wharton

Hy3是国内当前最好用的，没有之一。

@python_xxt 我尝试vibe输入法才发现这玩意多复杂

我有一个疑问： 输入法的移植很困难吗 ？ 豆包输入法的Android 、iOS、MacOS出来都很久了 但是Windows 端迟迟没有发布。 在有LLM加成的这个时代，我着实有点不理解 有没有大佬，能够解答一二 ？

meta 的AI真拉垮，用起来啥也不是。

@aliez_ren 上下文才是最重要的质量保证，只有速度对效率而言还差一半

这下真的舒服了，本地 70 tps 跑 GLM 5.2 几乎满血版！ huggingface.co/madeby561/GLM-…

rtx3060+qwen3.6 35B A3B,用llama.cpp能达到稳定的tg 30+，而且智商够用，本地模型能到这个程度真是很不错了，希望还能进步

A few patterns we frequently use with Fable 5: Use Fable 5 as an "advisor." An executor (Sonnet 5) calls Fable 5 for guidance. Most tokens are billed at the lower executor rate.

1M tokens burned on gemma4-12B-QAT 🔥 Another milestone for my local llm server, running on 3060/3090. Top 3 models by total tokens processed: Qwen-3.6-35B (Q5,Q6,Q8) - 7M Qwen-3.6-27B (Q4,Q5) - 1.5M Gemma4-12B-QAT - 1M Getting close to 10M mark now. This includes context tokens, not just generated output. And even then, it's probably on the lower side, actual usage is most likely higher. Thanks to @NousResearch Hermes and the Pi coding agent, which are where most of these tokens are being burned. I’m increasingly leaning toward local models for most of my day to day tasks, and the usage is starting to reflect that.

Loop Engineering 现在很多人用 Claude Code、Codex 或者 Cursor 的时候， 还跟聊天机器人一样操作： 先扔个 prompt → 等它回 → 复制出来 → 改改 bug → 再扔新 prompt……循环往复。 这个 repo 直接告诉你： 别再自己一直 prompt 了， 你去设计一个 loop， 让这个 loop 自动去给 agent 下指令、指挥它干活。 里面已经帮你准备好好几种现成的 loop，比如： • 每天任务分流 loop • PR 自动审查 loop • CI 检查 loop • 依赖更新 loop • 写 changelog loop • 合并后自动清理 loop • issue 分类处理 loop 还附带了好几个 CLI 工具，能让你： • 把 loop 放大规模 • 估算大概花多少 token • 检查你的 repo 适不适合用这套东西 • 给 agent 加记忆/状态 • 加人工接管环节 • 加验证关卡 • 还能安全地通过 GitHub Actions 跑 以前 Prompt Engineering 是教你怎么写好指令， 现在 Loop Engineering 是让你搭一个系统， 让 agent 自己跑、自己检查、自己改、自己扩展， 你只要在关键节点看看就行。 github.com/cobusgreyling/…

免费试用Fable 5， 点这个链接，咱们两个都有5美元的额度zenmux.ai/invite/ZZI3D8

pi+herdr+lakebed clutch af 🐀

ClaudeflareでGLM5.2無料で使えるヤツ、秒で設定できた。クレカもなんもいらんから楽。 Claudeflareログイン Workers AIクリック REST APIクリック Create a Workers AI APIToken クリック 適当に入力してAPIキーゲット アカウントIDコピー Chatboxで試す場合、OpenAI API互換カスタムAPIから新規選ぶ APIキー さっきのやつ APIホスト https:　//api.cloudflare.com/client/v4/accounts/ここにアカウントIDコピー/ai/v1 ※スペース抜いてね 新規モデルから モデル名 @cf/zai-org/glm-5.2 だけで動いた。けど1日の制限は厳しそう。使い放題とは程遠い感じ。

转发一下 B 站博主的锐评 PPT skills： 注意：有些 skill 不是专门做 PPT 的，所以评分会有点低，只是需求不同，想专门做 PPT 的看最前面的。 1. hugohe( 3.1万 star) | 顶级天花板 👑 全场唯一正经做 PPT 的！元素全可编辑，自带音色克隆和旁白生成，纯纯降维打击。 🔗github.com/hugohe3/ppt-ma… 2. 张咋啦 @zarazhangrui ( 2.3 万 star) 顶级主观审美最佳，完成度极高！目前呈现为 HTML 格式，稍微考验一点使用者的基础。🔗github.com/zarazhangrui/f… 3. 花叔 @AlchainHust ( 1.9 万 star) 顶级审美极佳，关键是能输出可编辑的 PPTX 格式！ 🔗github.com/alchaincyf/hua… 4. 歸藏 @op7418 ( 1.5 万 star) 人上人瑞士风审美非常棒，自带快捷键，很适合线下分享。 🔗 github.com/op7418/guizang… 5. Lewis ( 6500 star) 功能设计十分贴心，自带计时器和逐字稿等实用组件。github.com/lewislulu/html… 6. 宝玉 ( 2.2 万 star) NPC风格偏可爱，主要以纯图片形式呈现。 🔗 github.com/JimLiu/baoyu-s… 7. 乔木 ( 5400 star) 偏向于纯图片卡片的输出，更侧重于内容的初步呈现。 🔗 github.com/joeseesun/qiao…

minimax code是我尝试各家agent code中最难用的。它比这些都差：codex kiro cline agy crush zcode kilocode agent (cursor) mimo

There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL. RLing an agentic coding model isn’t rocket science. In simplified terms: 1. RL needs trajectories — rollouts where the model actually completed a task in some env 2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem 3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet 4. Now it produces positive trajectories on those tasks 5. RL on those trajectories and hill climb agentic coding 6. At that point you no longer need to distill and can solely hill climb RL to better models This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier. GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.

用下来感觉mimo code有点被低估了，还是说glm 5.2太拥挤导致使用起来感觉降智了

This is a new paradigm for interacting with Claude that is significantly more "inline" with all the other human activity org-wide. Once you do all of the under the hood engineering work to make this "just work" (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads. Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It really takes a while to wrap your head around it, but it works and it is awesome.

NVDIA：Rubin全面液冷技术

Tutorial on how to use GLM-5.2 in Claude Code (bookmark this) ~4.5x faster & ~5x cheaper compared to Opus 4.8! 1. Install the latest Claude Code npm install -g @anthropic-ai/claude-code 2. Create an account at baseten.co. 3. Grab an API Key from app.baseten.co/settings/api_k…. Save it for the next step. 4. Edit ~/.claude/settings.json. Open with vim or another editor. Paste the following with your key. "env": { "ANTHROPIC_AUTH_TOKEN": “your_baseten_api_key", "ANTHROPIC_BASE_URL": "inference.baseten.co", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "zai-org/GLM-5.2", "ANTHROPIC_DEFAULT_SONNET_MODEL": "zai-org/GLM-5.2", "ANTHROPIC_DEFAULT_OPUS_MODEL": "zai-org/GLM-5.2" } 5. Enjoy GLM-5.2 in CC!

BERNSTEIN: Global Memory > Analysts forecast a 2-2.5x increase in HBM Average ASP year-over-year (YoY) heading into 2027. This price hike is expected across all generations of HBM (HBM3, HBM3E, HBM4, etc.). > DRAM industry prices are forecasted to continue rising after 2QCY26 (though at a slowing quarter-over-quarter rate) and are expected to hit their cyclical peak in CY2027. > Conventional DRAM currently generates considerably more revenue per wafer capacity than HBM. > The gross margin gap between conventional DRAM and HBM is expected to significantly contract by 2027, with conventional DRAM GM approaching the mid-90% range and HBM GM climbing near 90%. > Samsung is projected to gain HBM market share by bit shipment (climbing from 28% in CY2025 to 46% by CY2027E), driven by better HBM4 performance and increased capacity. > Due to the massive expected price increases, HBM's contribution to total DRAM revenue is forecasted to experience a sharp, V-shaped rebound for all major players starting next year, hitting an industry total of nearly 40% by CY2027E and up to ~45% by CY2028E. > Across all three major memory makers, earnings are modeled to reach a peak in the second half of calendar year 2027 (2HCY27). Samsung's EPS is expected to spike just past 20,000 KRW. SK hynix's EPS is projected to max out near 150,000 KRW. Micron's Non-GAAP quarterly EPS is forecasted to push close to $47.00–$50.00 USD.