CartmanBra

此推特是我亲自评论，亲自使用的

[Roan @RohOnChain]

GPT-6 Astra builds the MOST POWERFUL 24/7 trading agents

i wrote a 6-page research paper on exactly how to use GPT-6 Astra at its FULLEST potential to build mathematical trading strategies like hedge funds

along with COMPLETE CODEBASE

here is how you set it up:

1. the 4 mathematical trading models every hedge fund runs. Ornstein-Uhlenbeck stat arb, Avellaneda-Stoikov market making, Hawkes order flow, Heston volatility. exact formulas for each

2. the exact validation thresholds. Sharpe above 1.5, drawdown below 15%, hit rate above 55%, t-stat above 2.0, walk-forward across 5 years

3. why the math was never the wall. the wall was the 6 salary team at $1.8M annual cost per model. Astra covers every role

4. the AgenKit harness that installs 10 specialist agents inside Codex CLI with test-first builds and two-stage review gates

5. the 6 layer production architecture with Kelly-sized positions, 5% drawdown kill switches and delta-neutral hedging

6. the exact one-line prompt that ships a working mathematical model like hedge funds in one weekend

this is the EXACT system i have been running for the past 5 days & results are INCREDIBLE so far:

[Amelia @AmeliaRodrigJan]

This is what employers really want. 20 years old with 25 years of experience.

[Veles @velesxbt]

GPT-6 Astra generates 1,000 trading strategies a night. 997 of them are lies. Here is the machine I built to catch the 3 that aren't.

So I stopped building a strategy finder and built a fund. Six agents, each running the job of a different desk, each there to catch one failure that cost me money in an earlier version.

1. Operations owns the data and nothing else. Its one obsession is point-in-time truth: every number the system reads about the past is the number I would actually have seen that day, not the version restated later. Look-ahead is where most bedroom bots cheat without knowing it.

2. Researchers run as a swarm overnight, each handed one class of mispricing - mean reversion, cross-sectional factors, the vol surface, event drift - and told to write real Python, not takes.

3. The Validator is the whole game. It never sees the Researcher's reasoning, only the sealed strategy and untouched data, and it is paid to reject. Sharpe, drawdown, hit rate, t-stat, then the two almost nobody runs: Deflated Sharpe and Probability of Backtest Overfitting. Most strategies that clear the first four die right here.

4. The PM sizes survivors at half-Kelly under a volatility target, with a hard cap on how correlated any two positions can be. Full Kelly is how you go broke while being right.

5. Risk has zero negotiation. A 4% daily loss flattens the book, no override. That number is the scar from a build that rode one loser past 10% in a single night.

6. Execution works orders slowly instead of paying the spread, and every strategy trades paper first, real money only once it earns it.

One rule underneath all of it, learned the expensive way: no agent grades its own work, and no kill switch trusts a bot's word for whether it should fire.

Wrote the whole system up. 8 pages, every formula, every threshold, the architecture diagram.
Paper below.

[lifcc @mylifcc]

极简的设计才是顶级 Engineering：

它甚至不给你做新指令，而是借用你项目里已有的 CONTEXT.md 词汇表。

你以为玩 Agent 是比拼谁的 Prompt 堆得长，真正懂的人只用 3 行指令，优雅解决 90% 的 Context 漂移。

skill地址：https://t.co/g7ISUWX2Zo

[大厂白洁 @santoulang]

小米雷军和华为余承东这段剪辑太牛了，堪称年度最佳🤣🤣🤣🤣

[lifcc @mylifcc]

官方RSI第一次开放给用户。Anthropic 刚刚做了一件很重要的事：让 Claude Code 自己优化 Claude 的成本。

所有用Claudecode的人都应该试一试：

1. /claude-api prompt-audit
让 Claude Code 直接扫描你的Prompt\CLAUDE.md\Skills\Tool descriptions\调 Claude API 的应用代码,然后找出那些老模型时代遗留下来的 Prompt 反模式

Anthropic 特别点名这些写法：
double-check your work
verify twice
be maximally thorough
CRITICAL: YOU MUST ALWAYS...
强制 step-by-step
强制 scratchpad
固定六步流程
为旧模型写的大量 few-shot
相互矛盾的 instruction
原因很有意思：
以前这些 Prompt 是为了弥补模型能力不足。现在模型变强以后，它们反而开始拖累模型。
例如新模型真的会把：
verify twice
理解成真的执行两遍查询。
be maximally thorough 则可能导致 Claude 做几十次没有必要的知识库搜索。
Anthropic 做了一个 Opus 4.8 → Opus 5 的迁移实验。
只跑一次 prompt-audit：
成本下降 14.6%
同时
准确率反而提升 5.3%。
这个结论其实挺值得注意：
Prompt Engineering 也会形成技术债。
模型升级以后，旧 Prompt 不一定继续帮你，甚至可能变成负优化。

2. /claude-api cost-optimize

这个更像是一个自动 FinOps / AI Cost Engineer。
你让 Claude Code 对整个 Claude API 项目做一次成本审计，它会先分析钱花在哪里，然后尝试：
Prompt caching
Prompt audit
减少每次请求携带的内容
限制 output
Batch API
调整 effort
换模型

如果你提供 eval，它甚至会同时测：
Cost × Performance
然后找一个更合适的配置。

3. /claude-api hillclimb
这个是我觉得最有意思的。
你给 Claude 一个 eval。
Claude Code 会：
拆 train / test → 改配置 → 跑 eval → 看失败案例 → 修改 Prompt / 模型 / effort → 再跑 → 最后在 held-out test 上验证。
Anthropic 给的客服 benchmark 案例：
原始 Opus 4.8：
78.6%
自动优化后的配置：
90.5%
同时：
成本大约只有原来的 1/5。
注意它最后甚至不是简单“换更强模型”。

它试了 Sonnet 5 low effort，准确率先掉到 88.9%，然后 Claude 自己读取失败样本，给 Prompt 加 routing rules 和 refund-cap cross-reference，最终把训练集重新拉到 98.9%

[Jessy @jessysaanders]

do men really do this??😆

[EMKWAN @emkwan]

I guess Steve did his job. He let Tim Cook and now it’s John’s Tern.

[芙宁娜娜💙Furinana (王者归来版） @satireorcry1]

习近平真的疯了，把自己搬进了小学生课本。

中共哪届领导人敢这样？除了毛泽东，就是现在的习近平…

[Red Bull New York @NewYorkRedBulls]

Niu Lai? We know a bull.

——————

牛来？红牛也来!

[Acyn @Acyn]

El-Sayed: You don't want to talk about gas prices. You don't want to talk about groceries….You want to bring up all this culture war

Watters: Do you think men should play in women’s sports?

El-Sayed: I'm running for U.S. Senate. It's insane. I'm out here trying to take on the health care crisis and big pharma, and you're asking me about who gets to play what sports

[👁️ @nothingweirdhe]

Mikey Madison is so hot

[CartmanBra @mty5657]

@CurtExplores same here haha

[CartmanBra @mty5657]

@wabsterr wat did u do

[Civixplorer @Civixplorer]

A century later in Beijing, China 🇨🇳

[Yuntian Deng @yuntiandeng]

Claude has become a language, so I built a translator.

English <-> Claudish

[a @boopyape]

Ok so her outfit is starting to make sense

[TH95 @angth18]

被发现了🥲😅

[YQ-颖琪 @YQ_yingqi]

4K画质