Yiqun Wang

[Yiqun Wang @PatrickWang2022]

@711Cypher @MaxForAI 哈哈哈哈就是

[Max For AI @MaxForAI]

🚨据WIRED 披露，特朗普一个电话，直接打断了英伟达全员会。

上周四黄仁勋正在英伟达总部开全员会，现场还有数千名员工在线收看。

讲到一半，黄仁勋手机突然响了。

电话那头是特朗普。

黄仁勋直接走到一旁接电话，但现场麦克风还开着，员工甚至能听出特朗普的声音。不到一分钟后，他挂掉电话，又回来继续开会。

几个小时后，特朗普就在 Truth Social 上公开祝贺英伟达最新财报。

更巧的是，就在同一天，英伟达正式成立了自己的联邦政治行动委员会 NVPAC。今年英伟达在华盛顿的游说支出也已经超过250万美元。

现在的英伟达，显然早就不只是一家芯片公司了。

当一家接近5万亿美元市值、控制全球AI核心算力供给的公司，CEO开全员会都可能被美国总统一个电话直接叫走，这个画面本身已经很能说明问题了。

黄仁勋现在每天处理的，可能已经一半是GPU，一半是白宫的政治。

[Uber Engineering @UberEng]

https://t.co/4E91KTbzDU

[孙宇晨（去过太空版）🧑‍🚀 @sunyuchentron]

我的女友景甜

一颗卵子的重量，三点五微克。
五千万美元现金的重量，两点五吨。
景甜在蒙太奇拉古纳海滩的电话里向我要的是后者，抵押的是前者。
她去那里是为了取卵。代孕的孩子预计2027年出生，快一点的话还能属马。这是她主动提的。到了诊所，她说没有五千万，不取。

蒙太奇拉古纳海滩在海崖上，底下是太平洋。那一层包下来，三十天，一百万美元，为了她的隐私。她住了八天就走了。
剩下的二十二天，酒店的清洁工每天照常上去。是我的助理讲的，她去交钥匙的时候撞见过一次：一个墨西哥女人推着车，一间一间进，换毛巾，擦台面，把被角掖好，退出来把门带上。做完一整层，两个小时。
助理问她，没人住，为什么还要做。
她说，登记上写着有人。
一座失去皇后的凡尔赛宫。
我突然意识到，景甜喜欢这个。景甜喜欢一个地方因为她而空着。什么都不发生。
与蒙太奇拉古纳海滩一起空置的，还有一架湾流G700尴尬地停在Van Nuys，机组每天照常报到。
第八天她打了那个电话。之后没有人再见过她。

---

2007年，已经有人替她烧钱了。那年我刚进北大。
当然，烧钱的不是我。
QQ弹出一个窗口，代言人是她。就在那台联想IBM的老机器上，分辨率糊得很。我把图片存下来了。
我在一年房租一千块的北大宿舍上网，在校内网上找到她，发了好友申请。校内网要填一句留言，我写了删，删了写，弄了大概四十分钟，最后发出去一句“你好”。
她没有通过。
她也许根本没用过校内网。
后来这些年，我做每一件事的时候都在想一个问题：到了什么程度，她会通过。第一个十亿的时候我想过，一百亿的时候我想过。
后来我不想了。人人网早就倒闭了。
我没有再申请第二次。
2011年，北京冷。我买了一件羽绒服，一百五十块。我记得这个价格，因为我犹豫了三天。
同一年她拍了一部一亿五千万的电影《战国》。孙红雷、吴镇宇、金喜善、中井贵一、姜武就像十九年后的蒙太奇拉古纳海滩与湾流G700一样被请来做了陪衬，他们不知道自己为什么出现在这里，来都来了，总得发生些什么，但是什么都没有发生。
再后来是一亿五千万美元的《长城》。张艺谋，刘德华，马特·达蒙。再后来是成龙。景甜宇宙的名单越来越长。
我那时已经在美国了，读书，创业，越来越忙。唯一没变的是，每换一台设备，无论多麻烦，我都把那张QQ图片导过去。
十九年后，名单上加了我。

---

瑰丽酒店我们见了第一面。我一直以为见到她那天我会紧张。
我没有。
她比照片上瘦，穿得很随意，靠在沙发上。我看着她，忽然想不起来我等的到底是什么。
我带她去看卡特兰的香蕉。她看了很久，问我，这个多少钱。
六百二十万美元。
她说，那它烂了怎么办。
我沉默了。她说，不管了，我们去看电影吧。
我说好。
她说放映厅里除了我们两个，不能有任何人。
我说好。
我出去打了个电话。回来的时候她还站在那根香蕉前面。
那部电影是《疯狂动物城2》。
两个人的放映厅，这种场景一般用于中宣部审片。
我让助理把电影院所有的座位都买下来。
我看过批判刷假票房的文章，说票买光了只有两个人看。现在我明白了，也许不全是为了票房。
把座位全买下来不难，麻烦的是已经卖出去的票。助理和保镖提前到影院门口，遇到买了票的观众，把现金递过去。
一对带孩子的夫妻来了，孩子已经在门口高高举着爆米花。
助理递上一沓钱，孩子不懂为什么不能进去，大人接过钱，道谢，什么也没问，走了。
那天领钱的有二十六个人。
看完她把微信头像换成了朱迪警官。
晚上我用手揉她的奶子，被她推开了。
我以为她不愿意。我把手收回来，准备说点什么。
她没说话。她坐起来，从包里翻出一个小袋子，拉开，里面一排东西，像一套很小的工具。
她说，你的指甲扎得我痛。
我看了一眼。我一直是拿指甲刀随便剪两下，边上是尖的，平时不觉得。
她说，过来。
她把我的手放在她腿上，一根一根磨。
先是粗的那一面，磨完换一面，再换一面。磨一会儿举起来对着灯看看，再接着磨。
我说这个要磨多久。
她说，磨到不刮人为止。
十个手指，她花了快一个小时。
中间她教我，这一面磨形状，这一面抛光，最后一面不要来回蹭，要顺着一个方向。
我说好。
磨完她把我的手翻过来看了看，摸了一下。
她说，行了。
摸吧，想摸哪都行。
中间她说过一句话。她说，要是以后我不在了，就再没有人给你抛指甲了。
我笑了一下，没说话。
认识她三个月的时候，我让人做了一份东西。她那些年的项目、合作方、公司股权、景国庆，田心爱，所有的几个关联账户，北京租的房，上海买的房，还有一些别的。一周就做完了，一共四十几页。还有些男男女女的事情，我觉得这是最不重要的。
里面有一些事她后来跟我讲过，讲的和上面写的不太一样。我没有说。
有一些事她一直没讲。我也没有问。
我全部相信她说的。
那份东西我留着，一直留到现在。
但是我一个字也不信。

---

[fin @fi56622380]

https://t.co/Wrfh9lUhOa

[Yiqun Wang @PatrickWang2022]

@steviegmarshall @Thom_Wolf I was going to order. Saw the duties, decided to not buy

[clem 🤗 @ClementDelangue]

BIG ANNOUNCEMENT FROM HUGGING FACE TODAY:

We're unveiling Microduck 🐥🤖

It's a tiny $399 open-source robot you can teach new tricks with reinforcement learning. It can walk, pick things up, get back up when it falls, and even roller-skate.

Welcome to the era of open-source affordable robots to democratize physical AI and world models!

🤗🤗🤗

[avrl ☘ @avrldotdev]

Anthropic Forward Deployed Engineer Interview Guide 2026

[almonk @almonk]

The worst part of the otherwise excellent Logitech mice & keyboards has always been the software. Nice to see an open source, local-only desktop app alternative launch https://t.co/H3R2aR18ge

[Shay Boloor @StockSavvyShay]

The AI compute market is splitting into two very different businesses with neoclouds like $CRWV & $NBIS pricing the same GPUs far below hyperscalers like $AMZN, $MSFT & $ORCL.

At the same time you can see how newer chips are still commanding higher hourly rates while delivering much more output so customers can pay more per GPU while still seeing cost per token fall sharply.

And that’s why availability matters just as much as price because the providers that can actually get the newest GPUs online are the ones with the most leverage.

[☀️AliquisNovus☀️ @PalmyrPar]

>runs a hedge fund
>doesn’t hedge

[Percy Liang @percyliang]

In addition to forecasting our final loss, we can also forecast the loss of intermediate checkpoints, so we can see if we’re on track over the voyage:

[Andrew James @ThesilenceJames]

@HazelAppleyard Oh but we can and have, as the text message illustrates.

[kuno @kunoo]

Nobody is an atheist at 50x leverage.

[TrendSpider @TrendSpider]

RIP 🪦

[Tesla @Tesla]

Like having a fighter pilot behind the wheel

[atomic.chat @atomic_chat_hq]

Run Qwen3.8 27B locally via Atomic Chat💥

We released Atomic Dynamic GGUF quants, from 8-bit (28.9 GB) down to 1-bit (8.5 GB), and measured all other Qwen3.8 GGUFs in the community

AD-IQ3_S runs on a 16GB MacBook Air and picks the same next token as the BF16 original 92.4% of the time

[Margo Martin @MargoMartin47]

Walked in to @PressSec typing with one hand, and feeding her son with the other 💙

Super Mom!

[Yume_X @yume_arasaki]

I will teach you how to run Qwen 3.8 27B Dense at its optimal configuration.

If you have an RTX 3090, 4090, or 5090, you can now have frontier-level AI on your desk.

The model is free, open source, Apache 2.0. But the defaults are not the optimum. The community spent the first 24 hours digging the real config out of it, and a handful of flags now separate "it runs" from "it runs right." Here is each one and why it exists.

The one that matters most.

--spec-type draft-mtp

Qwen trained a draft head directly into the weights. A small attached brain guesses the next couple of tokens, the big model checks all guesses in one pass, every accepted guess is a free token. The head already ships inside the GGUF you downloaded. You do not download a drafter, you do not build anything. Someone found unused tensors in the server logs at 2am, tried to build the draft file, and discovered there was nothing to build. One flag connects what is already there (sudoingX found this, paired A/B, open sourced the probe before sunrise).

The depth cap. The head has exactly one layer. n=4 breaks it.

--spec-draft-n-max 2

n=2 is the sweet spot. n=3 is the ceiling. The model has one MTP layer, so pushing the draft depth to 4 or 5 crashes the head and it starts emitting junk tokens. People hit this on the Spark and documented the whole ladder: n=1 gives 1.75x, n=2 gives 2.37x, n=3 gives 2.85x, n=4 does not exist. Respect the cap.

The memory flags. MTP brings its own luggage.

--cache-type-k q8_0 --cache-type-v q8_0
--spec-draft-type-k q8_0 --spec-draft-type-v q8_0
-np 1

Three flags, one purpose: fit it on 24GB.

The KV cache is the model's running memory of your conversation, and it is the thing that eats your card at long context. q8_0 halves it with no visible quality cost.

The second line does the same for the draft head's own cache, which defaults to full fat and quietly eats 2GB.

And parallel slots set to 1 means requests queue instead of reserving a second pool. Single card, single lane, everything fits (AJ runs this exact trio on a 3090).

The quality flag. Past 100K the model gets dumb, this is the fix.

--kv-cache-dtype bfloat16

The quantized cache saves memory but degrades reasoning at long context. One person ran it all day past half the window and called the full precision fix night and day. Slight tok/s cost, real quality gain. If your sessions stay short, skip it. If you live past 100K, do not.

The trap that generates "this quant is broken" reports.

--jinja

Qwen 3.8 ships its own chat template. Load the model without this flag and there is no reliable marker for where your turn ends and its answer begins. Two failure modes: it rambles past the stop token, or it answers clipped and loses the thread between turns. Both look like a broken quant. It is not the quant. Several packs now ship a corrected template file because the official one nests empty think blocks across turns.

The Blackwell lane, if you own a 50-series or a Spark.

NVFP4 instead of GGUF. The MTP flag translates to --speculative-config '{"method":"mtp","num_speculative_tokens":3}', same cap. FP8 KV cache doubles your context window (a full 1M token session costs about 32GB of cache).

Two gotchas documented in the first 24 hours: stock vLLM cannot load this model's MTP architecture on a Spark, you need the community GB10 build. And FP8 KV requires a specific attention backend on the Spark, the default one silently cannot serve it.

Set reasoning to medium unless you want it thinking at maximum depth on every reply. Default is xhigh and it burns your tokens.

None of these came from the model card. Every one came from someone's server log, 2am session, or paired benchmark. Flip the flags, then come tell the community table what your card did.

Drop in parameter flags and sources for your technical DD in reply 👇

[Serio_ai @Multi_Serio_Ai]

Qwen 3.8 27b를 쓰시는 분들은 이 이미지 저장해 두시면 좋겠습니다.

- 다들 Q4_K_S 많이 쓰실 거 같은데, 여유가 있어 램 용량 4G 정도 더 쓰면 BF 16기준 정확도를 약 3% 더 끌어올릴 수 있습니다.
- Kv 캐쉬는 Q4가 아니라 Q8을 쓰면 속도와 정확도의 균형을 잡을 수 있습니다.