Here's a stat that doesn't get talked about enough: the global median income is only around $4-5k a year. Most of the world simply cannot afford a $20/month ChatGPT Plus subscription or an Anthropic/Claude plan, let alone stack multiple of them to actually get real work done. That's a huge structural advantage for DeepSeek. It's free to use, genuinely competitive on a lot of tasks, and it can be run locally if you have decent hardware, no subscription, no per-token API bill, no dependency on a US company's pricing decisions. For the vast majority of people on the planet, that's not a nice-to-have, it's the difference between having access to frontier-level AI or not having access at all. If you're building anything global-facing right now, defaulting to ChatGPT or Claude as the only option quietly prices out most of your potential users. DeepSeek (and open models like it) close that gap. This is why I think it's worth paying close attention to DeepSeek specifically, and why running it locally is worth setting up even if you're not hurting for money yourself. It's a hedge against pricing/rate-limit changes, and it's a much more realistic default for the actual global population. Curious if anyone here has DeepSeek running locally already, what's your setup and how's performance been for real tasks?

Been building no-code automations for a while now, and I've always defaulted to ChatGPT or Claude for the heavy lifting. That was just the assumption American frontier models were the obvious choice, no real debate needed. Ran a side-by-side comparison recently using Qwen on some of my actual automation tasks against Claude and ChatGPT, and honestly, Qwen came out ahead. Not "close enough," but genuinely better output on the specific tasks I threw at it. That was not what I expected going in. What's crazy to me is the trajectory here. It feels like open-source/Chinese models are maybe 2-2.5 months behind the top US labs at this point, and that gap keeps shrinking every release cycle. A year ago that gap was massive. Now it's basically a rounding error for a lot of practical, real-world tasks. Curious if others here are seeing the same thing in their workflows. Which tasks are you finding these newer models are actually winning on vs still lagging?

It solved 10 open math/CS problems that have stood for decades. Including proving non-sofic groups exist — a question that’s been open since 1999. OpenAI published a 249-page paper with machine-checkable proofs on GitHub. Total compute cost for all ten: ~$2,000. Let that sink in — problems that stumped mathematicians for 27 years, solved for less than a MacBook. It’s not a chatbot, it’s a system. Multiple agents working together for hours or days on one problem planning, testing, revising, without you babysitting it. Compare that to 5.6, which (per leaks) still gets “ADHD” on long tasks — sets a goal, comes back 10 hours later, hasn’t finished it. Sam already showed it to DC policymakers. Astra is expected to be one of the first models run through the new federal pre-release review process. That alone tells you they think this one’s a different category of capability. No release date. No confirmed name. Could be GPT-6, could be a GPT-5.7 branch. OpenAI’s own naming ladder so far: Luna → Terra → Sol → Astra (each tier faster/dumber to slower/smarter). Here’s my take: we’ve spent this whole year optimizing prompts and routing to cheap models. That era’s ending. The next unlock isn’t “which model is cheapest” — it’s “which model can actually be trusted to run unsupervised for days.” That’s a completely different skill set to build around. Get comfortable with agentic/long-horizon workflows now. That’s where this is headed. 🧵

I built 8 apps alone. I made $3. And I bought it myself. That's not a joke. My only "sale" was me buying my own in-app purchase to test it. Had to cancel and refund it. That's the entire revenue history of seven months of building. I just posted the full breakdown on YouTube real App Store Connect screenshots, all 8 builds, no fake numbers. But I wanted to say the important part here first, because you're the reason I made it. I thought vibe coding was my advantage. It let me build fast and it made me feel productive. One app pulled thousands of impressions and made nothing, because there was nobody on the other side of it. No one to tell me what was broken. No one waiting for v2. TikTok gave me reach, but reach is one-way a video, then a funny clip, then some news, then you're gone. You can't build trust in a feed that fragmented. What I was missing wasn't code. It was a room. People I see every day, who'll tell me the truth, who I can actually launch something to. So that's what this is. I'm building the room first this time, and I'm building it in front of you. Question for you: what's something you've built that nobody's using? Doesn't matter if it's an app, a landing page, a spreadsheet, a half-finished repo. Drop it below I'll look at every single one and tell you honestly what I think is missing. That's the deal. Luke