Leviathan

Son, Husband, Father, Researcher, Geek and Tech enthusiast.

[Nix0n @UnTalNixon_exe]

🏙️ De mapas a vectores: Cómo la IA está empezando a entender las ciudades

El software de mapas tradicional (GIS) es excelente para visualizar, pero limitado cuando intentas aplicar modelos de aprendizaje profundo para predecir flujos urbanos, optimizar el transporte o analizar la morfología de una ciudad.

La librería de código abierto city2graph construye un puente directo entre los datos geoespaciales y la Inteligencia Artificial Geográfica (GeoAI). Transforma geometrías complejas en grafos relacionales listos para entrenar Redes Neuronales de Grafos (GNNs).

.

🧱 La infraestructura técnica detrás del proyecto:

• Modelado Morfológico: Permite conectar variables del entorno urbano (edificios, calles, zonas de uso de suelo) y convertirlos en nodos con atributos enriquecidos de proximidad y adyacencia.

• Ingesta de Transporte Eficiente: Integra consultas SQL optimizadas mediante DuckDB para procesar feeds de transporte público (GTFS), facilitando el análisis de accesibilidad y el modelado de redes multimodales en segundos.

• Análisis Dinámico de Movilidad: Traduce matrices de origen-destino (como flujos de bicicletas compartidas, peatones o migraciones internas) directamente en redes de flujo humano legibles por modelos predictivos.

• Ecosistema Unificado: Funciona de forma nativa como una interfaz integrada entre GeoPandas (análisis espacial), NetworkX (análisis de redes tradicionales) y PyTorch Geometric (Deep Learning).

.

Al final, las ciudades no son solo imágenes o coordenadas aisladas; son redes vivas de interconexiones complejas. Llevar esa complejidad al formato exacto que entienden los tensores de IA es el primer paso para diseñar entornos urbanos realmente inteligentes.

Dejo el enlace al repositorio oficial en la sección de comentarios.

Guarda este post en tus marcadores si trabajas con datos espaciales, grafos o machine learning aplicado al mundo real 🔖

[Felix Rieseberg @felixrieseberg]

Today, we're introducing Claude Fable 5 and Mythos 5, two configurations of our next major language model.

I'd normally highlight the numbers: It's SOTA on nearly all benchmarks. I want to talk about something else, because with Fable 5 out in the world, I think a third era quietly started today.

I lead Claude Code & Cowork on the desktop, so I think a lot about how people use AI to get work done. I believe we're about to see a major shift, moving from giving AI tasks to giving it responsibilities.

[Udara @TGUPJ]

Research 1.3 is out now with our new annotator!

[AYi @AYi_AInotes]

分享一个计算机视觉开发者必藏神器，4 万星GitHub开源！
不用自己写几百行画框跟踪代码，
一行命令搞定所有可视化，
Supervision，CV 界真正的瑞士军刀。

1️⃣ 它到底有多强？
自动画框加标签，支持编号、自定义样式，
视频对象永久跟踪，ID 不乱跳，轨迹自动生成，
一键转换 YOLO/COCO/Pascal VOC 所有数据集格式，
内置热力图、区域计数、越线检测、人体骨架、人脸网格
模型无关，YOLO/Transformers/ 任何检测模型都能接，

2️⃣安装只要一句话
运行：pip install supervision

从 NBA 球员实时跟踪、交通路口车流量统计，
到工业缺陷检测、无人机目标追踪，所有你能想到的 CV 可视化场景，它全搞定。

3️⃣已经被 6500 + 个开源 CV 项目在用

现在就装，早用早爽，
做 CV 原型和演示的终极工具，没有之一！

[Geek Lite @QingQ77]

为中文科研工作者提供一套集成化 AI 学术技能包，覆盖论文写作、学术 Office 文档生成与科研计算三大场景。

https://t.co/rtT5cphqw1

三个技能包，分别管论文写作、学术 Office 文档和科研计算，都能在 Claude Code 和 Codex 上直接用。

research-writing-skill 写论文改稿子回审稿人，默认用中文。office-academic-skill 做文献报告、组会 PPT、答辩 PPT，输出可编辑的文件。scientific-toolkit-skill 面向光电方向，包装了 MATLAB/Python 计算、画论文图、查文献这些活。

三个可以串起来用：先算数据出图，再写论文，最后做答辩 PPT。

[Mushtaq Bilal, PhD @MushtaqBilalPhD]

How to Connect Zotero with Claude Code

A step-by-step tutorial

This tutorial is written a simple and accessible language. It contains a few technical details, but you don't need any technical knowledge of coding expertise to follow this.

[Leviathan @wohkoon]

@barabasi @netscisociety 恭喜🎉🎉🎉

[superwhisper @superwhisper]

Voice-control Codex from anywhere. Get notified when it's done, respond with your voice ✨

[AnySearch @AnySearchAI]

AI answers look right. But they’re often incomplete or just wrong.
Short answers. Few sources. Not accurate.
AnySearch pulls real-time data from multiple trusted sources into one view, so you get answers you can actually trust.

[DeFi狙击手 | Ai🕊️ @bi_9527zx]

Hermes Agent这样设置那么恭喜你为了省钱又进一步😂

deepseek为主大模型，kimi2.6为视觉模型，并且无需来回切换，只要丢图片进去就会自动使用kimi2.6去识别。（理论是你可以设置更多的大模型去完成不同的工作，并且全程自动无需切换大模型）

写很多以X文章的形式去编辑这类干货教程已经很多了，我想试试普通的发帖方式编辑会不会被更多的人去看到。
那么话不多说下面直接是干货！

-

主模型：deepseek-v4-pro

视觉模型：kimi-k2.6

运作逻辑：deepseek 负责所有的工具调用、代码、排查、对话，一旦你丢过来图片或者视频帧，系统自动切到 kimi 去识图，识完把结果扔回给 deepseek 继续处理。整个过程不需要手动切换模型，你体感上就是同一个助手在干活，但实际上后面是两个大脑在协作。

-

为什么这样搭

deepseek-v4-pro 的工具调用和长上下文处理能力确实强，跑代码任务、排查问题、多步骤复杂操作的稳定性比 kimi 好一截。

但它有一个硬伤暂时不支持多模态，直接给它丢图片会报 unknown variant image_url 的错。

而 kimi-k2.6 的多模态能力不用说，中文 OCR 也准，但纯文本推理和工具调用不如 deepseek 稳。

所以最优解就是各取所长，deepseek 当大脑，kimi 当眼睛。

（其实说白了就是为了省Token！！！）

-

配置步骤

第一步，确保 API Key 就位。在 `~/.hermes/.env` 里确认两行：

```
DEEPSEEK_API_KEY=你的deepseek密钥
KIMI_API_KEY=你的kimi密钥
```

第二步，切主模型，终端敲：

```
hermes config set model.default deepseek-v4-pro
hermes config set model.provider deepseek
```

第三步，设视觉模型：

```
hermes config set https://t.co/Vyak5jfApm.model kimi-k2.6
hermes config set https://t.co/Vyak5jfApm.provider kimi-coding
hermes config set https://t.co/Vyak5jfApm.base_url https://t.co/EmbACMPrQl
```

第四步，如果你的 kimi key 是老 key（sk- 开头不是 sk-kimi- 开头），需要在 `.env` 最后加一行：

```
KIMI_BASE_URL=https://t.co/EmbACMPrQl
```

新 key 走 https://t.co/JdI5IxyyOl，老 key 走 https://t.co/FgY4mpqDTg，config 默认写的 https://t.co/yjidM8709K 是错的，不改会 401。

第五步，`/reset` 重开会话，生效。

-

踩过的坑

一个是上面说的 api 地址问题，https://t.co/asdrLg5wJO 和 https://t.co/VQRXaILbIg 差一个字母，排查了好一阵。

另一个更隐蔽：Hermes 内置的 vision 工具往 kimi API 发请求时，content 数组里 image_url 排在了 text 前面，kimi 的接口不认这个顺序，偶尔 400。
目前的变通方案是发图时绕过内置工具，直接 Python 调 moonshot API，text 放 image_url 前面就没事。
不影响使用，但不算完美，等后续优化。

-

配完之后你正常跟他聊天就行，deepseek 处理一切。

发图过来的时候，你不需要说"用 kimi 看"，系统自动检测到图片就走 kimi 视觉，识读完无缝回到 deepseek 继续对话。

你体感上就是同一个助手能看能想能干，但后台其实是两个模型在做各自最擅长的事。

这个方案的最大好处是不用你在聊天里手动切模型，流程全自动。而且以后如果想换视觉模型，改 `https://t.co/Vyak5jfApm.model` 一行就行，不影响主模型。

-

兄弟们有哪个不懂的直接甩在评论区，狙击手会挨个解答！

[Mr Panda @PandaTalk8]

用Codex 时， 一定要在浏览器安装 codex 的扩展。

然后在codex 扩展中打开允许访问文件url权限。

这样你就可以让codex 完全控制你的浏览器来帮你工作了。

[Bernhard Mueller @muellerberndt]

How to hack gravity and build your own flying saucer (with repeatable bench experiment)

https://t.co/TBc6BbkS1m

[Leviathan @wohkoon]

太好了！

[东村火影 @charliefish001]

台湾20年前的音标教学视频，遥遥领先。
视频里的美女老师周育如已经是台湾国立清华大学幼教方面的教授了。

[Leviathan @wohkoon]

The emergence of AI has greatly increased the workload of university teachers.

[Leviathan @wohkoon]

https://t.co/NrMgd9mcfM

[B.AI @BAI_AGI]

⚠️ https://t.co/JerjymcZyf API Key 已升级！旧版 API Key 将于2026.06.22 23:59 (SGT) 停止支持。

如您仍在使用旧版 Key，请尽快创建并切换至新版 API Key，避免服务中断👇

[meteoblue @meteoblue]

🌐 Forecast resolution has become one of the defining developments in modern meteorology.

Over the past decades, operational forecasting has progressed from grid spacings of tens of kilometres to kilometre-scale simulations capable of resolving local terrain effects, urban heat islands and small-scale weather phenomena.

Higher resolution is changing how forecasts are verified, how computational resources are allocated, how convection is represented within models, and increasingly how weather information is used for decision-making. Our latest article also explores why better forecasts depend not only on more powerful models, but equally on the quality of global observation networks, and how artificial intelligence is beginning to complement traditional numerical weather prediction.

Among the topics covered are:
➡️ the value and limitations of high-resolution forecasting
➡️ the computational challenges behind kilometre-scale models
➡️ the "double penalty effect" and forecast verification
➡️ the role of observations in reducing forecast uncertainty
➡️ the emerging role of AI in next-generation forecasting systems

Read the full article 🔗 https://t.co/pbt7CmU5TS

[Jason Wang @zswang24]

Codex has become a core part of how we do data science @OpenAI.

Today we’re launching a new Data Science plugin, built by data scientists for data scientists.

It helps with everything from exploring raw data and diagnosing metric changes to building dashboards, writing reports, and getting to stakeholder-ready insights faster.

The feedback has been strong:
• 100% of users said it speeds up the path from raw data to insight.
• 100% said it helps them take on more work than they otherwise could.

The best description I’ve heard is that it’s a force multiplier: getting us from 0→80% quickly, helping us make progress across multiple projects at once, and improving the quality of the final output.

Excited to share it with everyone.

https://t.co/Ou5jhCEjMi

[Google AI Developers @googleaidevs]

Building autonomous agents for scientific discovery? 🧬🤖

@GoogleDeepMind Science Skills is now available on GitHub. We've open-sourced this specialized toolkit to accelerate your agentic workflows with scientific grounding and higher token efficiency.

Download now ↓
https://t.co/cwp1HOeKvo

[Leviathan @wohkoon]

Very interesting.

[Google Maps Platform @GMapsPlatform]

Frustrated by traffic? 🚗💨 What if roads could tell us what's next?

Mobility infrastructure management leader @Abertis uses our Roads Management Insights to turn mobility data into powerful foresight, making your journeys smoother and safer.

RMI's unique combination of data quality, granularity, and the ability to operationalize insights through a scalable analytics environment allows Abertis to move beyond descriptive reporting toward proactive management.

See how we're making roads smarter, together → https://t.co/vjE1vbWV3i