王一 retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/con…
Got the latest Steam Arm64 client up and running on the Nintendo Switch 1. Games run via Proton Arm64 (Fex).
Valve increased the HW requirements for Steam to Armv8.1, so I had to emulate and patch the missing instructions on the Switch, hence the initial lag on Steam bootup.
王一 retweeted
这次真冤枉Argon了,真正拖后腿是反重力!
最近真的被Gemini折磨疯了,Gemini每次发布看似跑分挺高,但是在agent中的表现极其拉跨,最大的槽点就是经常性失忆,跑长任务写着写着就跟脑梗了一样,以下几点尤为突出:
明明没执行过的单元测试,张嘴就敢说全部通过;
开局反复强调的业务红线,写着写着全当耳旁风;
刚解决的依赖冲突,转头又改回去踩一遍。
现在全网都在吐槽Argon额度少性能一般还降智。
今天我实在受不了,挂代理抓包看了一下客户端请求,才发现大家全骂错对象了。
Google宣称 Gemini Argon上下文窗口为1M,但是事情却没那么简单!翻出反重力内置调度器配置那一刻,我整个人都绷不住了。
界面上大字标着 1,048,576 Token的超大上下文,合着全是幌子,
反重力底层的调度器里,清清楚楚写着这几行配置:
{
"strategy": "CHECKPOINT_STRATEGY_SAME_MODEL",
"token_threshold": "50000",
"max_token_ceiling": "256000",
"max_overhead_ratio": "0.15",
"max_user_requests": 10,
"is_sync": true,
"include_last_user_message": true,
"include_conversation_log": false,
"include_running_task_snapshots": true,
"include_subagent_snapshots": true,
"include_artifact_snapshots": true
}
解释一下最重要的几处配置:
token_threshold: 50000:在消耗50K token之前agent处于安全区,自由发挥,上下文历史原汁原味地保留。一旦累计到50K,系统并不会粗暴突然掐死,而是进入待命状态。
is_sync: true:意思是同步等待。系统在等一个合适的时机,也就是逻辑断点,一旦等到,立刻强制执行压缩!
所谓逻辑断点,就是Agent在写代码时告一段落的停顿点,主要有这三种:
完成 task.md 的一个小步骤:比如它刚刚把清单里的 [ ] 修改勾选成了[x];
切换子代理:主Agent把活派给子 Agent,或者子 Agent 把结果汇报给主 Agent 时;
用户对话达到 10 次(max_user_requests: 10):你跟它来回扯皮聊了 10 轮。
也就是说只要 Token过了50K,Agent 一旦停下来触发逻辑断点,系统反手就强制压缩。
max_token_ceiling: 256000:意思当Token消耗到 256K时,系统不管它正在改代码、读日志还是在干嘛,直接强制截断。这是用来兜底、防止服务器被无限死循环撑爆的防线。
然而就像刚才说的,现实是绝大多数会话在50K至70K时就被强制压缩,远早于256K上限。
include_conversation_log: false:它会直接把你前面所有的聊天记录、工具调用、报错日志全部删干净,一条都不留。
搞清楚配置后现在还原一下业务场景:
0 ~ 40K:你提了需求,Agent读了项目结构、3 个相关代码文件,写出了task.md计划。此时一切正常。
40K ~ 55K:agent修改了第一个文件,跑了一次终端测试,产生了一些报错和输出日志。此时token消耗来到50K!
55K: Agent解决完第一个小报错,顺手在task.md上打了个勾 [x],准备开始改第二个文件。
逻辑断点此时被触发!系统发现当前上下文已经消耗55K的token,而且刚好处于完成一个小步骤的停顿点,立刻强制执行压缩,把你前55K窗口里的报错历史、聊天记录直接删除!只保留一个16K的残缺总结。
这就是为什么大家普遍觉得Agron思考挺全面,但是跑长任务跑着跑着,模型就突然跟脑抽了一样失忆的根本技术原因。
然而最难绷的是看他们内部的系统提示词,Google自己的工程师都快被搞崩溃了,在提示词里求爷爷告奶奶地加:“求你分清文件是真改了还是只读了”、“千万别把没测的代码标为通过”。但是依然毫无卵用,该忘还是忘。
基于此我总结了一套3个避坑操作:
1. 重要的规矩约束千万别只在聊天框里说。核心规范、禁动文件,全部老老实实写进项目的 rule 文件或 task.md 里,因为只要有文件它能一直读到。
2. 坚决不要开长会话
别指望一个对话窗口跑完整个大工程。写完一个小功能或模块,直接开新会话,把任务拆开了搞。
3. 永远不要信它嘴里说的测试通过
它说改好了别当真。必须逼它在终端里敲真实测试命令,或者切其他agent进行审查。
用了 Tailscale 好久,只限制了 ACL 相关权限,没注意还有 Tailnet Lock 这玩意 🤡
刚才查了一下,这功能相当于给设备进入 Tailscale 网络加了一层签名验证,就算未来 Tailscale 官方被黑,也能防止黑客往你内网添加恶意设备
如果不开这功能,一旦 Tailscale 官方服务器被黑,并且如果你 ACL 权限没限制好,黑客完全可以利用恶意设备一路打穿你内网····
鉴于目前ai打洞和喝水一样,我果断光速打开了🤡····
王一 retweeted
Game plan is still the same: Touch-Screen MacBook Pro, 14-inch M6 MacBook Pro, new iMac, iPad mini OLED all set for next week.
王一 retweeted
Mi primer día con el Tesla teapaw.org/ssssscca/status/211098…
王一 retweeted
I got this dialed in...
24/7 X Community Manager
w/ @bot 🫡
If you manage a brand account on X, this is literally a cheat code.
Here's what to do with this update...
create 3 bots:
1) 24/7 Community Manager - Scan X daily (or more frequently) for mentions of your brand. Surface tweets worth amplifying, highlight product feedback (or stuff worth flagging internally), share team posts that are trending (or performing abnormally well). I also got mine doing a daily digest, notifying me when there's an urgent alert (like a trending tweet/crisis, monitoring our launches, and doing ongoing sentiment reports.
2) X Content Strategist - Read every single tweet your brand has done. Analyze all of the engagement, replies, bookmarks, etc and recommend content buckets, ideas, memes, etc. based on top performing posts. Also worth analyzing content from the founders/team. Bonus: Have the bot build out a voice/tone skill that could be used for future content sparring.
3) X Analyst - Analyze all of the engagements and replies on your past 30/60/90 days of brand content. Tell me which accounts jump out. Any influential people? Industry leaders? Creators? etc. Then share these with the team - provides good signal.
王一 retweeted
爽,无限Astra来了,开源 dot2api,完全安全不封号
对于dot的额度消耗,ChatGPT官方的说法是:
dot 聊天不会消耗套餐的额度;
那问题就变成了:怎么在 Dot 入口之外用上它? Dot 没有 API,但它能通过 MCP 连接器被唤醒。
所以我写了 dot2api: 把你的 Dot 包成一个 OpenAI 和 Claude 兼容的端点。
适合:定时任务、n8n 这类自动化流程、自建 Telegram 或飞书机器人、让别的 Agent 把一件事委托给 Dot。
开源地址:github.com/Pluviobyte/dot2ap…
王一 retweeted
openai just released Decisions API, Jev-like "system 1" multi-modal model based on GPT6 Luna!
Each egg gets cropped and sent to OpenAI (~270ms latency) and returns P(clean, dirty, cracked).
One image crop consumes ~280 input tokens (~$0.03/1K imgs). @roboflow RF-DETR for detecting eggs
王一 retweeted
「音と動きをビジュアライズ」
Opus 5.5×Blender×Python
■素材
・人物:MiniMax H3
・キービジュアル:Midjourney
・音楽:MiniMax Music 3.0
■Blenderとコードでやらせたこと
・音が大きいほど、絵が立体になって手前に飛び出す
・速く動いた所ほど、粒になって浮く
・打った瞬間に時間が止まり、カメラだけが回り込む
・止まっている間、空間ごと色と照明が切り替わる
■わかったこと
・生成だけでは出ない絵も組み合わせるとできる
・やりすぎるとだめ
・映像の動きにも反応させたら、映像そのものが鳴っているように見えてきた
・飛び出す量は曲の音と、素材の動きに任せた
実験的な映像ではありますが、何かの表現に使えそうだなと思います。AIをより拡張できる方法を探っていきたいです。
音と動きの両方で絵が動くと、なんか面白いですね。
王一 retweeted
Here's a Grok Bot template I use to challenge my ideas
A Devil’s Advocate bot that won’t cheerlead your ideas
Socratic pushback while it helps you shape the offer. Weak assumptions get questioned before you commit
Here's the bot link:
x.ai/bot/Uj8zF0iprBX0AsSB3mU…
1/ How I infiltrated a Chinese organized crime syndicate that has laundered $1B+ across multiple exploits for Lazarus Group.
Posing as a client, I gathered intel that helped action freezes for the Feb 2025 Bybit exploit and attribute illicit activity onchain.
Just published my "Explainer Bot" 🧠
I built it for myself first. It helps me understand things deeply and see things more clearly
I love learning, but some topics are really hard to understand properly. I'd see something on 𝕏, ask Grok for context, and get a quick summary. A summary isn't understanding, though. I wanted to know why it works, what's actually going on underneath, and whether it even makes sense
So I made a bot that does that, with first-principles thinking built in
You bring it a question, a link, an 𝕏 post, a paper, or a document. It reads everything (the full thread, the quoted posts, the sources, the data) and searches the web to check the claims. Then it explains how reality works underneath:
• The physics, meaning what's truly impossible and what's just a habit, a cost, or a regulation
• The mechanism, step by step
• The real numbers, worked out in code and compared to things you already know
• The trade-offs, and the hype separated from what's proven
• How Elon would look at it, plus the strongest counter-argument
• Clear diagrams and real reference images
You can ask it anything, like "Do space data centers actually work?", and it'll take you from the basic physics all the way to a clear answer
The goal isn't a summary. It's understanding that changes how you see the world
Try it here...this is my favorite bot actually
x.ai/bot/_G_I5jXyXTNLten5Mnt…


































