Wuggyan

喜劇 藝術 電影

[CyrilXBT @cyrilXBT]

I'm genuinely shocked this is FREE

YOU CAN BUILD YOUR FIRST AI AGENT IN ONE AFTERNOON. FOR FREE.

My first one took me two weeks.

This guide gets you there in a few hours:

Step 1: Pick one boring task you repeat every week.

Step 2: Give the agent tools: read files, search, send a draft.

Step 3: Add a Jev decision layer, so the small calls (which tool, which file, done or not) cost a fraction of a cent.

Step 4: Put a human checkpoint before anything that sends, deletes or pays.

Step 5: Run it, watch it, tighten it.

One builder with this setup ships like a small team.

Full guide below. Bookmark it.

follow @cyrilXBT

[Ronin @DeRonin_]

which board should you actually start with?

Arduino vs ESP32 vs Raspberry Pi

4 months into robotics as a hobby, here's what i'd tell myself on day one:

1. start with arduino

use it when you're learning:

- motors
- servos
- sensors
- basic circuits

atmega328p, 16 mhz, 2 kb of ram, and 5v logic so most hobby parts plug straight in

when your code runs, nothing interrupts it, and that makes timing easy to reason about

best place to learn the fundamentals without debugging an operating system at the same time

you outgrow it the day you add wifi or a camera

2. move to esp32

use it when you want:

- wifi
- bluetooth
- wireless control
- anything that reports back to your phone

dual core at 240 mhz, 520 kb ram, radios on the chip, around $5 a board

more connectivity and still hands on, so you keep writing the same kind of code

two things that cost me a weekend each:

3.3v logic, so your 5v sensors need a level shifter or they read garbage

and on the original esp32 the adc2 pins stop reading the moment wifi turns on

3. go for raspberry pi

use it when your robot needs:

- computer vision
- ai and ml
- ros 2
- heavy processing

now you're running a computer inside your robot, with all of the upside and all of the cost

linux decides when your code runs, so pwm timing jitters and servos twitch

and pulling the power without a shutdown corrupts the sd card, which i learned twice

also worth knowing, raspberry makes a real microcontroller too, the pico, about $4

so the quick version:

never wired anything before -> arduino

want wifi and the best price -> esp32

need vision or a camera -> raspberry pi

and yes, you can use more than one

arduino -> control
esp32 -> connect
raspberry pi -> compute

that combination is what took me 4 months to understand

the pi is the brain and the microcontroller is the spine, and a real robot wants both

also, if you want to learn robotics engineering for 6 months, read the article below:

[Azlen @azlenelza]

Continuation of my old video-as-3d-object demo

Always wanted to extend it further with segmentation so you can search for any type of object within the scene

[Mia @MiaAI_lab]

Full GLM 5.3 running on 3x DGX Sparks with TensorFold

Currently 24-28 tok/s on prose

[Hytidel聊商业（文章更多干货） @HytidelLegend]

用 AI 做了一个背单词神器🛠️，已经开源了 https://t.co/XUyLLlHYcT

[Codez @0xCodez]

Andrej Karpathy predicted the future of AI once again:

“Everyone’s renting frontier models for jobs a 3B model could do. Small models are the future.”

this 18-page PDF breaks down Karpathy’s case for working with small LLMs.

the real question isn’t “Is the small model as good?” It’s “Which of my 1,000 calls ever needed a frontier model?”

And @thewebai just answered it for formal logic.

TwIL-LM3-Pro:
→ 3.6B params, on par with Qwen3-8B on formal logic
→ leads VibeThinker-3B on all 6 formal-logic tasks tested
→ 95.4% on BBH logic, 95% on SVAMP
→ 2.09 GiB in Q4, runs on CPU or 4GB VRAM
→ no API bill, no data leaving your machine

The secret isn't size. It's post-training.

PDF below. Model 👇
https://t.co/T9lgygHgNS

[Adam @_overment]

bro aint wrong

[Alexandr Wang @alexandr_wang]

🚨 MUSE ECOSYSTEM ALERT 🚨

today we are announcing Muse Gadgets!

this is an open-source ESP32 firmware and Linux SDK for anyone to make hardware that works with Muse

we are also releasing our own gadget—Muse Home Link—to enable your muse to work with your smart devices (TV, speakers, etc.)

[Grok Bot @bot]

Grok Bot can now suggest ways to help without you needing to ask.

[Manus @ManusAI]

Introducing Manus 2.0

[jiangkoumo @jiangkoumo_]

如果说 Pi 1.0 的全屏 TUI 只是换了身皮，那随版本一起上架的Pi Durable，才是这次版本更新的的重磅炸弹

这个由 mitsuhiko（Flask / Sentry 作者）和 Mario Zechner 联手操刀的包，彻底撕开了当前 Agent 行业最心照不宣的遮羞布：

以前大家写的 Agent，本质上只是“内存里的临时玩具”

跑了 20 步复杂工具��，突然断网、OOM 或进程被杀，整条状态直接灰飞烟灭。如果重新跑，不仅白烧昂贵的 token，非幂等的写库、发请求、跑脚本还会把生产环境彻底搞烂。

Pi Durable 没搞花哨的抽象，而是把数据库内核的 WAL（预写日志）和事务提交，硬生生塞进了 Agent 运行时：

第一，意图先于副作用（Intent-Before-Effect）。任何对话、模型响应、工具调用与上下文变更，在展示或执行前，必须先原子提交落盘。副作用还没发生，意图已经持久化了。

第二，故障原地断点续跑。进程崩溃重启后，只有显式标了 replay: "safe" 的只读工具会重跑；其他写操作直接回填 interrupted 错误并保留已提交上下文。调一句 harness.resume()，直接从上次跌倒的地方原地爬起来。

第三，严格的子任务因果所有权树（Task Tree）。支持 allSettled 等待和 failFast 熔断，一个子任务崩溃立刻自底向上（Bottom-up）递归清理，绝不残留悬空孤儿进程。甚至每 100ms 自动刷盘一次，崩溃最多只损失 100ms 的窗口。

为什么说它重要？因为这是 Pi Agent 第一次从“终端单兵小玩具”，蜕变成“抗崩溃、长生命周期的云端工业级底座”

行业嗅觉最灵敏的 Vercel，已经在第一时间上线了 @ai-sdk/harness-pi，直接拿它作为官方运行时。

用写数据库和分布式系统的底层工程思维来重构 Agent，这才是开源 Infra 真正的降维打击。

[a16z @a16z]

Where does money invested into the AI buildout actually go?

For every $100 flowing into the supply chain:
- $50 to chips
- $20 to power
- $15 to networking
- $15 to cooling, buildings, and land

More charts in State of Markets II: https://t.co/MTaxKUxa2w

[洛克船长 @slashui]

太酷啦！以后的系统架构图都以这个标准来做！！

[༺ཧคlคฝคཊ༻ @rybinfx]

Web crawlers /[•]\ #js x #css

[Pi @pidotdev]

People of Pi: We've shipped Pi 1.0 with Pi Durable. Go make them yours.

https://t.co/oUPfVfvAZl

[Andrej Karpathy @karpathy]

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:

Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:

Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:

Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!

In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

[uudon @uudonX]

🎁 Karpathy 又给行业发福利了！codex / claude code 给你生产一大堆文字，你还一个字不落地读吗？ 别这样，效率太低了，试试下面 Karpathy 提出的几个方法。

你只能感慨，牛人终究是牛人，聪明的同时，还努力，还不断想办法让自己高效。

Karpathy 提了一个很多人忽略的问题：当 AI 的产出速度超过人的理解速度，继续让它写更多文字，只会把瓶颈越堆越高。

模型可以在几分钟内生成几万行代码、研究报告和复杂方案，但人仍然要判断它做了什么、逻辑是否成立、结果能不能使用。

所以 Karpathy 提出的几个方案：怎样把模型的输出转换成人更容易理解和验证的形式。

1️⃣ 用 ASD-STE100 约束文字。
普通模型喜欢使用长句、抽象词和模糊表达。同一个概念写了很多字，读完仍然不知道重点。
ASD-STE100 最初用于航空维修文档，它限制词汇、句式和表达方式，要求每句话清楚、直接、含义唯一。
可以提示模型：“用 80% 的 ASD-STE100 风格解释这个问题。”

它的收益是降低语言噪声，减少歧义，让技术文档和复杂概念更容易检查。

2️⃣ 生成图表。
有些内容的难点来自关系复杂，例如系统架构、调用链、因果关系和模块依赖。线性文字很难同时展示这些结构。
让模型生成流程图、架构图或概念图，可以直接看到节点之间如何连接、数据如何流动、问题可能出现在哪里。

它的收益是降低理解成本，也更容易发现缺失关系和逻辑断点。

3️⃣ 生成交互式网页。
如果一个问题包含状态、参数或动态变化，静态文字和图片都不够。
这时可以让模型输出 HTML，通过按钮、滑块和动画展示：参数变化后结果会发生什么，系统在不同状态下如何运行。
它的收益是把“阅读结论”变成“亲手验证”。用户可以修改参数、观察结果，快速建立因果理解。

4️⃣ 生成解释视频。
算法执行、物理过程和抽象概念都包含时间顺序。视频可以逐步展示变化过程，同时结合动画、旁白和重点标注。
Karpathy 建议直接让模型制作 3Blue1Brown 风格的解释视频。

它的收益是帮助人建立直觉，尤其适合只看文字很难想象的概念。

这几个方案背后的真正变化，是“可丢弃软件”的成本正在接近零。过去，为一个问题单独制作网页、动画或视频不划算。

现在模型可以根据你的知识水平临时生成，用完就扔，别再傻傻完整读完模型给你输出的所有内容了，效率太低了，试试上面的方法。

[Bear Liu @bearliu]

新西兰人做 AI 有一点我特别佩服：特别实在。

上周去 NZ Tech Expo，最好玩的一家初创团队是用 AI 给牛称重的，产品名字就叫 Scan a Bull。扫一下就能估出重量，误差大概两三公斤，农场主觉得完全够用。后来翻参展商名录，发现他们 CEO 就姓 Bull。

有人说：“为什么不用磅秤呢？称得更准。”

真实情况是，你养牛就知道了。你把牛赶到磅秤那里，一只牛还可以牵着走，但如果是 20 头牛、50 头牛呢？

随时用一个拍照 App 来监控体重，省时间，还能随时知道自己的资产。

能解决实际问题的 AI 才是好 AI。

推上天天在吵哪个模型最强，这边有人在认真研究一头牛到底有多重。

[Pranav @PranavOriginals]

My favourite footer section I’ve made so far.