jupiter @jupiter186

We use GLM-5.3-Flash build a dream kitchen. A 3D world built in Blender. This is not a generated video.

这个台词水准有点高，三四层楼那么高。👍

x.com/i/article/2093…

Inspired by @percyliang's CS336, I built a static performance model for LLM inference — no dynamic batching/chunking, just clean analytical bounds. Pick model × GPU × batch × seq length × parallelism(DP/TP/EP/PP)  → VRAM check + TTFT + TPOT + prefill/decode breakdown + throughput vs batch size. Covers 5 KV cache variants (GQA/MLA/SSM/sliding window/linear attention), speculative decoding, multiple quant precisions. Calibrated against ~100 public benchmarks (TRT-LLM, Splitwise, MLPerf, Koyeb…). Demo: …inference-calculator-delta.vercel.app Repo: github.com/pochenai/llm-i…

Top open models by use case 👇 chat - Kimi K3, Qwen3.8 2.4T A95B reasoning - Kimi K3, DeepSeek V4 Pro coding agents - Kimi K3, DeepSeek V4 Flash, GLM 5.2 small + fast - Qwen3.8 27B, gpt-oss 20B mid-size general purpose - DeepSeek V4 Flash, MiniMax M3 function calling - DeepSeek V4 Flash, GLM 5.2 vision - Qwen3.8 2.4T A95B, Kimi K3, MiniMax M3 speech-to-text - NVIDIA Nemotron 3.5 ASR, Whisper Large V3

Hotchips 2026 干货和私货分享 - 内存 - 知乎 zhuanlan.zhihu.com/p/207561757526…

From Zero to GPU: A Guide to Building and Scaling Production-Ready CUDA Kernels huggingface.co/blog/kernel-bu…

我小侄子大学三本毕业了，看了四年剧，天天吃外卖，问我有没有赚钱好的路子？ 我说那肯定有，体面三件套。考研、考公、考编，他摇了摇头说太卷了，读书太累考不了。 我说那就牛马三件套。进厂拉货、做销售，他说想自由点的。 我说那就铁人三项。快递、外卖、网约车，他说有没有安稳点的。 我说那就只有吉祥三宝：保安、保洁、保姆。他说来钱太慢了，没有前途。 他说就没有可以赚快钱的吗？我说有浮动三件套：基金、股票、期货。他说风险太大了，不敢碰。 他说有没有可以躺平就能赚钱的？我说也有躺平三件套：收租、啃老、富二代。他说家庭条件又不是不清楚，哪里有资本？ 我说那就只有走捷径三件套：找富婆、认干妈、吃软饭。他说干不了，性格内向，吃不了那碗饭，还跟我生气。 他说没有骨气的事不干，还不死心的问我，难道就没有一夜暴富的吗？ 我说有做梦三件套：体彩、福彩、刮刮乐。几十条路摆在你面前，总有一条适合你，他转身就走了，似乎下定了某种决心，再也没有回头看我一眼。社会的毒打会比来自校园的毒打更加刻骨铭心。

ilovelife @ilovek8s

a week ago

我小侄子高考落榜了，考了236分，问我有没有赚钱好的路子？ 我说那肯定有，体面三件套。考研、考公、考编，他摇了摇头说学历太低了，考不了。 我说那就牛马三件套。进厂拉货、做销售，他说想自由点的。 我说那就铁人三项。快递、外卖、网约车，他说有没有安稳点的。

92 89 443 95K 259

GLM-5.3 vs Fable 5 breakdown of DeepSWE performance. Coding quality parity with 1/4th the cost is a sweet deal! > GLM-5.3 at 69.0% vs. 69.7% > $3.99/task vs $21/task > 80k tokens vs 114k tokens > 124 turns vs 85

最近爆火的OpenRouter 上冒出一个匿名模型 Ox Alpha（GLM），其实管道不会撒谎，会撒谎的是人 Crawl4AI的作者unclecode这句话,道破了怎么验一个API背后到底是哪家的模型 厂商在别名后面偷偷换模型,不会告诉你。他两天前就发现DeepSeek的deepseek-chat这个老名字下,悄悄换成了另一个模型 他的思路是不听模型自己说,只看基础设施的指纹,九个探针: 01｜喂英文全字母句、中文段落、代码、emoji,只数prompt_tokens。分词器是每家自己训的,中文分词差异最大,藏不住 02｜temperature设成2.0,看报错原文逐字。那句话是这家工程师亲手写的 03｜GLM的1301错误码、finish_reason的词表、上下文天花板是1048576还是262144,每一项都能定位到具体厂商和变体 前几天OpenRouter上那个匿名模型Ox Alpha,全网都在猜。modelprint一测,分词器四项归一化计数全中,直接指向GLM家,其它厂商最好也只中一半 纯前端无后端,API key只留在你自己的浏览器标签页,不经过任何服务器

吴思先生的《血酬定律》

Jeff Li @jefflijun

2 weeks ago

好像是有这样的情况...Compromise也是一种艺术

542 131 2K 289K 410

15分钟一局的王者荣耀，成了数学博士生的避难所。 mp.weixin.qq.com/s/kUWXxWqfryq2…

Read full blog: lmsys.org/blog/2026-08-1…

别人给她台阶是让她下的，不是让她上的。 所以有些人被讨厌是有原因的。

Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions. The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed. Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter. Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it. This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count. Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.

大模型是残酷的制造业 mp.weixin.qq.com/s/I-U0lOHs_uHf…

回顾90年代中国改革与“江朱体制” | China's 1990s reforms and the legacy of the 'Jiang-Z... youtu.be/maM7D1BffT0?si… via @YouTube

Ten advances in mathematics and theoretical computer science | OpenAI openai.com/index/ten-adva…

Mcore Bridge：迈向Megatron训练"零门槛"时代_魔搭ModelScope社区-ModelScope魔搭社区 modelscope.csdn.net/691185110e4c46…

何清漣專欄：潘石屹追稅風波——資本跨境套利時代結束 - 鏡報 mirrordaily.news/story/79231