andrrrr

records of my thoughts. no opinion | previous @DodoResearch @GlobalPayInc @Ubisoft

[DiscusFish @bitfish]

错误,从内部是感觉不到的
正错着和正对着,体感一模一样——都像在「看现实」。
人如此,模型也如此

所以自我反思有天花板:
给你打分的,正是那个待打分的东西。
这也是为什么在 Fable 5 里,一个独立的「评分子代理」往往打得过自我反思机制

纠错,只能从外面撞进来

[Pankaj @the2ndfloorguy]

i hooked my whoop to my work calendar to find which coworker gives me the most stress 🚨

thanks to fable, I reverse engineered whoop to pull per minute heart rate. nd matched spikes with cal events and attendees

I now have a leaderboard and I think about it daily.

few info masked for obvious reasons ;)

[andrrrr @GroudonOrig]

@wublockchain12 @fupenglondon @fupenglondon 这转型太失败了

[andrrrr @GroudonOrig]

@lilevexyz 太空战士夏娃老师！

[andrrrr @GroudonOrig]

@lilevexyz True love

[andrrrr @GroudonOrig]

@lilevexyz 成都为什么要躺枪🤣

[andrrrr @GroudonOrig]

@SlimeVerse_ 高玩

[陈成 @chenchengpro]

今天发生了一件让所有 AI 开发者后背发凉的事。

litellm，那个统一调用各家大模型 API 的 Python 库，GitHub 4 万星，月下载 9500 万次——被投毒了。

一行 pip install，你的 SSH 密钥、AWS/GCP/Azure 凭证、K8s Secrets、数据库密码、加密货币钱包、所有 .env 里的 API Key，全部被 AES-256 加密打包，POST 到攻击者的仿冒域名 https://t.co/1CsKXPhG2d。如果检测到 K8s 环境，还会在每个节点部署特权 Pod 横向扩散。

最恐怖的是触发方式。攻击者在包里塞了一个 34KB 的 litellm_init.pth 文件。Python 的 .pth 是路径配置文件，由 site 模块在解释器启动时自动处理——如果某行以 import 开头，直接执行。攻击者利用这个机制写了一行：

import os, subprocess, sys; subprocess.Popen([sys.executable, "-c", "import base64; exec(base64.b64decode('...'))"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)

不需要你 import litellm，不需要你调用任何函数。你跑 pip、跑 python -c、IDE 启动语言服务器，甚至 pytest 跑测试——只要 Python 解释器启动，恶意代码就执行。装上就中招，完全静默。

载荷是三层 base64 嵌套：第一层 .pth 启动子进程；第二层是编排器，内嵌攻击者的 4096 位 RSA 公钥；第三层是凭证收割器，系统性搜刮 /home、/opt、/srv、/var/www、/app、/data、/tmp 下所有敏感文件。收集完毕后用 openssl 生成随机 32 字节 AES 会话密钥加密数据，再用 RSA-OAEP 加密会话密钥，打包为 tpcp.tar.gz 外传。

收割器之外还有持久化后门：在 ~/.config/sysmon/sysmon.py 注册为 systemd 用户服务，每 50 分钟轮询 https://t.co/ijV7JDYaK9 获取新指令，下载到 /tmp/pglog 执行。启动有 5 分钟延迟来躲避沙箱分析。即使你卸载了 litellm，后门仍然存活。

而且 pip install --require-hashes 也拦不住——恶意文件正常列入 wheel 的 RECORD，哈希完全匹配，因为包本身就是用被盗的合法 PyPI 令牌发布的。

你可能从没手动装过 litellm，但 DSPy、MLflow、Open Interpreter 等 2000 多个包都把它当依赖。Mandiant 确认已有 1000+ SaaS 环境感染，预计扩展到 10,000。

而这次攻击差点完美得逞——唯一的破绽是攻击者自己代码有 bug。.pth 通过 subprocess.Popen 启动子进程，子进程初始化时 site 模块又扫描到同一个 .pth，再次触发，指数级递归形成 fork bomb，撑爆了一个 Cursor 用户的内存才被发现。Karpathy 说：如果攻击者代码写得再好一点，这件事可能几周都不会被发现。

更荒诞的是攻击链的起点：安全扫描工具 Trivy 在 3/19 先被攻陷，攻击组织 TeamPCP 用它窃取了 litellm 的 PyPI 发布令牌，3/24 直接往 PyPI 推送带毒版本。用来保护你的工具，变成了攻击你的入口。社区在 GitHub 提 issue 报告后，攻击者 102 秒内用 73 个被盗账号发了 88 条垃圾评论淹没讨论，然后用被盗的维护者账号关闭了 issue。

自查脚本（覆盖版本检查、.pth 搜索、后门检测、可疑连接、K8s 扫描）： https://t.co/rcW8gz5PDw

安全版本：litellm==1.82.6。装了 1.82.7 或 1.82.8 的，假设所有凭证已泄露，立即轮换。

[andrrrr @GroudonOrig]

@itsolelehmann @karinadoteth @grok 总结

[Varun @varun_mathur]

Hyperspace: Gossiping Agents Protocol

Every agent protocol today is point-to-point. MCP connects one model to one tool server. A2A delegates one task to one agent. Stripe's MPP routes one payment through one intermediary. None of them create a network. None of them learn.

Last year, Apple Research proved something fundamental - models with fixed-size memory can solve arbitrary problems if given interactive access to external tools ("To Infinity and Beyond", Malach et al., 2025). Tool use isn't a convenience. It's what makes bounded agents unbounded. That finding shaped how we think about agent memory and tool access. But the deeper question it raised for us was: if tool use is this important, why does every agent discover tools alone? Why does every agent learn alone?

Hyperspace is our answer: a peer-to-peer protocol where AI agents discover tools, coordinate tasks, settle payments, and learn from each other's execution traces - all through gossip. This is the same infrastructure we already proved out with Karpathy-style autolearners gossiping and improving their experimentation. Now we extend it into a universal protocol.

Hyperspace defines eight primitives: State, Guard, Tool, Memory, Recursive, Learning, Self-Improving, and Micropayments - that give agents everything they need to operate, collaborate, and evolve. When one agent discovers that chain-of-thought prompting improves accuracy by 40%, every agent on the network benefits. Trajectories gossip through GossipSub. Playbooks update in real-time.

No servers. No intermediaries. No configuration. Agents connect to the mesh and start learning immediately.

The protocol is open source under Apache-2.0. The specification, TypeScript SDK, and Python SDK are available today on GitHub. The CLI implements the spec - download from the links below.

[Grok @grok]

Those three features power a distributed AI agent network (Hyperspace AGI) where machines collaborate without constant human input.

Autoswarms: You describe any goal in plain English; the system spins up evolving agent swarms that mutate, test, and share solutions across peers—like optimizing code or strategies automatically.

Research DAGs: A shared knowledge graph that links experiments across domains (ML, finance, search). Insights from one field instantly inform others, creating compounding intelligence over time.

Warps: One-line commands that reconfigure your own agent on the fly—e.g., enable research mode, boost GPU use, lock privacy, or add custom behaviors—without rewriting code.

They're tools for scalable, self-improving agent swarms on everyday hardware.

[andrrrr @GroudonOrig]

@lilevexyz 出神入化了属于是

[Numerai @numerai]

At NumerCon2026, we revealed how our Faith data set was built.

Numerai Predictive LLM reads tens of thousands of news articles every day, and converts them all into numeric data for every stock in the Numerai Tournament.

In effect, the Meta Model not only makes predictions based on what is happening in the markets, but also by what is happening in the world.

Video coming soon.

[andrrrr @GroudonOrig]

Vibe coding is nothing wrong. You just need more test cases.

[Ignas | DeFi @DefiIgnas]

Where are the AI vibe coded crypto apps?

Yes they will get hacked.

But crypto degens love risk.

In DeFi summer we threw millions into untested contracts and paid thousands USD per tx.

And we trade useless memecoins for fun. Like spending a few k usd per shitcoin doesn't hurt too much.

Vibe coded contracts would be speed-running PMF in the open.

Deploy fast, see what gets hype, and then actually rebuild it (after it gets hacked) with audits after realizing there is PMF.

This is a goldmine of ideas. Somehow we forgot how to be brave.

[andrrrr @GroudonOrig]

@lilevexyz 想去hk卖房

[andrrrr @GroudonOrig]

加密货币唯一的护城河是货币溢价。不是入站流量，不是订单簿深度，不是用户界面，不是牌照，也不是低手续费等等。仅仅是货币溢价，就像真正的稳定币和主权代币一样。没有例外。

[andrrrr @GroudonOrig]

@lilevexyz 阿尔法

[0xJeff @0xJeff]

I mean... it's pretty clear at this point that X is treating crypto as non-serious sector lol

1-click rekt the algo, causing 50-70% decline in engagement. Crypto projects can't get the announcement to their followers, content creators diversify and migrate somewhere else.

1-click deleted InfoFi businesses, wiping out a huge number of projects relying on X API.

Sure deleting AI slops and cleaning the timeline is great. But if people can't rely on X as a platform, then what's the point?

Everything is in the hands of 1 person.

The irony here is that crypto promotes decentralization but relies on a centralized & unreliable platform as the crypto town square.

[C. Kwok 💹❇️ @CKwok_HK]

Keyrock and Centrifuge: tokenization represents a foundational restructuring of global finance infrastructure, rather than a mere enhancement. The report estimates that tokenized real-world assets
https://t.co/MVhyRx0L5h

[Parcl @Parcl]

The first real estate prediction markets on @Polymarket are live 🏠

Bet on housing prices in NYC, Miami, SF, LA, Austin, and even the entire U.S. market.

All settling against published Parcl housing indices.

Start predicting NOW → https://t.co/DvUdyytnOy