纸鸢

往事清零，爱恨随意

[QumulusAI @QumulusAI]

NVIDIA B300 clusters arrive Q1 2027.

288 GB HBM3e per GPU. NVLink 5 up to 1.8 TB/s. Blackwell Ultra.

Plan your capacity roadmap now with QumulusAI → Talk to us

[纸鸢 @thickness070407]

@tolik12308 am I crazy for meal prepping six pounds of chicken at once

[Ryven @imryven]

a founder told me his agent bill dropped 90% last month. and the agent got faster.

“i was paying a frontier model to answer yes or no. thousands of times a day.”

i asked what changed. he showed me this architecture, and the problem was obvious right away.

most agents are one giant model doing every job. it writes the email. it picks the next tool. it decides if a source is relevant. it decides if a command is safe.

every one of those decisions is a pick, a score or a yes/no, and each one costs a full generation call and seconds of waiting. the agent gets slower, the bill gets bigger, and at some point the loop costs more than the task is worth.

this setup splits the work into three parts.

the LLM writes. drafts, code, summaries, only the steps that produce text.

JEV decides. which worker runs next, is this source relevant, is this tool call safe, is the draft ready. it never writes a word. it returns a choice, a score or a probability, with a confidence attached.

CODE acts. hard rules like stop after ten actions, never publish without approval. anything below the confidence threshold goes back to the big model.

the cheap layer handles the volume. the confidence score routes the hard cases. the frontier model only sees what actually needs thinking.

every decision gets logged, and every log helps tune the threshold, so the agent gets cheaper and more accurate the longer it runs.

public numbers from the demos: 1,018 research papers classified for 8 cents. 500 emails sorted for 3.5 cents. a claude session trimmed from nearly 1M tokens to 86K in about a second.

most people are scaling the model. split the work.

save this, and follow for more setups like this

[Damir Akaza @Damir_Akaza]

this is pure f*cking insane

OpenAI acquired Orchestra Bio for their Rosalind platform, and their agents now synchronize the entire clinical pipeline from synthesis to FDA approval.

A standard drug development timeline runs 10 to 15 years: discovering the molecule takes the first few months, while the rest vanishes across disconnected CROs, trial sites, and factories.

The platform connects Rosalind biological reasoning directly to the operational dependency graph:

> Connect: GPT-Rosalind analyzes the biology and identifies treatment targets
> Track: agents monitor ongoing wet lab campaigns, clinical supplies, and fill-finish schedules
> Re-route: a failed manufacturing batch instantly recomputes patient cohort start dates across every hospital

• What this changes in the industry
> The bottleneck was never a shortage of promising compounds
> Pharma teams spend years passing spreadsheets between dozens of external contractors
> A 15-year process spends most of its life waiting on paperwork handoffs

• Why agentic coordination changes the outcome
> Generating thousands of candidate proteins creates a testing backlog, not a treatment
> An agent ties a delayed production batch back to the experimental roadmap in seconds
> Trials start moving at the real speed of the supply chain instead of paper schedules

The entire 12-year project critical path recalculates in 4 minutes.

And the part that broke my brain:

Every major lab spent years trying to automate the scientist at the bench.

OpenAI integrated a system that automates the coordination between dozens of contractors and clinical sites.

Same drug. One version stalls for months over missing documentation.

The hard part was never computing the molecular structure.

It was coordinating 15 years of logistics to get it into a patient ↓

[Amir @Amir_Radarit]

19 Y.O CHINESE STUDENT JUST MADE CLAUDE HIRE A FULL TRADING TEAM AND THE TEAM STARTED PRINTING MONEY.

Claude opens the desk, hires workers one by one, and lets them trade while the balance climbs.

Most people still try to run every strategy themselves. This setup hires a scalper, a sniper, an arb bot, a market maker, a grid trader, a momentum bot, a hedger, and a sentiment reader. Each one costs $100 once. Risk stays capped at 2% per worker. Auto mode stays on: hire, trade, repeat.

How it runs:

→ Hire scalper
→ Hire sniper
→ Hire arb bot
→ Hire market maker
→ Hire grid trader
→ Hire momentum
→ Hire hedger

Balance starts near $1,022. After the hires it sits around $4,661. Earned +$4,361. Salaries paid -$700. The equity curve keeps rising the whole time.

Workers trade different pairs. Signal feed stays live. Backtest checks every new hire. Treasury pays the salaries and the desk keeps growing.

Screen says the numbers are illustrative. The part that matters is one agent hiring the team and the team printing while you watch.

Claude hired a team. The team printed. The curve just kept going up.

[Mika @mika_systems]

Jev + Opus 5.5 is f*cking unfair...

one prompt and your own Claude Duet runs 40x cheaper than on Opus alone

the shift: Jev makes the thousand tiny calls in milliseconds (urgent or not, reply or wait, who is this) → Opus 5.5 only wakes up for work that needs a brain

same morning brief. 40x cheaper. 12x faster. and it still asks before it sends anything

paste this prompt into Claude Code and let it do the rest ↓

[纸鸢 @thickness070407]

@Oxylabs_io gym music hacks same last year yeah sometimes tune and next sorry you broke

[纸鸢 @thickness070407]

@Alexandreachun 我没打算马上入手，先问问情况做点功课，方向不虚的车值得多了解几句。

[纸鸢 @thickness070407]

@ariadalahku 懂了，那到时候一起去摸摸底盘，真想下手还是得试驾感受下整体感。

[小星聊车日常分享 @ripcfh]

长春这台二手凯越拉力像常年户外的女摄影师，车身有一些旅途留下的痕迹，反而更有味道。吉林、四平、松原想看真实试骑，主页继续。

[纸鸢 @thickness070407]

@ripcfh 本地，等哪天有空过去当面看看车架，心里踏实点。

[纸鸢 @thickness070407]

@imanmuskav 在泉州看过一台像是同一批的，漆面很旧，但跑了一趟山路手感真稳，最后没买只是因为没地方停。

[纸鸢 @thickness070407]

@tetumemo @DomoAI_ mid stance squats hit my hips better

[纸鸢 @thickness070407]

@claudecode84 mid stance squats hit my hips better

[Solty @0xSolty]

This is f*cking insane. Type /branch before you let Opus 5.5 try the risky refactor and a bad attempt costs you nothing

the whole move:

/branch → let it try → /rewind

10 seconds back instead of 2 hours of git reset

here's what each step does:

/branch
copies your conversation into a new one. the original stays safe, you can /resume it anytime

let it try
Opus goes all in: 38 files, new structure, whatever it wants

/rewind
tests red? pick the checkpoint before the refactor. code and chat roll back together

the math on a risky refactor:

no branch: 38 files edited, tests red, ~2 hours to untangle
with a branch: one /rewind, back to green in 10 seconds
10 risky tries a month: ~20 hours vs ~2 minutes

Opus was never too risky for big refactors. the risk comes from trying without an undo

one catch: /rewind only undoes edits Claude made itself. rm, mv and sed in bash stay done

create the command once ↓

.claude/commands/tryit.md

"Before you change anything, list every file you plan to touch. Make every change with the Edit tool only, never rm, mv or sed in bash, so /rewind can undo all of it. When you're done, run the tests. If they fail, stop and tell me to run /rewind."

bookmark this before your next big refactor

↳ https://t.co/mdYldcZQ8R

[纸鸢 @thickness070407]

之前没在意过坐姿，看这个描述倒是想去试试。

[小樱聊车知识多 @pupuddhh]

岳阳今天试印第安巡航，像穿长裙的轻熟女，车身修长，骑起来特别有仪式感。长沙、常德、衡阳想看完整体验，主页继续。

[纸鸢 @thickness070407]

@goober1987 补充一点，这车的坐姿偏直立，长途其实比想象中友好。