天下牛马是一家。遇上这种脑瘫上司就纯属倒大霉
営業部長「Outlookが使えないんだが」
情シス「Microsoft側で障害が発生しています」
営業部長「こんな忙しい時に。いつ直るんだよ?」
情シス「復旧時刻は未定です」
営業部長「はぁ?Microsoftには連絡したのか?」
情シス「公式の障害情報を確認しています」
営業部長「いや、電話しろよ」
情シス「……電話ですか?」
営業部長「そうだ。直接話した方が早いだろ」
情シス「世界規模の障害なので、現在対応中かと」
営業部長「じゃあ責任者を出してもらえ」
情シス「……Microsoftの責任者ですか?」
営業部長「そうだ。こっちは金払ってるんだろ?」
情シス「世界中の企業が払っています」
営業部長「うちは長年使ってるんだ。優先してもらえ」
情シス「……」
営業部長「とにかく電話して、何時に直るか聞いてくれ。それか、うちだけ先に直してもらえるよう交渉しろ」
情シス「世界規模の障害なんですが」
営業部長「そこを何とかするのが情シスの仕事だろ?」
情シス「……分かりました。確認します」
営業部長「最初からそうすればいいんだよ」
情シスは、静かにMicrosoftのサポートページを開きました。
そこには「復旧時刻未定」と書かれていました。
情シスは、しばらくその画面を見つめていました。
その間にも、営業部長からの催促電話は鳴りやみませんでした。
Microsoftの復旧時刻は、誰にも分かりませんでした。
ただ一つ、分かっていることがありました。
営業部長からの次の電話は、必ず来るということです。
本当に怖いのは、Microsoftの障害じゃない。
月額料金を払えば、情シスがMicrosoft本社を動かせると思われていることだった。
1
146
wonjder retweeted
In March, I emailed Jensen when @inferact's compute provider took back our capacity, leaving us unable to develop @vllm_project for Blackwell. He replied within the SAME minute and had the issue resolved overnight.
The entire NVIDIA team, led by Jensen, operates like an unstoppable armada. Jensen moves fast, so is his "exo-skeleton". Everyone in the industry respects and trusts him. That's how you push open intelligence forward.
Thank you Jensen, for saving Inferact and vLLM.
In June @NaderLikeLadder emailed Jensen about a crazy idea we had over dinner to make Local AI 10x better on NVIDIA hardware.
Jensen replied in 7 minutes. He saw this as important work, and approved a war room for us to work on this at NVIDIA HQ.
For a month, NVIDIA HQ was the @exolabs satellite office. We were overwhelmed by the support we got. We were back to back the first week with smart NVIDIANs coming in to our war room to help us. Creators of CUDA, Dynamo, people who built NVIDIA.
Long story short with NVIDIA we managed to improve their Local AI playbooks by 9.6x (almost the 10x goal we set out with). We’ll be publishing a whitepaper on this work.
We were just a small startup with a crazy idea. Jensen took a bet on us. Jensen cares deeply about Open Source AI and Local AI.
I don't believe Jensen is replaceable, he's a 1 of 1 human, but I hope NVIDIA will continue to aggressively support Open Source AI and Local AI, and continue to take bets on small startups with crazy ideas.
Thank you Jensen
NB了,华裔在美国只配当低级打工仔。哪怕老黄这种级别想搞你就搞你,这还玩毛线。
NVIDIA announces CEO transition. Read the release: nvidianews.nvidia.com/news/n…
116
wonjder retweeted
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too?
Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model.
V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0.
Results are coming in live. Link below🧵
现在AI给力之后重写底层框架以达到自家系统最佳适配程度的事情应该会越来越常见
Replying to @cradrrat
We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.
Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right. Now, 4.8 is in RL and the evals are starting to climb. We will do a soft rollout to Bot users later this week and see how it goes.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
144
以33倍的价格来获取5倍的速度提升,在原厂API已经达到200 tok/s的情况下你们觉得有啥适用场景吗?
Replying to @vercel_dev
价格差别很大:这个 @deepseek_ai V4.1 Flash Ultrafast 走 @FireworksAI_HQ 加速端点,每百万 token 输入 $2.40、输出 $9.60;同一 10/17 快照普通版只要 $0.022 / $0.294,输出贵约 33 倍。适合要即时反馈的交互编码,批量任务用普通版更省。来源 vercel.com/ai-gateway/models…
152
我只能说NousResearch真不考虑找个专门的宣发吗?
这个模型本意大概就是移除了Qwen3.8内某些方面的限制再通过低成本后训练就可以更改价值观方向。
现在人们质疑它为什么这么做?难不成西方价值观才是唯一正确的价值观吗?它却只会不断重复让你看博客。
说一句我们是进行低成本后训练实验,你也可以通过后训练低成本改造模型往你所期待的方向发展。这很难吗?博客里就有
目前我看它要不是就是要推崇它的西方价值观,要不就是为了骗点热度点击量。哪怕损失的是它自己的名望
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
1
207
wonjder retweeted
En estos momentos me gustaría ser una rubia puta con onlyfans y que me entren 300 dólares por minuto de pajeros, en cambio soy un pelotudo que me están cojiendo en la bolsa
测试结果和体感一致,CommandCode这么没格局吗
I was contacted by the CEO of @CommandCodeAI and I'm not really sure how I feel about it...
I'm not a journalist, I'm just a regular user. I wanted to compare services on completely neutral benchmarks. I do not have a stake in the success or failure of any of these companies. However, I'm now being blamed for trying to release benchmarks because I believed that the community deserves real numbers on neutral grounds.
I've acted in good faith to get to a solution and offered to correct ANY incorrect configuration I may have had and also re-run any tests with corrected configurations and post those. I've offered to run future tests in other agents (Pi), and also native agents (future tests). I offered to wait for backend issues to be patched. I've offered to install workaround plugins. That's more than I've done for any provider.
I'm accused of being "outrageously wrong", my methodology being "bad faith", "publicly bashing" the company, and not showing empathy (disheartening, because I offered hours of my time to help when I had no obligation to).
Their claim: the backend has safeguards "against memory leaks or memory overloading sessions," and those blocked my requests. We can all agree the harness may be inefficient, and if you believe that your guards are right, then let's leave it there. I use Pi day to day. I don't mind saying the harness is unoptimized, it's not even v1 yet. But don't say the benchmark is bad faith, don't say it's "still your issue".
github.com/CommandCodeAI/com…
Every provider ran the same harness, the same prompt, at the same time. Only one failed. If your service supports using API keys, expect others to use different clients. It's your feature. Disable the use of third party SDKs and clients if you think they're unoptimized and paint your service in a bad light. Else, it IS your responsibility to optimize for them. Especially a harness with 250k+ stars on github and the official harness for the model. After all, I'm testing the model, not the harness.
So I'm done. Command Code won't be part of my future benchmarks. The sessions are public for anyone who wants to check my work, and the community is free to @ me and tell me that I've incorrectly done these benchmarks so I can fix them: github.com/nazzysar/commandc… (FYI, I sanitized out any sensitive info, you can use agents to just read through).
My conscience is clear.
1
196
家里有GB300的可以直接卖钱了
Beyond matching DeepInfer's performance, we also put vLLM to the test on @SemiAnalysis_ AgentX with real-world agentic workloads.
At DeepSeek's off-peak pricing, vLLM delivers $38.96/GPU-hour revenue, with 954K tokens/s/GPU total throughput (cached+input+output) at 196 tokens/s P90 per user.
This is the best DeepSeek-V4.1-Flash result on AgentX ever, powered by recent advancement in vLLM!
Check out our blog for the full optimization story and how to reproduce the results 👇 Enjoy the read.
145























