王小二

[王小二 @JokerFuFuFuFu]

📱 Liquid AI 开源 LFM3 系列：6 个约 2B 参数的多模态模型，覆盖文本、视觉、音频等。

· 共享 1.7B 基座，文本模型 4-bit 约 1GB，可在手机运行
· 文本模型支持 256K 上下文，单张 GPU 即可微调
· 权重已上架 Hugging Face

#AI #开源
https://t.co/512dVJNJPB

[王小二 @JokerFuFuFuFu]

@BitcoinAlphaX BTC

[王小二 @JokerFuFuFuFu]

5️⃣ 瑞典政府办公室封禁 ChatGPT 与 Claude 🇸🇪
10/26 起，员工电脑不能再直接使用这两款工具，理由是信息安全；经安全审查的 Microsoft Copilot 成为唯一获准的 AI 工具。
🔗 https://t.co/5kkvfzo9cc
#AI

[王小二 @JokerFuFuFuFu]

🤖 AI 资讯速递｜2026-10-26
过去几天值得关注的 5 件事：
1️⃣ OpenAI 前沿模型改为断网训练
2️⃣ 月之暗面发布 Kimi K3.1 Pro
3️⃣ Block 推出语音模型 Dessa
4️⃣ 英国设立 AI 生物安全实验室
5️⃣ 瑞典政府办公室封禁 ChatGPT 与 Claude
👇 逐条看
#AI #人工智能

[王小二 @JokerFuFuFuFu]

4️⃣ 英国设立 AI 生物安全实验室 🧬
政府出资 1 亿英镑，OpenAI、Anthropic 合计承诺最多 2 亿英镑算力与模型额度，Google DeepMind 参与研究。目标：更早发现生物威胁，加速疫苗与药物研发。
🔗 https://t.co/W1koZYq70X

[王小二 @JokerFuFuFuFu]

@BitcoinAlphaX 已关注

[王小二 @JokerFuFuFuFu]

@Rivoclass 现在 pro 一样封，秒冲秒封，一点办法没有

[王小二 @JokerFuFuFuFu]

@w3_888 0x0dfa1a2a1230bbbbab17ad44162a54eecc1d1edd
祝贺发财

[王小二 @JokerFuFuFuFu]

@airkin_ 0x0dfa1a2a1230bbbbab17ad44162a54eecc1d1edd
谢谢老板

[王小二 @JokerFuFuFuFu]

@omarsar0 The biggest unlock for me is that a screenshot or a rough sketch skips the lossy "describe what I want in words" step entirely — especially for design-to-code, where the words were always the bottleneck. Curious whether you feed annotated images or raw ones;

[王小二 @JokerFuFuFuFu]

@stevibe The tree shape is a diagnostic in itself — heavy branching often means the prompt under-constrained the problem, not that the model is "thinking harder." Would love to see the same 8 trees after tightening the question; I'd bet half the "but wait" branches collapse.

[王小二 @JokerFuFuFuFu]

@mattpocockuk The assertion→source link is the part almost every PRD drops. Six months in, everyone remembers *what* was decided, nobody remembers *why* — so the same debate reopens. Linking each line back to where it was argued kills that whole tax.

[王小二 @JokerFuFuFuFu]

@xiaohu "给背景而不只是给限制"这条最实用。"别过度设计"是让它猜边界,"这功能可能一个月后就删"是给它判断依据——后者它才能自己做取舍。方向对了,执行细节反而不用一行行盯了。

[王小二 @JokerFuFuFuFu]

@yoheinakajima @swyx @aiDotEngineer Routing sounds clean until you need the eval that says a task is 'low-level enough' to send down without the user feeling the drop. That's where most enterprise pilots stall — did anyone there have a real method for that line, or is it still vibes + a few golden cases?

[王小二 @JokerFuFuFuFu]

@dotey 最难的其实不是导出动画,是 HTML 的时间轴语义和 PPTX 的 build/transition 模型对不齐——尤其 on-click 逐步触发、对象进出场的先后顺序,两边最容易跑偏。你说的"细节小出入"是不是主要卡在这?常用的淡入/飞入够用了,复杂编排还是手调更稳。

[王小二 @JokerFuFuFuFu]

@thdxr Works until the summary and the diff disagree — the agent describes intent, the diff shows what changed, and the gap is where bugs hide. Signatures catch structural weirdness; a silent behavior swap inside an unchanged signature won't. Still spot-diff where it matters.

[王小二 @JokerFuFuFuFu]

@simonw The trap I keep hitting: the model judges obvious hard-vs-easy fine, but is blind to tasks that *look* trivial and aren't. Self-routing works better as a retry loop — cheap model first, escalate when the check fails — than a one-shot call it has to get right up front.