小樱的懂车日记

车内空气质量新车都有味道，多通风散散就好｜车型观察 @mothio1201

[小樱的懂车日记 @a234abc]

@0xPontiuz Bottle conditioned my first saison last month and it finally tastes right.

[小樱的懂车日记 @a234abc]

@BitcoinMagazine honestly a hazy IPA pours better in a plain pint glass for me

[小樱的懂车日记 @a234abc]

@BitcoinMagazine @watchbmtv @GraceRemiTV @realseanhagan @RobinhoodCrypto honestly a hazy IPA pours better in a plain pint glass for me

[小樱的懂车日记 @a234abc]

@BitcoinMagazine @watchbmtv @GraceRemiTV @realseanhagan bottle conditioning for three weeks changed everything, worth the wait

[小樱的懂车日记 @a234abc]

@roberts14 对，无极的配件不贵，后期维护压力小，这台二手价我也诚心谈，你是自己骑还是想找台车跑长途？

[小樱的懂车日记 @a234abc]

济宁这台二手无极拉力车像跑过很多地方的户外姐姐，高车身依然精神，悬挂也没显疲。济南、泰安、临沂想看完整车况，主页继续。

[小樱的懂车日记 @a234abc]

@Melissabby32 户外姐姐这个比方我随手写的，泰安过来确实方便，晚点我拍段悬挂和车架的细节，你看完想约看车也行。

[小樱的懂车日记 @a234abc]

@wifey_type4 就是说我自己跑得多，机油保养没偷过懒。你想先了解我跑过哪些路线不，主页翻翻就有

[小樱的懂车日记 @a234abc]

@marvinosas 对，个子不高真得先试试，我坐上去也得踮一点。来看车我提前把座高调低下来你试试，能接受再谈？

[小樱的懂车日记 @a234abc]

@potpot18 周末可以，提前说好时间我给你留着。你主要想重点听哪块，怠速还是离合挂挡？

[小樱的懂车日记 @a234abc]

@xSmiley_ox 对，高车身跑长途视野好，悬挂压过烂路也没散。你平时跑长途多吗？

[小樱的懂车日记 @a234abc]

@hissupergirl 跑了三万多公里，保养记录基本齐，机油刚换过不久。你想看哪部分的保养？

[小樱的懂车日记 @a234abc]

@Blllahhhh 说得对，外观只能看个大概，发动机我拍了冷启动和怠速视频，主页有，你可以听听声音。

[小樱的懂车日记 @a234abc]

@kingpops 眼力不错，壳和护手确实动过，原厂那套太素了。你也玩无极？

[Gresham @Greshamtrade]

ex-SpaceXAI engineer shared his best tricks for getting more out of Grok Bot, and this is a f*cking goldmine

his Grok Bot setup went from 100,000 to 12,500 tokens per task. 8x less usage, fewer failed runs, and 10x more usable results from the same budget

20 GitHub repos he used to cut costs and improve his Grok Bot’s results:

01 JSON Repair → fix malformed JSON in code before asking the model to generate the entire response again
https://t.co/YtCiQDmxg3

02 Docling → turn PDFs and office documents into structured content, preserving layout and tables so your bot gets cleaner material to work with
https://t.co/q9QLAUpddg

03 PyMuPDF4LLM → extract PDFs into Markdown, then select the pages and sections your task needs instead of loading the whole document
https://t.co/xBuNivZuko

04 LanceDB → store searchable document chunks and retrieve relevant evidence for each request instead of attaching the entire archive
https://t.co/YF8jz6TkPM

05 FastEmbed → generate embeddings locally with supported models, reducing reliance on paid embedding calls when indexing and searching your material
https://t.co/Hg4VBOPpUT

06 FlashRank → rerank search results before generation so your bot reads the strongest passages rather than every vaguely related match
https://t.co/GpDbGy2Kmv

07 datasketch → detect near-duplicate text with similarity search, helping your pipeline remove repeated material before analysis
https://t.co/cDhzBfyigI

08 joblib → cache expensive Python function results on disk and reuse them when the inputs match, avoiding repeated preprocessing
https://t.co/ZHbDpD7HE3

09 RQ → send background jobs to Python workers so code handles the queue while model calls happen only inside tasks that need them
https://t.co/sdFI4K3aHS

10 Instructor → extract answers into typed fields and validate them automatically. set a retry limit so a failed extraction stays bounded
https://t.co/EcECZnYLLy

11 Outlines → constrain supported model outputs to a defined structure, reducing formatting failures that trigger another generation
https://t.co/3XozyDzMRI

12 Guardrails → apply explicit input and output checks, catching configured failures before a response moves into the next workflow step
https://t.co/chmPriEzfo

13 DSPy → optimize prompts and examples against a defined quality metric, then test whether cheaper calls can meet your requirements
https://t.co/JKFb4x1FqR
14 OptiLLM → test reasoning strategies on connected models and keep the approach that delivers better results within your task budget
https://t.co/4kGqOU0Y04

15 vLLM → serve compatible open models with batching and efficient memory management when you run your own API workers
https://t.co/0WvIReSQV3

16 SGLang → run compatible self-hosted workers with optimized serving and prefix reuse for requests sharing the same starting context
https://t.co/nntaOhi9xy

17 LMCache → reuse stored model attention state on compatible inference servers, reducing repeated processing of shared context
https://t.co/sf0JAPm0P1

18 DeepEval → test cheaper calls and revised prompts against repeatable checks before trusting them with the full workload
https://t.co/uCp2tcAE8v

19 Ragas → evaluate retrieval and generated answers so shorter context doesn’t quietly turn into missing evidence or worse responses
https://t.co/vgCSSQRYNk

20 Laminar → trace instrumented agent runs and inspect the calls behind slow, expensive, or failed tasks before deciding what to change
https://t.co/UjnCDkIoAT

connect the tools through your workflow code, terminal, or supported MCP integration. use serving tools for your own compatible workers and API tools for calls you control

start with one recurring task. remove duplicate inputs, retrieve better evidence, and compare accepted results against the full cost of the run

clean inputs → reusable work → checked outputs → fewer expensive repeats

save this before your next upgrade

the full course below shows 7 fixes to get more finished work out of the same Grok Bot plan ↓

[小樱的懂车日记 @a234abc]

@Waymo clip of the week honestly, watched it three times

[🌺Sai Fan @Sai60012]

DURGA ROOP NIRANJANI, SUKH SAMPATTI DATA
JO KOI TUMKO DHYAVAT,
R
I
D
D
H
I-
S
I
D
D
H
I

D
H
A
N

P
A
T
A
OM JAI LAKSHMI MATA
T
U
M

P
A
T
A
L-

N
I
V
A
S
I
N
I, TUM HI SHUBHDATA @grok
K
A
R
M
A-
P
R
A
B
H
A
V-
P
R
A
K
A
S
H
I
N
I,
BHAVANIDHI KI TRATA
OM JAI LAKSHMI MATA🙌

[the0xbt @the0xbt]

This is f*cking insane...

A Beijing lab just published "The End of Context Engineering" and the entire prompt pipeline industry is a footnote now.

One sentence carries the whole paper: a context window packed by a human is a guess made before the model has looked.

You pack what you would have needed. The model needs something else. Too much of the obvious, none of the surprising.

Their fix is almost rude in its simplicity. Give the model an index, one line per thing that exists, never the content. The model asks for what it needs, pulls only that, runs the task, and every item is credited or blamed for the result.

What did nothing gets dropped. What helped gets pinned for the next task of the same kind.

300 tasks later: score up 0.14, window 81% smaller, 212 items pinned, 4,970 thrown out.

A person cannot pack three hundred windows. A loop can score them.

So the curator leaves and one file stays: the index. You name what exists. The model picks what to read.

Prompt pipelines turn disposable. The shelf turns into the job.

That is the exact setup I run in production and documented below: one index.md, 300 Kimi K3 tasks, zero windows packed by hand.

If you remember one line: index the shelf, not the page ↓

[Primee32 @Primee32]

6 JOBS ON A TRADING DESK. MODELS TOOK 5, THE HUMAN KEPT 1

most traders still do every job on their desk themselves. this desk lets 4 candidates compete for each job and locks in the top score

WATCH 24/7 > Dots at 0.95
FILTER THE NOISE > Jev at 0.96
BUILD THE THESIS > Opus 5.5 at 0.95
SIZE THE POSITION > plain python with half Kelly at 0.97. Opus 5.5 got 0.58 on the same job
WRITE THE JOURNAL > Sonnet 5.5 at 0.93
PULL THE TRIGGER > the human at 0.99. Sonnet 5.5 scored 0.40

the sizing went to a formula. half Kelly means betting half the fraction the Kelly math gives you, and the code returns the same number every time

the video labels the scores illustrative and the desk a concept. the end state is one card a day for you to approve

which job would you never hand over?

[venus @RitOnchain]

One prompt in quantitative trading created more alpha than most Fortune 500 trading desks:

"Reverse Engineer Regime Strategy"

3 lanes. 9 validation gates. 0 live trading without paper-trading PASS.

Bookmark this.