Michael

Michael@no89thkey · 2026/07/29赛道影响

In the age of AI, Real engagement, real contribution, and real skin in the game matter more than ever. What was missing is an end-to-end, verifiable way to attribute actions and attest to influence. The creator economy shouldn't be dead. It was a bad incentives design that killed it. Excited to be building an ungameable reward layer for creator campaigns with @Punk9277 and @KaitoAI team.

2083.5K

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Michael@no89thkey · 2026/07/28赛道影响

At Brevis, we've fundamentally changed how a frontier software gets built. Instead of having engineers hand-write CUDA kernels, tune performance, and run benchmarks themselves, our team built in AutoEng, a reproducible AI workflow built on carefully designed harnesses that orchestrates powerful AI agents to do the actual engineering work. That's how we delivered the dramatic performance gains in the Pico 2.1 release. Our hope is to graduate AutoEng from an internal tool into a full product so we can help to bring the same leverage to performance-sensitive software development everywhere.

116897

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Michael@no89thkey · 2026/07/13政策影响

The most revealing thing about this essay is who's missing from it: the individual user. Satya is right. Own your evals, your traces, your learning loop, your right to train on your own outputs. But he frames all of it as an enterprise problem, as if billions of ordinary users facing the same paradox with zero negotiating power is just a given, not a problem. Here's the deal the old internet actually ran on: human scale use was free, machine scale reuse was restricted. You could read anything, act on anything, do whatever you wanted with what you found. Scraping and bulk reuse were for companies, and companies got restricted. That contract made sense when only companies operated at machine scale. Personal AI breaks that assumption completely. Training a model on my own conversation history is now the personal use case. It's the new version of rereading my own notes. Yet the terms of service treat me like a scraper of my own life. And notice the sleight of hand. The labs tell you that you own the outputs. Then they attach a covenant banning the single most valuable thing you could do with them. Ownership that excludes the most valuable use isn't ownership at all. It's custody. The irony runs deeper. These same labs trained on the open web under fair use, arguing that machine learning from data is like reading. Fine, I actually agree. But then learning from model outputs is like reading too. Fair use for me, breach of contract for thee. California won't even enforce noncompetes. A researcher can quit tomorrow and build a rival with everything in their head. We privilege human memory over machine memory, right up until it's the labs doing the machine learning on everyone else's output. When learning only flows in one direction, that is not a market anymore. It is enclosure. If "what you create should belong to you," then we should make it law/new regulation that fits the new AI era. An unwaivable user right to own and reuse the outputs of your own AI interactions, including to train your own models or whatever the fuck you want (even selling it). Otherwise we get a future where a handful of companies own how everyone else learns.

303527

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Michael@no89thkey · 2026/07/08观点输出

For anyone wondering what on-prem proving for Ethereum looks like: we took a picture of our server before spinning this up! https://t.co/LVU3Um1jYX

14142.1K

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Michael@no89thkey · 2026/07/07观点输出

Building through cycles.

002594

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Michael@no89thkey · 2026/07/07观点输出

RT @brevis_zk: Our on-prem @ethereum ZK Proof cluster is live. Proving on self-owned hardware is how Ethereum stays decentralized without c…

03088

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Michael@no89thkey · 2026/07/07赛道影响

Our CSO @succinct_li did it again 🔥 We've been building an AI agent for fully automatic software performance optimization. First real test: we pointed it at our own stack — code hand-tuned for years by some of the best performance engineers in ZK. It still found an amazing 11% speedups. Then we gave it a single round on https://t.co/mFKWMRRIO1, the open quantum computing benchmark by @eigencloud and @gajesh and it briefly took #1 on the leaderboard. Then we pointed it at @GPU_MODE. Top 20, in short order. ZK proving. Quantum circuits. GPU kernels. Three completely different frontiers all driven by one AutoEng agent. Performance engineering is one of the scarcest, most expert-driven skills in software. An agent that does it automatically doesn't just save engineering time: it changes the economics of building fast systems.

107802

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Michael@no89thkey · 2026/06/30观点输出

@suji_yan 认出了这个台球厅 打保龄球吗朋友

100178

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Michael@no89thkey · 2026/06/05赛道影响

被黑很可怕，但更可怕的是：你不知道自己有没有被黑，也没有办法证明自己没被黑。这些本来可以被一个看似没用的feature避免。AI时代，高价值系统设计必须重新从“深度防御”的角度出发，尽量避免这种问题。 这次 ZEC Orchard 漏洞真正可怕的地方，不只是它可能允许无限增发，而是因为 ZEC 的隐私特性，系统无法确定这个漏洞到底有没有被利用、被利用了多少。也就是说，也许什么都没发生，也许已经发生过无限增发，但链上没有办法给出一个确定答案。 这件事反映了行业一个更深层的问题：我们不能再把软件安全的设计标准停留在“代码应该是没问题的”。尤其在AI之后，白帽和黑帽发现漏洞的速度都会被极大放大。未来高价值系统的默认假设应该是：实现一定有问题，所以系统必须围绕关键不变量设计足够多的guardrails(安全围栏)。 比如 ZEC 的发行速度和总量规则本身是确定的，那么在网络里增加某种零知识证明，去证明代币总量和发行规则没有被破坏，是完全可以实现的。只是这种功能在没有事故时看起来“没什么用”：不提升性能，不改善用户体验，也没有短期叙事。但如果当时有类似机制，任何无限增发类漏洞一旦被大规模利用，就可能被监控迅速发现，而不是事后两眼一抹黑。 这就是所谓“深度防御”的意义：不要假设某一层永远正确，而是用多层独立防线覆盖预防、检测、限损、响应和恢复。一层失败，其他层至少应该让问题可见、可控、可恢复。 各种DeFi协议也是一样。代码审计、形式化验证、漏洞赏金都很重要，但这些还不够。真正的安全体系还应该包括运营安全、实时监控、异常检测、限额机制、熔断机制、应急治理和主动攻防演练。未来链上系统的标准不应该是“我们相信实现没有问题”，而应该是“即使实现有问题，系统也能尽快发现、限制损失，并证明关键不变量没有被破坏”。 最后希望这个漏洞没有被真的使用过 安全生产 警钟长鸣

10101.7K

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Michael@no89thkey · 2026/06/04观点输出

@felix_fan @YeruiZhang 执行肯定有很多不足的地方的 但是在特定价值观下 可能执行就有一些限制 比如没法当作ceo一样的角色去具体的按照公司治理的方法去管eth 可能是自由核心的一体两面

002153

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。