ai_binbin

可以的

magpie简直是这个ai时代最伟大的发明 流畅度远超cc-switch bug远少于cc-switch 功能齐全，各家模型跟不同harness的适配度非常的高！ 正好我前几天的chatgpt pro的订阅取消了 我还特别喜欢codex的cua来做一些自动化的操作 这样的话我可以接入便宜的gemini或者qwen放到codex里面 甚至比gpt自己的模型还快 codex现在可以无缝切换各家大模型 至于开源的harness 本来就对各家model适配的比较好 所以在这种codex和claude app上 切换不同的模型体验非常好 上下文窗口也很好看 强烈推荐！ github.com/yetone/magpie

我也是这样认为的 我现在主力就是argon➕opencode 太像人了

支付宝终于下场干它最熟的事了：给 Vibe Coding 玩家直接接好收款。 出了个叫AI付 / Vibe Pay的东西，核心就一件事，让个人开发者和 OPC 能用 Agent 直接把网页和 App 的支付跑通。 以前写完代码，卡在商业化收钱这一步的人太多了。现在它把商品、订单、权益管理连带后台全打包进一个产品里，Web 和 App 一套搞定。 企业能申，个人也能申（独立开发者目前是定向开放）。 我还没来得及接，在搓独立产品的朋友可以先去探探路：aipay.alipay.com

I live in Claude and Codex most days. That does not mean the open source lane is dead. OpenCode is an open source coding agent. Terminal first. You can plug models into it and keep working without waiting for one vendor's UI mood. github.com/anomalyco/opencod… What people miss is the second layer. You do not have to fork the whole thing. Most of the useful stuff is plugins and SDK shells on top. Plugin for auth. Plugin for memory. Or a full desktop shell that sits on the same core. A few active ones worth knowing: OpenWork. Open source desktop shell with a Cowork feel. github.com/different-ai/open… OpenChamber. Workbench across desktop, web, VS Code. github.com/openchamber/openc… oh-my-opencode-slim. Lean multi agent setup if you want orchestration without a giant pile of config. github.com/alvinunreal/oh-my… The point is not "open source wins." The point is you get a real second cockpit when Claude or Codex is down, rate limited, or just the wrong shape for the job. If your stack only has one front door, you are renting your workflow. Keep Claude. Keep Codex. Also keep one open path you actually know how to start.

Lived in Argon this week. Not a vibe check. Actual work. Human feel is real. Agent work got better too. Less babysitting. More finishing the job. That is why I am staring at Carbon harder than the scoreboard. Near SOTA and still locked for most of us is a special kind of torture. I do not have it. I just want it. When does the door open.

I need carbon

Lived in Argon this week. Not a vibe check. Actual work. Human feel is real. Agent work got better too. Less babysitting. More finishing the job. That is why I am staring at Carbon harder than the scoreboard. Near SOTA and still locked for most of us is a special kind of torture. I do not have it. I just want it. When does the door open.

这下我要安装一个diplay了

Deep used Gemini 4 Argon this week. Not a vibe check. Actual work. It is genuinely good. The human feel is the part that stuck. Writing still sounds like a person. Long messy threads do not collapse into that polished nothing most models spit out. Agent capability also moved up for me. Less babysitting. More finishing the job. I used to roast Gemini for announce theater. Fair. Carbon is still locked for most of us. I am not in that room. But after living in Argon, I am actually waiting for Carbon now. Not hyping a chart. The current one already earned a seat in the rotation.

Everyone keeps saying Claude looks at screenshots and Codex controls the Mac kernel. Half right. Then wrong. Claude's loop is mostly eyes. Screenshot goes in. The model counts pixels. Click (x, y) comes out. Anthropic literally wrote that teaching Claude to count pixels was a big part of the training. On Mac, Claude Code still asks for Accessibility and Screen Recording. Accessibility is the hands that inject clicks and typing. Screen Recording is the eyes. That does not make Claude AX-first. Perception is still the image. Codex desktop is a different animal. That stack comes from the Sky acquisition. The better product writeups say the main driver is the macOS Accessibility hierarchy, the AX tree. Screenshots and AppleScript show up as backups, not the main path. It has its own virtual cursor. It can stay in the background. It can poke multiple apps without yanking everything to the foreground every two seconds. Almost everyone walks into the same trap. OpenAI's Computer Use API is still screenshot-family. Click coordinates. Get a screenshot back. Same genus as Claude's tool. Do not confuse that with the Codex Mac plugin. Same company. Different sensor stack. So Claude is mostly a vision-first GUI agent. Codex desktop is mostly AX-first, with vision as backup. The API line from OpenAI is vision-first again. When Codex feels sharper on native Mac apps, it usually is not deeper kernel access. It is structured UI state beating pixel guessing, plus a cursor that does not steal focus for sport. If your takeaway is OpenAI is more low-level, rewrite it. If your takeaway is the sensor layer matters more than the model card, keep that one.

Grok bot is so unbelieveable.

这次季候风~~~吹得那样凶~~~

Everyone wants a personal agent. Almost nobody wants to maintain the memory file. The model is the easy part. The hard part is writing down what you actually prefer, then updating it when you change your mind. Without that doc, your "agent" is just a chat with amnesia.

Gemini 4 Carbon: SOTA on benches, SOTA on "you can't use it yet".

Baidu is so great. Yes I said it. Fight me in the replies. In China, Baidu's reputation is... complicated. Memes. Dead search. "just use Google". Then they open-source an OCR model that tries to kill the "chop PDF into pages" agent pipeline. Unlimited OCR 2. MIT. ~3B MoE (~0.5B activated). Still updating. Repo: github.com/baidu/Unlimited-O… Same stack also ships as a Baidu Cloud API if you do not want to self-host. Most agent RAG still does: page loop → OCR → glue → pray reading order survives. Works until a table spans pages. Or a footnote. Or a 200 page contract. v1: R-SWA. Keep the visual reference. Only slide a short window over what you just wrote (~128 tokens). Output-side KV stops growing without bound. v2: SR-SWA. At hundred pages, even the visual side gets expensive. Each decode step keeps a fixed working set. Local window ~128 + sparse visual blocks ≈ 512 active refs. Page gate + token sieve: look at what you are copying, not the whole book. Unlimited-Horizon decoding windows pages and reuses boundary KV. Author claims (soft until third parties reproduce): ~7× cheaper per step than full attention at ~100 pages ~5.6× vs R-SWA peak KV ~47GB → ~1.77GB with windowed decode MarathonDocBench Overall 77.06 (their 100–300 page bench) Why agents care. Document IO is still the boring bottleneck. Bounded attention on long docs means less glue code and less lost cross-page state. Caveats: MarathonDocBench is self-built. Sparse attention can miss global heading structure. Self-host needs GPU + preferably SGLang. Free HF Space is not a speed demo. Hot take: Baidu shipping MIT long-horizon OCR while half the timeline argues chatbot vibes... yeah. Baidu is so great. (For this one repo. Calm down.)

Like playing playstation