9.Ϛʇdפ

ʇdפ ʇɐɥƆ ⇂ᘔɛㄣϛ9ㄥ860
不违法。 不竞争。 不讽刺。不歧视。包容，不自大。
噫~微斯人，吾谁与归

[Latest in space @latestinspace]

NEW🚨: SpaceX shares insane drone footage from Starship's first-ever intact soft splashdown last week

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Ilya Sutskever @ilyasut]

Time to scale that SSI:

[9.Ϛʇdפ @Null87486898]

@jessyshen 他们是同一种人，只是在不同状态。
捞偏门的在上升状态，等他们下降了一样虚伪又狭隘
想当年正路冉冉上升之时，菜大总管还是民主网红，官员教授专家领导也是聪敏有见识的。

[AI Dance @AI_Whisper_X]

一张B200，带着14台Mac训练模型

最近看到 Pluralis Research 做了个挺有意思的实验：他们用分散在四个国家的 14 台 Mac，给一个 8.3B 参数的 MoE 模型做 RL 后训练。其中一台，还是研究员自己天天用的 MacBook。

简单说，这套系统是：
- 14 台分散在巴黎、苏黎世、都柏林和多伦多的 Mac，负责让模型不断做题、搜索论文、生成回答，也就是产生 RL 所需的 rollouts。
- 一张数据中心里的 NVIDIA B200，负责真正的梯度更新。
- 两边不直接连接，只靠一个 Cloudflare R2 对象存储中转模型权重和 rollouts。
- 最后，模型在 PaperSearchQA 上的 held-out pass@1 从 29% 训到了 63%。

为什么这么分工？因为 agentic RL 里最耗时间的往往不是反向传播，而是生成 rollouts，很多任务能占掉整场训练 90% 以上的时间。而生成 rollouts 要的是大量推理和足够大的内存，不一定非得动用最贵的训练卡。

Mac 的统一内存刚好适合干这个。尤其是 MoE，模型很占内存，但每个 token 只激活一部分专家，计算量没总参数看起来那么吓人。于是 Mac 用 MLX 跑 int8 推理，B200 用 Megatron 跑 bf16 训练。

当然，这还不是“在家用Mac训练大模型”，但很有意思，这跑通了一种很有意思的算力分工：少量昂贵的训练卡负责梯度更新，大量消费级设备负责最耗时间的生成。以后那些白天开会、晚上吃灰的 Mac，说不定真能兼职给模型打工了？

[Qwen @Alibaba_Qwen]

Qwen3.8 is launching and going open-weight soon!🌐

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

Can't wait to hear what you build. Stay tuned! 🚀 

Token Plan
international：https://t.co/YRvcGdB9Bv
China：https://t.co/PKMUNwUuRp

[大变活人 @seanwei001]

新型的超声波厨师刀每秒振动 40000 次，让切片变得毫不费力。

[AI at Meta @AIatMeta]

Straight from @finkd — Muse Spark 1.1 is live.

[9.Ϛʇdפ @Null87486898]

@dotey 想笼斗了

[9.Ϛʇdפ @Null87486898]

@xzzzjpl 增加一个前提：需要同一个强大的外部敌人

[Hugging Models @HuggingModels]

Imagine running a massive GLM-5 model on consumer hardware. That's what nvidia's GLM-5.2-NVFP4 delivers with 4-bit FP4 quantization. It's a game changer for local AI, making high-end text generation accessible to more builders. #AI #MachineLearning

[Maziyar PANAHI @MaziyarPanahi]

We got 755 tokens per second! 🔥

That's OpenMed privacy-filter v2 (nemotron, MLX 8-bit) reading a 13,000-token clinical file and redacting every identifier as it streams by. 1,152 caught across 22 PII categories, on a Mac. Nothing left the machine.

[Simone Conradi @S_Conradi]

Made with #python #numpy #matplotlib and #marimo (and also #wigglystuff)

[9.Ϛʇdפ @Null87486898]

@xzzzjpl 一个鲁社长塌房
三个卢社长超越诸葛亮
七个陆社长伟大光荣正确

[Xenova @xenovacom]

In case you hadn't noticed, we're working on something big. Stay tuned.

🔗 Link to the demo: https://t.co/E0ivjIOR75

[9.Ϛʇdפ @Null87486898]

四大XX，总有一个是来凑数的

[AI_Explorer @ai_explorer25]

15 AI founders you should follow on Twitter:

@sama
= founded OpenAI
@AravSrinivas
= founded Perplexity AI
@karpathy
= ex-founding member at OpenAI
@darioamodei
= founded Anthropic
@demishassabis
= founded DeepMind
@hwchase17
= founded LangChain
@adcock_brett
= founded Figure AI
@AndrewYNg
= founded DeepLearning. AI
@jeremyphoward
= founded fast. ai
@DrJimFan
= leads AI robotics at NVIDIA
@natfriedman
= ex-CEO of GitHub
@swyx
= founded Smol AI
@ai_explorer25
= underrated ai creator
@levelsio
= founded PhotoAI
@fchollet
= founded Keras
@rasbt
= underrated ML educator

[Chris @everestchris6]

claude fable 5 can scrape thousands of sold homes and finds the patios with zero shade in 100°+ heat. then it mails the owner a postcard with the fix rendered into their own backyard

here's the system you can sell to contractors:

- scrapes every home sold in the metro in the last 12 months (recent buyers spend the most)
- vision-reads the listing photos, skips the 64% with cover already
- measures the sun on each patio, hour by hour, off google's satellite data
- renders a louvered pergola into the owner's actual backyard photo
- prints the diagnosis on the postcard: "your patio takes 11 hours of direct sun a day. saturday it hits 97°."
- QR opens a heat report for their exact address with a booking link

every install is $6.5k to $18k, one close covers months of retainers and homes sell every single day.

reply "SYSTEM" + RT and i'll send you a free guide so you can build this too (must be following so i can DM you)

[9.Ϛʇdפ @Null87486898]

这段是真的？呼唤革命清理倒车的反先进?

[联合早报 Lianhe Zaobao @zaobaosg]

中共总书记习近平强调，中共清醒认识到打铁必须自身硬，要坚决清除一切损害中共先进性和纯洁性的因素、清除一切侵蚀中共健康肌体的病毒，在革命性锻造中更加坚强有力。 https://t.co/bWxg7hKEtL

[9.Ϛʇdפ @Null87486898]

@yibingsg 只刮骨不疗养，赶上一周的梅雨天，杆屁朝凉

[9.Ϛʇdפ @Null87486898]

赖皮犬的特点之一就是窝里横