Arthur Lin ретвитнул
Arthur Lin
288 posts
中文
突然意识到前两天刚加入Anthropic的 @karpathy 也没有资格访问他们最强的模型Fable 5😅
因为他并非一名美国公民，他出生于斯洛伐克（前捷克斯洛伐克），15岁时移民至加拿大，目前是加拿大公民，同时也是美国EB-1杰出人才绿卡的持有者。
这太讽刺了哈哈哈🤣
Anthropic@AnthropicAI
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…
中文
Arthur Lin ретвитнул
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
English
求求大家不要办Founder's breakfast了，因为AI founder没有吃早餐的习惯（根本起不来
也许可以考虑Founder's 下午茶：）
中文
Warning to Claude Code account switchers: multi-account on the desktop app is NOT VIABLE.
You have to hard log out and you will lose all your sessions.
CLI doesn't care at all, you can just /login in a different tab and all traffic gets routed correctly on existing runs.
English
@amorriscode @theo Is there a way that I can resume a session I start from CLI to the desktop App?
Currently it doesn't appear to be the case
English
got a challenge for you. without checking slack...
how many times do you think I need to press ↓ to select #team-hype?
English
他講的應該是指這個 pi.dev
他的核心理念是，現在有很多家像是 Claude Code 或是 Codex 這些都是其他家的 Agent harness
但是 pi 不一樣，他是你自己客製化的 agent harness
就挺像 neovim 的，所有東西都可以客製化
所以如果你明白你要怎麼客製化你的 AI agent 而非只會用出廠 Claude Code 或是 Codex 給你的設定的話
那你要做什麼自己的 AI agent 應該也蠻容易的
pi 出廠只給一些蠻基本的設定 pretty minimal
算是一個 kickstart 一個 foundation
可以去玩玩看，依照自己的理解去做你的 agent harness
你可以登入你的 codex OAuth 就可以用 GPT 模型了
可以先去載來然後問 pi 他的核心理念是什麼，然後要怎麼客製化他（他的系統提示詞裡面有寫到，所以 pi 可以去查文件，然後告訴你他可以怎麼幫你打造他自己）
整體算是比較進階，能夠做到怎樣就看你對於 AI agent 的理解了
希望對你有幫助～
中文
@kushbhuwalka @steipete Most of the loopers won't understand this. But it makes a lot of sense.
English
I’m pretty AI pilled. This loop stuff is slop. I respect @steipete for his innovation - but openclaw is a bloated unstable pile of garbage because of stuff like this.
I’m all for loops of crons and webhooks where an AI agent wakes up and performs some task like cleanup, or updates the docs or triages errors. I think these are great for standard well defined tasks with a fairly deterministic route (a.k.a workflows).
I think what these guys are talking about now is jumping the gun. The models need to be guided, and you want to atleast skim their output so you don’t end up with slop. Humans are far better planners and architects than models.
You absolutely shouldn’t delegate away prompting and reviews in my opinion. this encourages the creation of crappy buggy unsafe software that actually hurts adoption.
English
@tanzhengmc97 我們用的 ai 是一樣的 ai 嗎 🤔
我覺得 Claude Code 或是 Codex 那一種已經算是蠻厲害的
已經有些味道出來了
還是這還算是小東西🤔
中文
@bcherny Could you fix all the render bugs instead of keep shipping broke products? 🤔
English
Just landed nested subagent support in Claude Code
Starting to experiment more with agents kicking off agents as a way to better manage context. Capped at depth=5 to start, going out in today’s release.
Lmk what you think!
English
English
Poke's system prompt finally leaked.
I once spent 30 minutes bargaining with it over price. it wouldn't budge, roasted me for lowballing, and i genuinely thought a person was messing with me.
After reading that prompt, every human thing it did was just a line in there. the roasting. the stubbornness. the lowercase.
so the personality was never really Poke's. drop that prompt into almost any agent and it'll feel just as alive.
That's the valuable part. not the bot. the recipe.
Burrito@BurritoCade
@0xUltraInstinct now they are main agent (as close as possible): gist.github.com/burritosoftwar… email processing subagent: gist.github.com/burritosoftwar… older ones (before apple messages): github.com/x1xhlol/system…
English
can a general agent like Claude Code actually reach into iOS apps and do what Siri does?
apple built App Intents as the action layer... but is it exposed as an API anyone can call, or only through apple's own Siri?
English
Agent Harness 的五个陷阱
1. Self-evaluation is a trap. Use an adversarial evaluator.
自我评估是陷阱。要用对抗式评测器。
很多团队都会这样做：
Agent生成答案
↓
Agent自己打分
↓
Agent说：95分
结果上线后翻车。
原因很简单：
模型和自己存在强烈共识偏见（agreement bias）。
2. Compaction doesn't cure coherence drift. Structured handoffs do.
Context 压缩解决不了一致性漂移，结构化交接才能解决。
这是很多 Agent 框架最大的误区。
长任务：
10k token
→ 50k token
→ 200k token
上下文越来越长。
很多人以为：
总结一下
压缩一下
继续跑
问题就解决了。
实际上：
压缩 ≠ 保真
压缩过程本身就在丢失信息。
最终出现：
目标漂移
约束丢失
需求变形
例如 Codex：
最开始：
重构支付模块
20轮以后：
开始修改登录逻辑已经跑偏。
真正有效的是：
Structured Handoff
每次交接固定格式：
Goal:
Current State:
Completed:
Pending:
Constraints:
Known Risks:
Next Action:
Anthropic 的 Claude Code、
OpenAI Codex Cloud、
Cursor Background Agent都在往这个方向演化。
3. Make subjective quality gradable with rubrics the model can apply.
把主观质量变成模型可以打分的评分标准。
很多需求本质上很模糊：
文章写得好不好？
UI漂不漂亮？
代码优不优雅？
模型根本不知道什么叫：
好，漂亮，优雅
所以必须变成 精确的打分表。
4. Read the traces. They're your primary debugging loop.
Trace 才是 Agent 的调试入口。
很多人调 Agent
输入
↓
输出
只看结果。
但 Agent 最大的问题都发生在中间。
例如：
Planner
↓
Search
↓
RAG
↓
Tool
↓
Reasoning
↓
Action
某一步已经歪了。
因此Agent 的日志比结果更重要。
例如 OpenAI Agent SDK：
Trace/Span/Event
LangSmith：Execution Trace
Langfuse：Agent Trace
Arize Phoenix：LLM Observability
本质都是可观测性
Agent 开发的黄金法则：
80%时间看 Trace
20%时间改 Prompt
很多团队正好反过来。
5. Delete scaffolding when the model catches up. The frontier moves.
当模型能力提升后，要敢于删除脚手架。
这是很多 Agent 产品的死亡陷阱。
2024年的标准做法：
Prompt
+
Planner
+
Reflection
+
Self Critic
+
Chain of Thought
+
Tree Search
2026年：
GPT-5、Claude Sonnet 5 可能已经内建这些能力。
结果团队还保留：
10层Agent，20层Workflow
系统越来越复杂。延迟越来越高。
成本越来越贵。效果却没提升。
典型案例：
早期很多 RAG：
Query Rewrite
↓
Multi Query
↓
Rerank
↓
Summary
↓
Answer
现在很多模型直接：
Search
↓
Answer
已经够好了。
所以Harness 不是越复杂越好。
优秀的 Agent 工程师有一个习惯：
每次模型升级：
增加一层能力
↓
删除两层脚手架
不断做减法。
所以，Harness 是一个随着模型进化的动态更新的过程。
中文
@peduarte Why do you think tramming up with Google is better for apple though?
English
i like it that apple teamed up with google and not openai/anthropic
English
I don't understand why switching from iphone to pixel the eh swiping or scrolling physic feels awful
Either swipe little(kinda sticky) or over scroll
Doesn't seem to be finger detection issue it's more about the physic
Just feel weird someone help because pixel is a great phone minus the scrolling physic 🥹
English