. In the news, a fifteen-year-old hand-building a trainer; in the deep dive, the open-source ledger debate—both are worth chewing on. Yeah, especially the fifteen-year-old—the younger generation is formidable. Alright, that's it for today. See you at noon.