小土豆

个人空间

[小土豆 @zhourui0125]

@plusxiaxia 恭喜脱坑

[小土豆 @zhourui0125]

没法跑agent，m5 pro，40k 上下文，90s首字

[Inco AI @inco_ai]

Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡

Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon.

Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.

[Inco AI @inco_ai]

Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡

Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon.

Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.

[琦玉老师 @KakaluoteW45042]

折腾 agent 这���年最大的感悟：模型聪不聪明是其次，成败基本取决于工具描述和报错信息写得有多清楚。大部分"agent 不听话"，其实是工具接口写得像谜语；再配上乱吞错误的 catch，神仙也救不回来 🙃

[Xudong Han @Xudong07452910]

AI Coding 现在很容易出现一种情况，系统已经做出来了，人却还没真正搞懂。

这篇文章里的读者几乎没有 CS 基础，却靠 LLM 搭出了 API、数据库、LLM Pipeline 和多模型 Workflow。

真正开始生产化以后，问题来了。

修掉一个 Bug，又冒出另一个。最后他意识到，自己做出来的系统，复杂度已经超过了自己的理解能力。

因为软件一直依赖抽象，我们本来就不需要理解所有底层。但以前至少还能慢慢建立一张「这东西为什么能工作」的脑内地图。

现在 AI 可以让系统以机器的速度变复杂，而这张地图依然只能以人的速度建立。

作者有句话我很认同，资料和老师已经不再是最大的瓶颈，真正难加速的是人脑吸收知识的速度。

所以 AI Coding 都会碰到的问题是：

当你已经能造出远超自己能力边界的软件时，怎样判断自己到底是在掌控它，还是只是在不断向 AI 借理解？

这个问题我感觉会困扰很多���开始用 AI 学编程的人。

文章：
https://t.co/fBa1NNvQvC

[小土豆 @zhourui0125]

devin 中 swe-2 很好，说人话，不过分思考。兜兜转转一圈，还是回到半自动模式。

[Cognition @cognition]

SWE-2 is post-trained on Kimi-K3 and proves that our RL recipe continues to scale on stronger base models.

On FrontierCode, SWE-2 achieves a score of 50.0%. It beats SWE-1.7, Grok 4.6, and GPT 5.6 Sol while matching Fable 5.1 at 64% lower cost.

[Techniques Spatiales @TechSpatiales]

Un thread passionnant de Terence Tao qui met en lumière un vrai risque pour les IA actuelles et futures: il existe peu de problème mathématiques difficiles à résoudre, et tous ont été identifiés par des humains.
En l'absence de méthode et d'historique pour évaluer où chercher, en l'absence de "gradient de la difficulté", l'IA pourrait bien assécher le pool de problèmes et se retrouver face à un paysage mathématique plat, avec simplement des murs à franchir sans trop savoir où chercher.
Pour l'instant, l'intuition humaine qui consiste à voir où on galère et où la solution pourrait mener à de nouveau problème n'est pas entrainable pour une IA, on ne peut pas tagger ce genre d'idées. Les IA qu'on fait tourner sur les vieux problèmes déjà résolus arrive à les re-résoudre mais part dans une infinité de directions ensuite.
Le nombre de problèmes mathématiques à résoudre est infini, mais le nombre de problèmes mathématiques pertinents à résoudre ne l'est pas, et des puissances de calcul et d'analyse décuplés ne sont pas une garantie de pouvoir identifier ces sujets, ces montagnes à gravir sans devoir franchir une falaise à l'aveugle.
D'ailleurs c'est déjà un sujet d'actualité, la source se tarit et une fois tous les "problèmes du millénaire" résolus soit des mathématiciens humains sauront comprendre les implications et trouver des pistes vers de nouveaux problèmes, soit l'IA risque de dire "bon, et maintenant je vais où, je ne vois plus rien émerger là ?".
Pour les ingénieurs qui lisent ça, ça serait une situation un peu similaire à demander à une IA de résoudre un problème sans rien lui spécifier. L'IA est aujourd'hui incapable d'inventer des spécifications pour un problème qu'elle ne connait pas. L'analogie a ses limites, mais c'est un peu là où se dirige les IA pour les maths. L'évolution et la façon de contourner ces problèmes sera ardue, nécessitera de nouveaux outils, mais sera passionnant à suivre.

[小土豆 @zhourui0125]

AI时代码农：理解整体结构与业务，掌握业务和系统关键约束，按风险检查实现，能够验证结果，并保留接管与排障能力。

换一句话说：建模、测试、可观测性和协作越扎实，就越能把更多实现工作交给 AI。

[小土豆 @zhourui0125]

codex 重置了，继续蹬

[huhuhang @huhuhang]

相反，我越发感觉到 Vibe Coding 就是技术工作者的短视频，对脑子和注意力的害处超过益处。

[小土豆 @zhourui0125]

@nicekate8888 还要等一个小时

[小土豆 @zhourui0125]

@imvihv 不耐蹬

[小土豆 @zhourui0125]

使用 Credits/积分 计费的编程/agent套餐，直接写额度不好么，何必为难大家呢？当大家算不清楚么

[Erick @ErickSky]

Tenían razón todos los leakers!!

Ox Alpha era el GLM 5.3 Flash

Y ahora yo pregunto...

¿Los empleados de Google son pendejos?

[小土豆 @zhourui0125]

@0xLogicrw 有传是 google 的，Z➗不是这种营销风格

[小土豆 @zhourui0125]

@9hills A➗已经不退款了

[小土豆 @zhourui0125]

@__oQuery 特别少，一天就刷完了

[小土豆 @zhourui0125]

glm快点把知识密度高+大参数量模型端上来���

[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.

[小土豆 @zhourui0125]

shit，换工作掉进坑里了