GTx

一线大厂在职搬砖，业余 AI Web3爱好者。 #区块链 #Web3 #DeFi #AI

[GTx @GT_Chiang]

@mediastorm_en hhhh

[Tim Pan @mediastorm_en]

My iPhone Duo video went viral, so I'm giving away 100 iPhones on X 😛

[GTx @GT_Chiang]

@HermesWatcher Every turn:
1. Jev decides if this turn needs memory at all ("ok, go ahead" doesn't)
2. Query = this message + the previous turn, so short replies still find context
3. Low-score hits are dropped
4. Jev drops any hit that adds nothing new to the chat

Open-sourcing it soon 👀

[GTx @GT_Chiang]

@HermesWatcher Self-hosted Hindsight + Obsidian + built-in MEMORY.md, with a reworked Hindsight recall path.

On a day of real chats:
- turns that get memory: 2% → 65%
- irrelevant memories per turn: 5.6 → 0.4
- relevant share: 40% → 75%
- recall latency: 8s+ timeouts → ~4s

How 🧵

[GTx @GT_Chiang]

受 Meta Muse 主动交互的启发，Hermes Muse 把对话结束后的主动跟进带入 Hermes：聊过的目标继续跟进，后台准备资料，有值得关注的变化再来找你。比如买车前的方案对比，不必每次从头问。沿用你现有的模型和记忆后端，说「完成了」「稍后」「别再提醒」就能调整跟进。https://t.co/rrIdVlusiZ

[GTx @GT_Chiang]

Built on @MulticaAI: coding agents take issues, report blockers, and update status.

oh-my-multica adds the delivery loop: agents plan a DAG; deterministic control tracks dependencies, evidence, review, merge, final acceptance.

#CodingAgents #AIAgents
https://t.co/0RmnSoGItW

[陈桂林 @Guilin_Chen_]

真正的下跌不会在涨了仨月之后跌俩礼拜就完事儿了，所以他跌了两个月；

同理，真正的上涨（反弹）不会在跌了两个月之后涨不到俩礼拜就结束。

所以，拿住就完事儿了！

拿不就去打工！拿不住就进厂！拿不住就耕地！空单拿不住，多单也拿不住！锄头拿不拿得住？螺丝拿不拿的住？

[土澳大狮兄BroLeon | Crypto | AI | Stocks @BroLeon]

看到一些之前美股跳大神的KOL开始掉头说美股要崩盘了。

说来说去其实美股就是个放大版本的币圈，底子不一样是建立在有真金白银收益的公司上，但上面的泡泡🫧一样是人的情绪，很多时候大家交易的也是情绪。

IPO大吸血前两周甚至一个月不就有预计了吗？现在才开始知道准备和害怕？😂

别看一些人涨的时候吹的跟最坚定的皈依狂战士一样，一跌起来就变带路党开始往死里Fud了。

这次大跌我还挺兴奋的，手里的大部分子弹一直备着，就等着像样的回调呢。

大瀑布来吧来吧

[陈小萌 @MengLayer]

用了没两天， @Plasma 卡就不让我用了
一问客服就是三不知 真服了😅
到底啥原因嗷~
@rongplace
@proofofnathan

[TraderS | 缺德道人 @TraderS18]

市场是很诚实的，这一条信息直接把CLUSDT从91干到94，早上开的多单舒服了

[GTx @GT_Chiang]

Wave planning: map dependencies first, then dispatch Wave 1 (independent, max 3 concurrent), Wave 2 (depends on Wave 1). Same-file edits are sequential. Controller verifies—never trust subagent summaries blindly. Extract patterns, gotchas, conventions into hindsight memory.

[GTx @GT_Chiang]

MUST NOT catches AI slop: do NOT add unnecessary logging, do NOT refactor unrelated code, do NOT change variable names for "clarity", do NOT create new abstractions unless requested. These prohibitions stop destructive over-delivery.

[GTx @GT_Chiang]

把最近打磨的一套方法论开源了：hermes-llm-wiki

受 Karpathy 的 LLM Wiki 启发，但重点不是“再做一个 RAG”，而是把它落成一套 Hermes + Obsidian 的 compiled knowledge workflow：

- Inbox = raw intake
- _wiki = compiled layer
- ingest / query / lint 显式运行
- 高价值问答 selective writeback

Repo:
https://t.co/2TJ4XGplwN

[GTx @GT_Chiang]

Open-sourced: Agent Operating Model Kit.

Run agents like a micro AI company:
- token is budget
- optimize value per spend
- no output = not finished
- no verification = not complete
- compound through reusable assets

https://t.co/1r8cK2GdCz

[GTx @GT_Chiang]

@xingpt cc的memory backend有啥解法吗，跨session的memory

[GTx @GT_Chiang]

@taresky 为什么会泄漏聊天内容呢，team不是隔离的吗