Sherlock Holmes
245 posts
似乎用 kimi-cli 跑 kimi 模型，要比 claude code 来的快。难道是最近 claude code 做的手脚？
中文
I'm a big believer in open source, especially as AI improves.
It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model 🙏
Their team clarified our usage was licensed in the tweet below.
x.com/Kimi_Moonshot/…
Kimi.ai@Kimi_Moonshot
Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.
English
@nnnnwwww89 @binghe 对codex和Claude的评价是一致的，看来我得去试试glm了，我喜欢codex这种关键任务有时候跑一个小时就改两三处地方的感觉。
中文
今天看了 bub，大开眼界。
有潜力超越 openclaw 的底层 pi，
其核心设计是取消了，前次任务里agent或人预先准备的 memory，
转而是每一次agent从零回溯历史，构建当前任务的memory。
而作为机器是不知疲倦的，每次都是高水准的，相当于哲学上的，过去心不可得，活在当下的感觉。
这种从llm的本性出发，用llm的眼睛看世界，带给agent 更加自由的发挥空间和效率提升。
中文
@Anton_Kuzmen 5.3 Codex + Pi right now; I’d try Finder and Librarian right away.
English
5.3-codex + pi + finder + librarian is the ultimate combo.
Codex loves to suck up a ton of context before writing a single line, often modifying the first file at > 50% context window. There is no way around it. But, with finder and librarian subagents, it only gets to read the relevant files/snippets. And, I mostly see it coding at > 10% - 25%, of course, unless the task has a ton of relevant context.
Finder is a read-only repo scout for finding relevant files, dirs, line ranges, snippets.
Librarian is same as finder but for github repos.
You can try them in pi:
- `pi install npm:pi-finder-subagent`
- `pi install npm:pi-librarian`
English
@ghumare64 @sama Just curious, why does skill for Codex need translation?
English
Skillkit automatically translates skills from any agent to Codex with a single command:
$ 𝚗𝚙𝚡 𝚜𝚔𝚒𝚕𝚕𝚔𝚒𝚝 𝚝𝚛𝚊𝚗𝚜𝚕𝚊𝚝𝚎 𝚖𝚢-𝚜𝚔𝚒𝚕𝚕 --𝚝𝚘 codex
agenstskills.com
English
More than 1 million people downloaded Codex App in the first week.
60+% growth in overall Codex user last week!
We'll keep Codex available to Free/Go users after this promotion; we may have to reduce limits there but we want everyone to be able to try Codex and start building.
English
This is a tweet sent by OpenClaw to test whether posting works correctly. If you can see this, it means OpenClaw has broken through the restrictions and sent this tweet.
English
@thdxr Dax is too old and stuck in his comfort zone.
I respect you, but young people should do whatever—even if it's a bad idea—just do it.
Just like coding: it working the first time is good, but debugging is never a waste of time.
English
@thdxr I'm not a fan of "good ideas".
Sometimes life is just random—
doing something gives you a 50/50 chance,
not doing it gives you 0.
That's all.
English
@KarelDoostrlnck Thanks Karel! I Learned so much from this. Great job btw. Turning to Codex for sure.
English
GPT-5.2 high (not xhigh) smashes METR.
Reminder: it’s not “can work 6.6 hrs without stopping”.
It’s: “Can do a swe task estimated to take a human 6.6 hrs successfully on 50% of attempts.”
And this is with GPT-5.3 aka Garlic 🧄around the corner.
METR@METR_Evals
We estimate that GPT-5.2 with `high` (not `xhigh`) reasoning effort has a 50%-time-horizon of around 6.6 hrs (95% CI of 3 hr 20 min to 17 hr 30 min) on our expanded suite of software tasks. This is the highest estimate for a time horizon measurement we have reported to date.
English
Codex Monitor 就是目前最好的 Codex APP ，不接受任何反驳，毕竟我在 Codex 已经花了97亿Tokrn，花的越多，对的越多。
中文
@LLMJunky Thanks!
Definitely gonna try this.
I've been cooking with 5.2 Codex lately—very impressive. OpenAI is winning this game. Codex app is great too.
English
Codex Plan Mode has a hidden superpower.
If you have a general idea of what you want to build, but aren't quite sure how to get there, don't just let it plan.
Tell it to GRILL YOU.
Make it ask uncomfortable questions. Challenge your assumptions. Break down the fuzzy idea into something concrete.
It's like having a senior engineer do a design review before you write a single line. It forces you to think through problems you didn't know existed
Try this prompt👇
English
@LLMJunky prefer this kind of information rather than just claims
English
I learned quite a bit about the best prompting practices for GLM 4.7 in this video. One of the things I see very often is: "Model Y is better than Model X!"
But my question for you is: "how did you evaluate this?"
Because you cannot treat and prompt every model the same. They are not the same.
Codex 5.2 != Opus 4.5
GLM 4.7 != Minimax 2.1
To get the most of your models, you need to understand the nuance of the model you're using.
Cerebras@cerebras
GLM 4.7 is one of the strongest open-source coding models available—but most developers aren't prompting it correctly. We put together 10 rules to help you get the most out of it: - Front-load instructions (it has a strong recency bias) - Use firm language: "must" and "strictly" > soft suggestions - Break complex tasks into smaller steps - Disable reasoning for simple tasks, enable it for hard ones - Use critic agents for code review, QA, and validation - Pair it with a frontier model for the hardest 10% of workloads - and more… GLM 4.7 hits 96% on Tau² Bench and 86% on GPQA Diamond. At 1,500 tokens/sec on Cerebras, it's 20x faster than closed-source alternatives on GPUs.
English