milong

一个无情点赞、转发、评论的Dirty Boy

[milong @simplemilong]

@slackerchat @MaxForAI 小红书有一个叫点点 AI

[milong @simplemilong]

最近很多人用 Claude 做动画，咱们的@qoder 当然也要出场比试一下喽。

[milong @simplemilong]

唯有积极投身其中才有可能真正理解这一变革的路径和意义。如何让Agent 从“能写代码”走向“可靠交付”所需要的系统性工程建设，阿里给出了自己的方法论：《AI Native 研发范式实践手册》

https://t.co/qwrwjeUmtJ

[milong @simplemilong]

@Jackywine ChatGPT 发明人出手，大模型时代终结。

[milong @simplemilong]

@realNyarime 管他叫什么 好用 好用就行了

[milong @simplemilong]

@zhishiai6 @Khazix0918 业务其实还是在微信，只是公司的人办公在飞书了，他这个应该更利好领导们利用飞书 AI 总结员工的一些工作等内容🤣

[milong @simplemilong]

@abskoop 是的，logo豆一样

[milong @simplemilong]

@xRog3r 你真的用过吗？我只能说这个搭配很垃圾。这两周我一直在用，效果是连 opencode+mimo ≈cline+53fladh 远远大于 zcode+53flash，问题不限于失忆，skill 遵循能力，记忆，速度等，打死我也不会用了，即使免费，只是在浪费我的时间

[milong @simplemilong]

@lxfater 现在还有人在用 skill 吗？🤣

[milong @simplemilong]

@Vincent_AINotes 是一个免费给我用我会用的模型

[milong @simplemilong]

我宣布，混元站起来了，超越 claude fable5 和 glm5.3。

[Arena.ai @arena]

Big news: Hy4 preview by @TencentHunyuan just landed ~#5 in the Code Arena: WebDev with 1633 pts (AutoEval).

This is a significant improvement from Hy3 at #31 overall (+115 pts)! Among open models, Hy4 preview is ~#3, compared to Hy3 at #7.

Note: this is an early AutoEval score, in which a Reward Model trained on Arena's human preference data casts automatic votes in place of live votes. We’ll continue to see how scores converge as more live human votes come in. See thread for more info on the methodology behind AutoEval.

Congrats to the @TencentHunyuan team on this release!

[Arena.ai @arena]

Big news: Hy4 preview by @TencentHunyuan just landed ~#5 in the Code Arena: WebDev with 1633 pts (AutoEval).

This is a significant improvement from Hy3 at #31 overall (+115 pts)! Among open models, Hy4 preview is ~#3, compared to Hy3 at #7.

Note: this is an early AutoEval score, in which a Reward Model trained on Arena's human preference data casts automatic votes in place of live votes. We’ll continue to see how scores converge as more live human votes come in. See thread for more info on the methodology behind AutoEval.

Congrats to the @TencentHunyuan team on this release!

[meng shao @shao__meng]

Qwen3.8-Flash 正式发布，
Qwen3.8-Flash-Next 正式开源！

Qwen 官方把 Qwen3.8-Flash-Next 定义为 Qwen4 架构的早期预览，提前开源的目的，是让 vLLM、SGLang 等推理框架和基础��施开发者在 Qwen4 正式家族发布前完成适配。
https://t.co/Of8v22tZKj

效率方面：相对上一代旗舰 Qwen3.7-Plus（397B/A17B），Flash-Next 以 1/3 激活参数、1/3 训练 token、约 1/9 训练 FLOPs，在 14 个预训练基准上 8 个领先、其余最多落后 2.6 分。

稳定性方面：报告给出了可复现的压力测试方法（恒定学习率加倍模拟大规模失稳）：在 4 倍最优学习率下，旧架构（AdamW + Qwen3.5 结构）频繁 loss 尖峰，而新配方（Muon + 门控残差）零尖峰；生产级训练全程未使用 qk-clip 之类的显式裁剪手段。

后训练模型的官方基准
对比对象为 Qwen3.7-Plus、DeepSeek-V4-Flash、Claude Opus 4.6 Max 等。
· 智能��编程：DeepSWE 1.1 得 58.7（Qwen3.7-Plus 仅 16.5，DeepSeek-V4-Flash 54.4）；SWE-bench Pro 62.5；SWE-bench 多语言版 81.0
· 长程办公/Agent：CoWorkBench 73.9、JobBench 55.7、Toolathlon 73.5，均明显领先对比模型
· 通用：GPQA Diamond 91.7、LiveCodeBench v6 91.9、IFBench 81.3
· 多模态：AndroidWorld 84.5、MathVision 带代码解释器 95.7、OSWorld 2.0 部分 52.3、长视频 LVBench 76.6

[milong @simplemilong]

@qingnianxiaozhe 我更喜欢你的

[milong @simplemilong]

一个池塘里的荷花，每一天都会以前一天的2倍数量在开放。 如果到第30天，荷花就开满了整个池塘。 请问：在第几天池塘中的荷花开了一半？

[jietang @jietang]

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.

[milong @simplemilong]

@vista8 千问输入法，因为不需要再多安装一个输入法。搜东西，家里写代码会用，但是写文章，比如给你这条评论下意识里不会用。

[Z.ai @Zai_org]

Introducing GLM-5.3: Built to Code. Ready for Cyber Defense.

- Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model
- A major leap in cybersecurity, setting a new standard among open models

Tech Blog: https://t.co/ekQkO83jCv

[Tianyi Cui @tianyi]

DeepSeek Harness was just released with MIT license. The current 0.1.0 version is a developer preview, and may still have many rough edges. Feedback is welcome!

DeepSeek Harness 已经以MIT协议开源发布。现在的0.1.0版本是一个面向Harness开发者的预览版，还很不完善。恳请大家多提提宝贵意见。

https://t.co/aBToa3b3L9 https://t.co/VMZC5i2sG0

[Tw93 @HiTw93]

世界上最好的工作，应该是持续学习者在自由市场，用自己的判断力、能力和审美持续创造价值，变成他人愿意买单的东西！