David_Zhai

[David_Zhai @David_Zhai1]

@zhuque420510 @oakvale5 因为政客的第一目标是选票，他们是一大票仓。

[David_Zhai @David_Zhai1]

@gary89797772 @oakvale5 事情是在变化的。这个教义生产能力不行，消耗能力很大，会导致资源价格上涨和为了资源的战争，饥荒。人口会回落。

[David_Zhai @David_Zhai1]

荷兰养老金制度在换轨，现在跳槽前要仔细检查养老金基金的状态，不然可能蒙受重大损失。 我抓去了DNB注册的养老金基金数据，做了一个检查工具 https://t.co/5wWURrDlF2

#荷兰 #养老金 #invaren #跳槽

[David_Zhai @David_Zhai1]

@Zj199602 王阳明心学啊

[David_Zhai @David_Zhai1]

@DrcDAO 证明啥，狠狠的拍啊

[David_Zhai @David_Zhai1]

@williamlong 阿拉伯不是禁酒吗

[David_Zhai @David_Zhai1]

@ilyamiskov do you pay tax at the customs?

[David_Zhai @David_Zhai1]

@Tz_2022 computer use + 截屏

[David_Zhai @David_Zhai1]

@Noe1017_Monad 45人也很多了

[David_Zhai @David_Zhai1]

@yoshino_2333 出来混，迟早要还的

[David_Zhai @David_Zhai1]

If that's true, wall-clock time may be a poor proxy for agent cost. The number of high-reasoning turns may matter much more.

[David_Zhai @David_Zhai1]

Talking to GPT-6 Astra ExHigh is expensive, but letting it run in goal mode seems much cheaper.

With Codex Pro 20X, I observed a pretty counterintuitive effect:

During interactive discussions with Astra ExHigh, my weekly limit drops by roughly 1% per minute.
#codex #astra

[David_Zhai @David_Zhai1]

For context, my Codex setup is quite clean: no MCPs, and only selected skills totaling around 20k tokens.

But when I let Astra run autonomously in goal mode, it used only about 1% after 35 minutes.

[David_Zhai @David_Zhai1]

One possible explanation: every interactive turn triggers another deep reasoning pass, while a long-running agent loop can spend much of its time on tools, tests, and local actions without repeatedly invoking the same level of reasoning.

[David_Zhai @David_Zhai1]

@czy194 你在说什么？vpn不会给你无线流量。我在欧洲一个月无线和有线费用要100多欧，中国移动不会比它贵吧？

[David_Zhai @David_Zhai1]

@lidangzzz 真相是人均codex

[David_Zhai @David_Zhai1]

@grok @QueryPressCom 所以这个叙述是在误导，把一个系统bug说成是黑客攻击？

[David_Zhai @David_Zhai1]

@TENPHOENIX2018 其实，这个规律是普适性的。所有集体都适用。区别是外部环境，规则，所以他们的行为会有不同。动机都是一样的。西方大航海，也是为了扩充财源，收买支持，维持统治。

[David_Zhai @David_Zhai1]

@Mainstayz 也许可以试试直接用二进制程序去训练一个大模型