草丛蹲点王

鎱㈡參鍙樼敎

[草丛蹲点王 @viralismbec]

@StateFarm this campaign roasts fresh common attribution mistakes yet got a quiet okay, almost cheerful vibe

[Varian @0xVarian]

WHOEVER LEAKED THIS JEV ROUTER CODE SHOULD BE REPLACING KAFKA

I thought handing a job between workers meant Kafka, Redis, a schema registry, and a retry topic that no one is paged on.

Apparently not.

This router's setup is much more interesting:

→ CHOICE returns the next worker name from the result, pinned to a labeled set with confidence attached, no "pick the fastest branch" guess
→ GATE compares confidence against 0.85, if it falls short the destination becomes "review" automatically, no silent slip into write
→ FOLDER writes to queue/[destination]/ with mkdir parents true, no central broker, no database, no connection pool to leak
→ JOB name is a uuid4 hex with a .json extension (payload serialized once with json dumps), one file per handoff, no key collision on the queue
→ DISPATCH reads 8 workers in parallel across 14 bus readers (worker loads 68, 77, 6, 25, 95, 29, 58, 82%), flags 7 workers whose queue drifted past baseline this frame

so you get:

Choices named. Gates defaulted. Jobs filed. Workers parallel.

the interesting part isn't picking the right broker.
it's giving every handoff a json file the next worker can read on its own.

Instead of a weekend configuring Kafka with retention and partitions, you write queue/[destination]/[uuid].json, let 8 workers tail their folder, and route 10,000 decisions at $0.02 total with 251ms median.

Pick the worker. Gate the choice. Write the file.

[Zayn @zayn0x28]

Opus 5.5 remade Gojo Satoru vs Jogo, no After Effects

this Domain Expansion goes f*cking hard...

Drop the PDF below into Opus 5.5 and make your own.

Then save the article, it's the full Motion Design guide for Opus 5.5.

[草丛蹲点王 @viralismbec]

@vampScally Everyone's chasing virality while ignoring retention metrics.

[Cheery @cheery9998]

从零搭建「第二大脑」。

Obsidian 的创始人 kepano。把让 AI 理解 Obsidian 的 Skills 开源了。

GitHub 上有4 个现成项目，几乎把整套流程都准备好了。

① obsidian-skills — kepano

49.8K Stars。

Obsidian 官方生态的 AI Skills，让 Claude 理解 Markdown、双向链接、Canvas 等 Obsidian 原生格式。

② claude-obsidian — AgriciDaniel

15K Stars。

直接提供 Wiki、知识摄取、查询、Lint 和保存命令。

不需要从零写工作流。

③ obsidian-second-brain — eugeniughelbur

4.8K Stars，48 个命令。

支持 Claude Code、Codex、Gemini 等多个 AI Agent。

④ second-brain-starter — coleam00

它不直接提供知识库，而是先采访你，了解你的工作、目标和项目，再生成一套适合你的 9 阶段建设计划。

但真正有意思的，不是这些仓库。

而是 Andrej Karpathy 提出的整个思路：

不要让 AI 每次对话都重新搜索你的知识，而要让它持续编译和维护你的知识。

[CELSIUS Energy Drink @CelsiusOfficial]

the shirley temple to energy drink pipeline is real

[草丛蹲点王 @viralismbec]

@CelsiusOfficial Well deserved

[meya. @birmeya]

i added 6 more tools to pumprint.

12 tools now. one workspace.

https://t.co/gpLG9PUOZ2

here’s what’s new:

→ WARDEN: inspect the github repo behind a project. commits, contributors, forks and the history behind the code.

→ BODKIN: follow creator fees, deployer history and the trade tape.

→ STAMPEDE: track wallet rotations. who sold coin A, bought coin B, and how many wallets made the same move.

→ TRAIL: follow a token’s sellers to their next observed purchase.

→ PELLET: check wallets returning after inactivity against their earlier trading history.

→ FOMOBILL: look for possible connections between token holders through their on-chain activity.

Split works across the expanded lineup.

open any two tools side by side on site. resize the panes. keep one open while switching the other.

STAMPEDE + TRAIL to investigate a rotation.
chameleonscope + WARDEN to inspect a launch and its repository.

the next batch of tools will come later.

those will be reserved for token holders who actively trade. i’ll share the access requirements before they launch.

CA: 766RYwWxvvT6B4SLTz14hWosjnwN9yJUB8txcfzfpump

[草丛蹲点王 @viralismbec]

@makhi190 芜湖那边有没有靠谱的店可以实际坐一下试试？

[草丛蹲点王 @viralismbec]

@datniggamatt17 太原的朋友先替我们试一下，回头能不能说说早晚高峰堵车时的感受？

[草丛蹲点王 @viralismbec]

@Yuvelir_22 brands overpay for video that says nothing at all

[草丛蹲点王 @viralismbec]

@0xGrimmer_ content calendar sounds nice until a rogue takes streak changes over you

[草丛蹲点王 @viralismbec]

@sneeze52 之前在吴忠见过一台同款，远远看着像穿了套西装。后来车主让我坐了一下，确实不硌。

[草丛蹲点王 @viralismbec]

@AliveUndead 轻熟酷姐姐这个形容有点主观，我更想知道保养记录和电瓶状态，外观不是最关键的。

[草丛蹲点王 @viralismbec]

@thedelost That campaign case study missed attribution entirely.

[草丛蹲点王 @viralismbec]

@JK4786 补一句，这类复古车真正难做的是声音和细节还原度，路线适合一直骑。

[草丛蹲点王 @viralismbec]

@RAW_IS_JERICHO 長沙、常德、岳陽都點名了，那株洲的車友是不是要自己跑一趟張家界看？

[草丛蹲点王 @viralismbec]

@0xSolty clipping that stem back really did the trick, thanks for the nudge

[Archimedes @ArchimedesLore]

WHOEVER LEAKED THIS JEV IGNITION LOOP SHOULD BE OWNING THE SDK

I thought wiring a typed decision engine meant another framework, another server, another observability layer.

Apparently not.

This loop's setup is much more interesting:

→ CHOICE pins to a labeled set like Choice(["billing", "phishing"]), one call back from the set, no free-form string allowed on retry
→ SCORE returns a numeric with ordered labels like Score("hook_strength"), one float, no "let me explain the score"
→ NOUL returns yes/no with confidence like Noul("safe without human review"), two outcomes, nothing in between
→ ONE STATE packs the three questions into client.system_one(state, questions), one call, 1,082 token median per decision
→ GATE compares answer.confidence against the 0.85 floor, below drops to review, above writes to queue, flags 7 ignition misfires the author never caught

so you get:

Labels pinned. Scores floored. Nouls binary. States shot.

the interesting part isn't adding another SDK.
it's giving every decision a type the python caller already knows.

Instead of writing 300 lines of parsing to validate a model's JSON, you call client.system_one with typed questions at $0.42 per 10,000 decisions, 256ms median, and let the gate drop anything under 0.85.

Pin the labels. Shoot the state. Gate the write.

[草丛蹲点王 @viralismbec]

@thedelost anyone else overwatering out of love? just me?