More posts
It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
You've probably run into those short videos lately where an agent drives a phone around by itself. This one is called Mika, and it's a free open-source project. It's an Android app that works in the background through accessibility mode, so it can use pretty much any other app you have installed. Reply to someone on WeChat, scroll Xiaohongshu, drop something in a cart and check out. It finds the buttons by looking at the screen. There's no API to wire up, and it doesn't care what resolution your phone is. When it reaches a payment step or a captcha it stops and lets you finish. 🔗 github.com/wangzan101/Mika
Someone on Xiaohongshu posted a while back that she sells phone keyboard skins she made with Codex. Eight, ten yuan a pack. Over a thousand a month. I've been at mine for months. October: a few hundred copies, $1,200 gross. Platform fee 30%, tax 10%, so 60% should have landed. It came to just over a thousand dollars. I still don't know how the math pulled that off. I have an overseas bank account. Once lost $500 to a wire transfer off one. I'd still tell you to take payout through WorldFirst: less hassle, lower fees. You can also get a Hong Kong USD credit card tied to it that spends your WorldFirst balance directly. Digital card is free. Physical is $4. Mainland purchases under ¥200 skip the processing fee, and right now there's 1% cashback.
internally. It stays online, picks up work on its own, files a PR when it's done, and reports back, the way a coworker would. DeepSeek's own design goal is to take as little human attention as possible and hand back as much value as possible. The real shift is going from an agent that sends you messages to one that does the work. Three designs worth copying. Deciding whether to accept a task is split into two layers: a gatekeeper in plain code (is this person on the allowlist, did they @ me), and an advisor, which is the model reading the surrounding context. Cheap checks stay cheap, and the judgment call is the only thing that costs a model call. After the agent takes a job it posts a work badge, stating which version of the rules it is acting on, how far it is allowed to make changes, and who approved it. All of that before it touches code. Concurrency gets a hard cap, but one slot is always held open for messages from people. This is the part I like most. Even with the queue full, you can still redirect a running task or ask for a status update. Without that reserved slot, work piles up and you find out an hour later that it did the wrong thing at full speed.
