Duong Nguyen

Solo Developer

[Arc @arc]

Arc Mainnet is live. Now look up.

Tonight, 850 drones will illuminate the San Francisco skyline above the Ferry Building, telling the story of Arc, the Economic OS for the internet.

The show points to what comes next: up to $1M for eligible teams building on Arc.

Fast settlement. USDC-powered transactions. Agentic payments. Builders around the world.

Watch live on X tonight.

https://t.co/BCNxhprAsF

[Suraj Sharma @suraj_sharma14]

If I had 6 months to become an AI Infrastructure Engineer.

I'd do this.

Stage 1: Linux, Networking, and Python Foundations
Linux, bash, SSH, TCP/HTTP, DNS, Python, async programming, systemd.

Stage 2: Containers and Orchestration
Docker, Kubernetes, Helm, GPU scheduling, resource limits, multi-node clusters.

Stage 3: GPU Fundamentals
CUDA basics, nvidia-smi, VRAM math, NVLink and InfiniBand, MIG partitioning.

Stage 4: Model Serving
vLLM, SGLang, TGI, continuous batching, KV cache management, quantization (AWQ, FP8).

Stage 5: Inference Optimization
Speculative decoding, prefix caching, chunked prefill, TensorRT-LLM, Triton kernels.

Stage 6: Model Distribution and Storage
Weight sharding, safetensors, model registries, CDN for weights, lazy loading.

Stage 7: GPU Autoscaling
KEDA, queue-based scaling, cold start mitigation, spot instances, multi-region failover.

Stage 8: Training Infrastructure
Ray and Slurm, FSDP and tensor parallelism, checkpointing, fault tolerance, gradient accumulation.

Stage 9: AI Gateway Layer
Routing, load balancing, fallback chains, rate limiting, per-tenant token budgets.

Stage 10: Observability and FinOps
GPU utilization (DCGM), TTFT and ITL metrics, cost per token, tenant dashboards, alerting.

Stage 11: Security and Multi-Tenancy
Tenant isolation, secrets management, sandboxed execution, audit logs, compliance controls.

Stage 12: Portfolio and Public Benchmarks
Ship a self-hosted inference cluster, publish latency and cost benchmarks, write architecture teardowns.

Most people stay stuck watching tutorials.

Builders get hired.

(Bookmark it)

[Bojie Li @bojie_li]

《深入理解 AI Infra》开源发布了！

写完《深入理解 AI Agent》后，在与读者交流的过程中，我越来越感到，要开发好基于模型的应用，还需要理解它赖以运行的基础设施。

最近几十年系统领域最重要的变化是编程抽象的上移：从操作系统到模型上下文。传统的操作系统、编译器和硬件要为事先未知的各种程序提供通用能力，系统优化总要在可编程性与性能之间取舍。

如今 LLM 成了最重要的应用，可以针对特定的模型和加速器架构优化；模型设计也开始反过来适应硬件，DeepSeek V4/V4.1 重新设计长上下文的表示方式，就是一例。

从某种意义上说，模型成了 LLM 时代的操作系统，AI Infra 成了 LLM 时代的计算机体系结构。《计算机体系结构：量化研究方法》是我在体系结构领域的入门书，而 AI Infra 领域还缺少一本从硬件约束和模型架构出发、量化推导系统设计的书，这是我写作本书的动机。

全书约 500 页，包括 12 章：
1. 初识 AI Infra
2. 模型架构
3. 推理与训练负载
4. 加速器架构
5. 算子与运行时
6. 超节点
7. 数据中心网络
8. 推理优化
9. 分布式推理
10. 训练系统
11. 资源调度与运行环境
12. 端边云协同

欢迎 star & 提 issue/PR！

https://t.co/R1DAPLMWVq

[danialhasan @danialhasan]

https://t.co/jCDoUQPbnl

[Harshit Nayan @HarshitNay80531]

So I decided to study inference engineering and started with the fundamentals first. Studied about how the operating system abstracts the programs through processes. https://t.co/PYrhbdXRU4

[Gauri Gupta @gauri__gupta]

https://t.co/cPJILNMC2l

[lumxss @bkdgiffug]

被裁之后，把整个求职流程交给 AI，最后拿到了 AI 工程师 offer。

这个 AI Job Search 可以从岗位匹配、定制简历、求职信，到模拟筛选、修改简历一路自动跑完，最后生成可直接投递的 PDF。

整个流程本地运行，个人资料不用交给第三方。GitHub 已经 3 万+ Stars。

如果你不想重复海投、改简历，可以看看这套思路。
🔗 https://t.co/GxDuUtEUEE

[周览资源 @grgerwcwetwet]

以前做一条电影感产品视频，分镜、动效、配音、音效、剪辑全得自己折腾。

现在推荐一个开源 Skill：video-shotcraft，直接把 Claude Code / Codex 变成 AI 视频导演。

把需求交给 Agent，它能帮你设计 Storyboard、镜头运动、动效和音效，最后直接生成视频。项目里还整理了 157 张分镜卡、214 种视觉风格，甚至能导出剪映草稿继续精修。

最爽的是：以前你得学剪辑，现在你只需要学会描述自己想拍什么。

GitHub：https://t.co/lNI5FhW8GV

[Lonely @Lonely__MH]

摸索了几个月，我目前的视频工作流是：

1. Chatcut
2.剪映专业版
3.MiniMax H3

居家过日子必备

我读书多，不会骗你们的

🙂

[LovartAI @lovart_ai]

GPT-Image 2.5 absolutely cooks at pixel art.

[@振り上げた拳を花束に💐 @furikobu]

なんだよこれwwww

[Meng To @MengTo]

I asked Astra to create a Catan game in three.js and it came out incredibly well.

I played it, and the animations make it so much fun. The instructions and info cards make the game easy to learn. It even has AI players and trading.

Astra used another game I had started building as a reference. The more games I build, the more skills and assets I have for the next one. It can remix the UI, 3D models, pages, and sounds instead of rebuilding everything.

Now I want to see how many other board games I can make from this foundation.

[Superman @thesupermannx]

China published the most uncomfortable paper on vibe coding.

ETH Zurich tested 100 developers in a controlled, commercial-grade vibe coding environment to see who actually succeeds.

The findings are brutal.

The researchers tracked computer science achievement, written communication skills, and general cognitive reasoning.

They wanted to see what actually predicts vibe coding proficiency when you never touch a line of source code yourself.

Two major predictors emerged.

Written communication proficiency mattered. The ability to structure thoughts and articulate intent unambiguously in text directly impacts what the AI builds.

But that wasn't even the main takeaway.

Computer science achievement was a massive, dominant predictor of success.

Even when researchers controlled for general intelligence and reasoning skills, CS background still heavily dictated who built working software and who completely crashed.

In fact, CS knowledge contributed roughly twice the unique predictive variance of writing skills alone.

Why? Because vibe coding isn't about writing code. It’s about debugging logic.

When an AI agent builds a complex application and quietly breaks an edge case under the hood, a non-technical user looks at the glowing UI and assumes it works.

They don't know what questions to ask. They don't know what logic to challenge. They lack the mental models to recognize architectural catastrophe.

You can prompt your way past syntax.

You cannot prompt your way past a fundamental lack of engineering intuition.

The hype told us that learning to code is dead because language is all you need.

The data just proved the opposite.

To truly master the vibe, you still need to understand how the machine thinks.

[GitHub AI Projects Community 🇯🇵 @trendtech33566]

【速報】
600以上の言語で「声をそっくり再現できる」音声AIが登場👀
OmniVoice(Python製)、⭐️約1万。

何ができるかというと👇
・600以上の言語でゼロショット音声合成
・音声サンプルから高品質に声をクローン
・年齢/声の高さ/方言/ささやき声まで指定
・リアルタイムの最大40倍速で高速生成

多言語コンテンツや動画の吹き替えを作る人に刺さります。

https://t.co/HyLgU3FAFX

[Paul Solt @PaulSolt]

I build all of my APPS now without opening Xcode.

Makefiles make it easy to test in ANY coding harness or Terminal.

[Scott @scottstts]

People seem to love the bouncy jelly, so I made this Jelly Baby game with Astra in @threejs, go play with it live here🥰:

https://t.co/mc4kSIRw9W

source code: https://t.co/kyJLNivwIc

[Prasenjit Sarkar @stretchcloud]

Spotify's engineering team cut Claude Code token usage by 90% with a two-model routing approach they call Portal. The mechanism is worth understanding because it's not prompt engineering, it's hard blocks at the architecture level.

Two cheaper "assistant" models handle the tasks that don't need the expensive model: opening files, writing repetitive code, boilerplate. The expensive model only touches novel reasoning. Files over 350 lines get blocked from the expensive model entirely, not suggested away from it, blocked.

They tried rules-as-instructions first. It didn't hold. Engineers (and the model) found ways around soft rules. Hard blocks at the routing layer worked where prompt-level guidance failed.

90% token reduction is a big number. The implication is that most of what people use frontier models for in a coding workflow doesn't actually require frontier-level reasoning. Routing to capability level rather than defaulting to the best model available is a legitimate architecture pattern.

Article is worth reading in full.

https://t.co/4Kd8y24ILo

[Yatharth Gupta @_yatharthg]

Astra solved launch videos!?

Building products has never been easier but now GPT 6 Astra can make marketing content like this in 5 minutes with @fal H3 Max!!!

[Rudrank Riyam @rudrank]

App Store Connect CLI 5.0.0 is out!! 🤩

One of the biggest releases I have ever shipped, and one I am so proud of. A whole AI week went into this.

There is a LOT here. Let me show you what all new you can do now:

https://t.co/RuXh8T3xK2