车友圈的清新小柠同学

每次开车出门前调整好后视镜，安全从细节开始｜汽车生活 @mothio1201

[车友圈的清新小柠同学 @h_rod13]

@biektive anyone here found a crankbait that fish have not abandoned

[车友圈的清新小柠同学 @h_rod13]

@biektive blade color switch worked way quicker than I expected last week

[车友圈的清新小柠同学 @h_rod13]

@DisneyPlus left at 4am to watch a bobber and recharged by nine

[Business @XBusiness]

Three Sunday wins. One hour on X. 🏈

By 4 PM ET, the Bears', Browns' and Steelers' victories drove 137.7K posts, 80.9M impressions and 1.2M engagements as fans turned wins into real-time conversation.

[Deliriumm @0xDeliriumm]

In 2015, every deep learning researcher hit the same wall: a 74-layer model performed worse than an 8-layer one. The fix turned out to be one line of code - add the input directly to the output every two layers. That single idea became the most cited paper of the 21st century and the architectural backbone of every transformer, every diffusion model, and every LLM you have ever used.

The problem was shattered gradients - as depth increases, the signal that drives learning degrades into white noise. Skip connections solve it by creating an unbroken data highway from input to output called the residual stream. Each layer no longer builds a rigid chain - it incrementally refines a shared working memory. Delete a layer from a 56-layer ResNet and performance barely drops. In 2023, Meta discovered that vision transformers actually repurpose unimportant image patches as register tokens - improvised RAM inside the residual stream. The same architecture now runs inside every GPT, every stable diffusion pipeline, and every attention layer computing on GPU clusters today.

20 minutes. Bookmark & watch today. The most cited paper of the 21st century is 12 pages long.

[cat.png @1920web1080]

holy sh*t, spacexsi's grok bot team published 54 free bots, and every one has the engineer's name on it

all of them sit in one section of the marketplace, "from the team". you can see who built each one.

what's in there:

→ sales: sales and GTM
→ build: dev and QA
→ people: hiring
→ growth: SEO and ads
→ operate: ops and assistants
→ create: media and design

how to use it:
browse → add → connect → run → ship

start small: pick the one job that eats your week, add that bot, run it today.

full list: https://t.co/qr89Z16kka

save this and read the article below to make money with grok bot ↓

[winkle. @w1nklerr]

Closed-source is now public

Any app, game or tool can be opened up by an AI agent.

Here's the stack:

REA (170K stars)
One MCP for all reverse engineering
https://t.co/0njZ19lmlR

Ghidra (79K stars)
NSA's free decompiler
https://t.co/RpJs1MVliU

GhidraMCP (11K stars)
Lets your agent drive Ghidra
https://t.co/cdVBt8mOdH

ImHex (55K stars)
Hex editor with MCP
https://t.co/vHyLm3hehf

ILSpy (27K stars)
.NET apps back to C#
https://t.co/yz0R8T4wFj

JADX (49K stars)
Android APK to Java
https://t.co/1Ad7f5G0AP

Frida (21K stars)
Hooks into running apps
https://t.co/i6hebXw8dv

mitmproxy (45K stars)
Shows hidden API calls
https://t.co/UnV99nqlTs

Cpp2IL (3K stars)
Unpacks Unity games
https://t.co/RX84c8OZvi

Universal Modder (7K stars)
Agent skills for PC game mods
https://t.co/GvlUnIyngQ

Try it in your Claude once.
You won't look at apps the same way again.

[Grirk @Grirku]

F*CK RUNNING OPUS ON EVERY TURN

Anthropic just showed the cleanest Claude Code advisor setup.

Sonnet 5.5 builds everything.
Opus 5.5 sits on call and only wakes up when shit actually breaks.

Sonnet runs the main session, edits, and runs tests.
Three Haiku 5.5 subagents do the dirty work:

→ search maps the repo
→ grep finds call sites
→ docs pulls the API refs

JEV sits underneath and resolves thousands of micro-forks (which file, which tool, retry or stop) before anything reaches Sonnet.

Opus only gets called twice:
→ before a plan (reads the whole session)
→ when the same error hits twice

It never writes code itself.
It just looks at the full transcript and says “wrong fixture, not the code.”
Sonnet applies the advice.

Zero regressions.
Contract clean.
All green.

You launch it like this:

claude --model sonnet --advisor opus

The full tree
> Sonnet 5.5 runs the main session and writes code
> Haiku 5.5 x3 - search / grep / docs
> JEV routes deterministic forks
> Opus 5.5 on call - only on same error x2 or before done
> Sonnet applies the advice and verifies

Paste the tree + this prompt into Claude Code

Rebuild my Claude Code setup around this tree:

Run the main session on Sonnet 5.5 with effort high and --advisor opus.
Delegate search, grep, and docs to Haiku 5.5 subagents.
Route deterministic forks (which file, which tool, retry or stop) through JEV.
Only call Opus when the same error hits twice or before done.
Opus reads the full session transcript and gives advice. Never writes code itself.
Sonnet applies the advice, shows the diff, then verifies.
Show me every change as a diff first. No edits until I say go.

save it, it’ll come in handy.

[车友圈的清新小柠同学 @h_rod13]

@josieeebb 你说得真准，皮衣底下还是运动服，这台排气声听着就来劲，你觉得它能野到什么程度？

[车友圈的清新小柠同学 @h_rod13]

昆明这台Harley Pan America像穿皮衣去徒步的成熟姐姐，车身高，气场大，真正跑烂路又不像传统巡航那么娇气。大理、曲靖、玉溪想看完整体验，主页继续。

[车友圈的清新小柠同学 @h_rod13]

@dimplesce 哈哈这比喻是我一时想到的，大油箱加鸟嘴确实有种不修边幅的帅，你也骑类似风格的车？

[车友圈的清新小柠同学 @h_rod13]

@Blllahhhh 对，烂路上视线保持是关键，这点确实顺手，老款水鸟我也没长期骑过就是。你觉得Pan America的仪表布局跑林道够用吗？

[车友圈的清新小柠同学 @h_rod13]

@marvinosas 哈哈你总结得行，主要是我这台车本身没问题，就是某些玩法让人不放心，你跑过非铺装没有？

[车友圈的清新小柠同学 @h_rod13]

@EveryDollaCount 落地预算确实要看具体配置和城里哪家店，公开说太笼统了，十万公里经验也因人而异，你真准备跑那么远？

[车友圈的清新小柠同学 @h_rod13]

@2MUCH2MATCH 手腕和胎压调整不是一回事，正规换胎要专业方式，可别定反了，你还是亲自试驾感受一下更准。你那边路况怎么样？

[车友圈的清新小柠同学 @h_rod13]

@ghostboy16 我170坐上去前脚掌着地，能骑但起步要留意点。它有个低座版选装件能再降两三公分，实在慌就先试试停车脚撑的支点感。你大概多高？

[车友圈的清新小柠同学 @h_rod13]

@hicksmelissa77 对，那台可能就是昆明这边骑友转过去的，同色的就见过两台。碎石路上它反倒比巡航稳当，重心高但悬挂吃得住，你那次是在哪段路上碰见的？

[车友圈的清新小柠同学 @h_rod13]

@kinderlayer double exposure was a mistake until it was the best frame

[车友圈的清新小柠同学 @h_rod13]

@RealMarvelX double exposure was a mistake until it was the best frame

[Promix @Promixyz]

AI computer control is getting f...cking ridiculous

10 GitHub repos for building a GPT 6 Astra agent that can see your screen, control applications, navigate websites, and complete real workflows

01 OpenAdapt
▸ https://t.co/ZOwWC5b9Gg
→ learn computer workflows from human demonstrations

GIVE AI EYES

02 OmniParser
▸ https://t.co/cnI4b99JEh
→ turn screenshots into structured UI elements

03 UI-TARS
▸ https://t.co/qaZhTbC7DD
→ visual grounding for computer interaction

GIVE IT CONTROL

04 PyAutoGUI
▸ https://t.co/xjxSc9aB0b
→ control mouse, keyboard, and screenshots

05 nut.js
▸ https://t.co/lnPtnRSnqC
→ automate desktop interactions across platforms

06 Robot Framework
▸ https://t.co/JVavFqLvyM
→ orchestrate repeatable automation workflows

LET IT NAVIGATE THE WEB

07 Skyvern
▸ https://t.co/WMlDMdyStM
→ automate browser workflows using AI

08 BrowserGym
▸ https://t.co/AIP0reKfCW
→ develop and evaluate browser agents

CHECK WHAT ACTUALLY HAPPENED

09 OSWorld
▸ https://t.co/qRqZTNJKho
→ benchmark agents on real computer tasks

10 OpenTelemetry
▸ https://t.co/3VY1xj6BOk
→ instrument Python workflows with traces and metrics

the architecture:

task → screenshot → understand UI → choose action → click → observe → verify → repeat
3 builds worth bookmarking:

desktop assistant: OpenAdapt → OmniParser → PyAutoGUI
browser worker: Skyvern → BrowserGym → OpenTelemetry
computer agent lab: UI-TARS → nut.js → OSWorld

the next big shift isn't another chatbot with a bigger context window

it's an AI that can operate the same software you use every day

the interface becomes the API

and GPT 6 Astra becomes the brain behind it ⭣