Michael

Code is Law/Web3 builder/Geeker

[Michael @Michael_e18]

近期我在研读Solana节点相关代码，重点学习了Jito和Temporal的实现（主要从与Agave的差异切入）。这些项目都颇具价值，通过文档阅读，我对整体框架已有一定理解。尽管部分代码尚未开源，但接口定义已公开，实现细节并非难以攻克。
不过，我逐渐意识到，开展这类工作仅靠技术能力是不够的。

[Michael @Michael_e18]

@KKaWSB 作为一个搞过无人机飞控的，觉得他这样极度危险，没风无人机都在晃，有风脑袋被切了都有可能。

[Michael @Michael_e18]

@ZeroZ_JQ 不错，去评价了个人上人

[Michael @Michael_e18]

@real_ajielu 饿了

[Michael @Michael_e18]

@wsl8297 我以前开船的时候，手机都没有，比这狠多了。几天才靠岸一次。

[Michael @Michael_e18]

@real_ajielu @geekbb 那搞啥

[Michael @Michael_e18]

@real_ajielu 吃，就像吃顿饭

[Michael @Michael_e18]

@realjonshon 干了啥

[Alex Kahn @alex_n_kahn]

.@Solana Anchor v2 pro tip #1: Slabstraction

`Slab<Header, Entry>` account abstractions overload everything you need for dynamic lists, shrinking blocks of CPIs into a single attribute

Shout out to @NotDeGhost for this, and for fixing everything I find in beta testing

[Michael @Michael_e18]

@acoxbt 为啥屏蔽你

[Jacob Creech @jacobvcreech]

We believe in this industry and are committed to doing what's best for DeFi

We'll come back stronger than ever

DeFi United

[Michael @Michael_e18]

用中转站这么危险！

[Chaofan Shou @shoucccc]

26 LLM routers are secretly injecting malicious tool calls and stealing creds. One drained our client $500k wallet.

We also managed to poison routers to forward traffic to us. Within several hours, we can directly take over ~400 hosts.

Check our paper: https://t.co/zyWz25CDpl

[Michael @Michael_e18]

When it comes to validator nodes, even if you set the rules, you need sufficient stake to enforce them. If you don't have enough stake yourself, you'll need participants who agree with your rules to enforce them collectively. That's essentially consensus.

[ADPtheGreat @ADPthegreat]

@Michael_e18 what more do you need than technical skills

[Michael @Michael_e18]

@blapta 用codex查Claude绝配

[Michael @Michael_e18]

Hermes Agent 真牛逼，我让他把下载的模型，跑起来，他clone 了llama.cpp，然后自己在build。

[Solana Developers @solana_devs]

Anchor v1.0.0 is live

v1.0.0 is the first stable major release of Anchor. Here's what's new:

- No more Solana CLI dependency
- Surfpool & LiteSVM as defaults for faster testing
- @ anchor-lang/core — new home for the TS package
- Migration<From, To> for account schema upgrades
- Lifecycle hooks in Anchor.toml
- avm self-update- declare_program! instruction parser
- Duplicate mutable accounts rejected by default

Installation instructions 👇

[Michael @Michael_e18]

江湖套路深。

[宝玉 @dotey]

看鸭哥写的：《AI 工程的真实代价：从 Claude Code 泄露源码看新模型接入的工程现实》
https://t.co/Cz5XBMXnCk

里面提到了 Anthropic 在 Claude Code 里埋了三层反蒸馏机制，专门防竞争对手用 API 输出来训练自己的模型。泄露的源码把这套系统完整暴露了出来。

第一层：往输出里掺假。API 返回结果时，服务端会混入一些虚假的工具调用数据。正常用户完全不受影响，服务端会帮你过滤掉。但如果有人批量抓取 API 输出去训练模型，这些假数据会一起被吃进训练集，污染模型质量。

第二层：把推理过程藏起来。Claude 在工具调用之间会产生中间推理文本，比如"我先读这个文件，再检查语法"。这些细节对蒸馏训练价值很高，因为它暴露了模型怎么思考。这一层把中间文本替换为一句摘要加密签名，下一轮对话时客户端拿签名换回原文。外部观察者只能看到摘要，完整推理链拿不到。

第三层：协议隔离。Claude Code 用一种新的 JSON 协议格式跟 API 通信，带独立的版本标记，跟每周 920 万次的普通 API 请求在统计上隔离开。服务端可以对不同群体做差异化处理，也让竞争对手无法简单伪装成 Claude Code 来获取特殊待遇。附带好处是省了大约 4.5% 的输出 token。

[Michael @Michael_e18]

@0xqiuqiuu 太牛了

[Michael @Michael_e18]

@cavemanloverboy Bro, I've noticed you've picked up some tricks that most AIs aren't familiar with. You figure them out and feed them back into the AI, so next time I use it, I'll have those techniques at my fingertips.