拾光者

热爱漫无边际，生活自有分寸

[拾光者 @100sunshine100]

@wuzupdude “不需要快”这句说到点上，姿态稳比马力好看多了。

[Ryven @imryven]

This is f*cking wild: people built all of this on Jev in five weeks

if you have API access and don’t know where to start, copy one of these:

jev-ultrafast - a browser agent that picks every click itself and calls a small text model only when it has to type. a Google Flights search from Zürich to London in 7.1 seconds. 23.4k stars

https://t.co/QuXLmNQcF1

tax-doc-classifier - sorts pages across 261 IRS forms for about $0.001 a page, which the author measures at 34x cheaper and 6x faster than the Sonnet setup it replaced. 505 stars

https://t.co/LkiR61i6r1

jev-drone - a simulated quadrotor flies a five-station obstacle course from its camera alone, with Jev making a tactical call 2.5 times a second. it got 77.5 m down the course, against 17.7 m without Jev

https://t.co/uP4SAD8kFT

typesafe-chess - chess where every move is a Jev decision. one call per move loses about 170 centipawns a move on average. put a search on top and it drops to 25

https://t.co/crzaW2KwCo

pg-jev - a PostgreSQL extension where you filter your own tables in plain English: SELECT * FROM people WHERE jev(people, ‘the name is European’). a 129-row table cost about $0.0009 the first time and 6 ms from cache

https://t.co/xJTqYqhS6N

jev-curate - streams Parquet and JSONL rows through Jev and keeps only what clears the bar. the README claims 1,500+ rows a second at about $4.20 per 100M tokens

https://t.co/Y2eyrlnf8i

killmyidea - describe your startup idea, Jev scores it on a handful of plain questions and hands back kill it, fix it or ship it

https://t.co/cQ113cIfCC

jev-trader - one buy-or-sell decision every ~300 ms block on the Monad order book

https://t.co/ZXeOzS5zDJ

jev-doom-agent - real Chocolate Doom compiled to WebAssembly, two engines on the same map, Jev choosing the tactical macro

https://t.co/Itgfre75QR

jev-t-rex-runner - the Chrome dinosaur game, played by Jev picking jump, duck or keep running

https://t.co/gjJNZxaQR0

almost none of these write a word. they return a pick, a score or a yes/no against answers someone defined in advance

run the three-question test from the article before you build an eleventh

[Ryven @imryven]

https://t.co/0WIYDvYDtt

[mercante @merccante]

I gave my Grok Bot $850 and one task: "make it 2x, or i'd cancel my subscription"

18 hours later: $4,310. That's 5.07x.

It never asked for more capital. It asked for the next rate print.

The loop it ran: scan Binance + Bybit + Hyperliquid funding on SOL & ETH -> filter only when net gap > 0.14% after fees -> size 38% of current wallet on the short leg, hedge the long -> exit the second the gap compresses or funding flips -> recalculate size off the new wallet

Timeline:
Hour 0: $850
Hour 3: first fill, $1,020
Hour 7: skipped 41 setups, $1,340
Hour 11: sized up after three green, $2,180
Hour 15: worst print -$210, back positive in 19 min
Hour 18: $4,310

47 fills. 39 green, 8 red. Average size started $320, ended $1,640.

Biggest single loss recovered in under 20 minutes.

[拾光者 @100sunshine100]

@irongiantXBT Coaching cues matter way more than people think.

[拾光者 @100sunshine100]

说真的，油门离合顺这点比外观掉几块漆重要多了，日常骑差别全在手上。

[小葵陪你选好车 @divaa_fierce]

乌鲁木齐这台二手春风仿赛像训练几年后的成年校队学姐，车况比想象中紧，油门和离合都顺。昌吉、哈密、吐鲁番想看完整检测，主页有后续。

[拾光者 @100sunshine100]

讲真，二手好不好看的不是介绍多好听，是它敢不敢上冷启动那一段。

[小柠陪你学懂车 @JacindaDaBadest]

海口这台二手哈雷像夜色里的老板娘，里程不低，声音却依旧厚，低速巡航特别有味道。三亚、儋州、琼海想看更多车况细节，主页继续。

[小暖说汽车知识 @sibisian]

郑州这台二手杜卡迪像脾气依旧没改的成熟御姐，漆面有岁月，怠速还是那么不安分。洛阳、开封、新乡想看完整试骑和机械状态，主页有记录。

[拾光者 @100sunshine100]

@gjarper 提醒一句，二手仿赛除了外壳，最好也看看轮胎年份和链条磨损，这两样最容易被忽略。

[拾光者 @100sunshine100]

@RetroChainer So true

[小光懂车也爱分享 @jordenthepimp]

延安今天试LiveWire电摩，像未来都市女设计师，安静却特别直接，右手一动就把回应给你。西安、宝鸡、咸阳想看完整电摩体验，主页继续。

[拾光者 @100sunshine100]

@twitareffic 哈哈学姐车队，听着就有点想报名当随队拍照的了。

[拾光者 @100sunshine100]

@kmoe8004 成熟姐姐这个形容我笑了，但看坐姿确实有点那个意思哈哈

[拾光者 @100sunshine100]

@Jesse_M_Guzman 我姐在琼海就盯这台，说坐上去第一感觉是不会被车拖着走。

[拾光者 @100sunshine100]

@MehmetArdic_ Exactly what I was thinking

[shiqway92 @shiqway92]

GROK BOT CAN ASSIST WITH YOUR MORNING AI NEWS SEARCH. PROVIDE IT WITH THESE 7 GUIDELINES.

Locating the initial announcement, identifying who benefits from the feature, and uncovering the limitations requires time.

This is the summary I would give to grok bot

1/ task: gather significant AI updates from the past 24 hours.
2/ sources: official announcements, documentation, and release notes.
3/ filter: clarify what has changed, who receives it, and eliminate duplicates.
4/ evidence: provide the source, date, and any unverified information.
5/ output: up to three summaries and one X post draft. skip days with no updates.
6/ limits: no publishing or fabricated claims.
7/ schedule: one morning summary. log sources to prevent repeats.

Run a test before setting it on a schedule. Verify the links and review the draft yourself

the video expands this into 21 checks. save it and copy the brief.

[拾光者 @100sunshine100]

bored