szl

Finding Calm in the Chaos, Together.

[szl @peterlszp]

Hello, Friday! #FridayFeelings

[szl @peterlszp]

@testingcatalog Pursuing a more visually appealing slide, or a slide that maximizes the effective transmission of information?

[szl @peterlszp]

@nicekate8888 期待 deepseek-v4-pro

[szl @peterlszp]

@nicekate8888 deepseek-v4-flash 超出意料的强。速度快，Agent 理解能力准，tools 调用不飘。手感似乎好过 glm-5.2。应该可以用来替换 composer-2.5。

做了个小实验，Review gpt-5.6-sol-xhigh 完成任务，竟然还能挖出4个缺陷（证伪1个，证实部分属实3个）。

[szl @peterlszp]

@nicekate8888 更新：分析能力较强，编码能力较 glm-5.2 弱。可以用来当项目经理。

[szl @peterlszp]

@thsottiaux Help ChatGPT Work <= Switch data => ChatGPT

Work for world
Codex for code
Still there are gaps between them

[szl @peterlszp]

What sets ChatGPT Work apart is its ability to convert messy, fragmented data into actionable output through structured multi-step execution.

I use it as a Commercial Intelligence Engine:
• Scans daily market signals, product launches & updates
• Distills noise into verified facts
• Preserves cross-day context to spot emerging trends
• Outputs a decision-ready brief before my day starts

It eliminates cognitive load, maintains operational continuity, and ensures I only act on high-signal insights. A true force multiplier.

[szl @peterlszp]

gpt-5.6-sol medium/high Human, have a joking time
gpt-5.6-ultra-high Human, leave me alone. Hey, job done.

[szl @peterlszp]

What happened to Codex gpt-5.5????
Unpredictable, incomprehensible, completely useless
@OpenAI @OpenAIDevs

[szl @peterlszp]

🚨 If you run LiteLLM, check this right now.

In docker-compose, "4000:4000" binds to 0.0.0.0 (publicly exposed). And if you don't set LITELLM_MASTER_KEY,
LiteLLM fails silently — no error, no warning — and becomes an open proxy with zero auth. Anyone who scans your
IP:4000 can use your key to call GPT-5 on your dime.

I just found one of my VPS instances had been abused for 11 days — tens of thousands of calls, peaking at 15k/day —
all from unknown overseas IPs burning my key. Not a key leak. A naked port.

Check yourself:
ss -tlnp | grep 4000
If you see 0.0.0.0, you're exposed.

Fix is one line:
"4000:4000" → "127.0.0.1:4000:4000"

Go check your bill and usage graph. An unexplained climb is the signal.

#LiteLLM #Codex

[szl @peterlszp]

@beyang Sourcegraph 是最早用智谱做自动补全功能的哦
印象里是这样的
也就2年，怎么好像很久似的

[szl @peterlszp]

After a night of '/Goal' battles, glm-5.2 won once again.

glm-5.2 vs gpt-5.5

[szl @peterlszp]

@beyang Why can’t both be smart?

[szl @peterlszp]

@nytchinese 帝国的崩溃从知识分子开始了

[szl @peterlszp]

@mdfujita 我第一次到苏州的时候，才发现喜欢的诧寂风其实是宋代美学。

[szl @peterlszp]

@Oscarrjavier95 看到老牌海盗国家指责中国海盗行为，实在忍不住哈哈哈

[szl @peterlszp]

@Gorden_Sun 拖够5秒，就有钱收了

[szl @peterlszp]

@keshiAIart @akuai33 我最好奇的就是这个节奏感是怎么来的

[szl @peterlszp]

@bidcapbro @jumperz Now I can say better than gpt 5.5 and opus 4.8.

https://t.co/dngLjTfODM

[szl @peterlszp]

Now I can say GLM-5.2 better than GPT-5.5, Opus-4.8.

Debugging a multi-Agent workflow node that relies on Trace-Reflection. GPT-5.5 and Opus-4.8 each went through 20 turns and still failed to solve the task.

Switched to GLM-5.2.

Before tner turn, It immediately spotted a field-loss issue in the interaction flow — cool, glm-5.2 got something smarter, a defect that none of the other models had detected.

More importantly, it solved the entire task flawlessly in just 8 turns.

After reviewing the runs, the biggest difference appears to be its understanding of Trace data. GLM-5.2 produced much more effective Reflections, allowing each iteration to build on the previous one with far higher quality.

That’s seriously impressive. Damn, this thing is strong. 🔥

@Zai_org Thank you!

[szl @peterlszp]

Now I can say GLM-5.2 better than GPT-5.5, Opus-4.8.

Debugging a multi-Agent workflow node that relies on Trace-Reflection. GPT-5.5 and Opus-4.8 each went through 20 turns and still failed to solve the task.

Switched to GLM-5.2.

Before tner turn, It immediately spotted a field-loss issue in the interaction flow — cool, glm-5.2 got something smarter, a defect that none of the other models had detected.

More importantly, it solved the entire task flawlessly in just 8 turns.

After reviewing the runs, the biggest difference appears to be its understanding of Trace data. GLM-5.2 produced much more effective Reflections, allowing each iteration to build on the previous one with far higher quality.

That’s seriously impressive. Damn, this thing is strong. 🔥

@Zai_org Thank you!

[Z.ai @Zai_org]

Intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere.

GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans.
https://t.co/AedZACyzej

As our new flagship model, GLM-5.2 delivers powerful coding capabilities, usable 1M-context support, and continued strengths in long-horizon tasks.

API and Chatbot services will launch next week. The model will also be officially open-sourced next week under the MIT License.

The future of AI is open, and it belongs to the people.

[szl @peterlszp]

Now I can say GLM-5.2 better than GPT-5.5, Opus-4.8.

Debugging a multi-Agent workflow node that relies on Trace-Reflection. GPT-5.5 and Opus-4.8 each went through 20 turns and still failed to solve the task.

Switched to GLM-5.2.

Before tner turn, It immediately spotted a field-loss issue in the interaction flow — cool, glm-5.2 got something smarter, a defect that none of the other models had detected.

More importantly, it solved the entire task flawlessly in just 8 turns.

After reviewing the runs, the biggest difference appears to be its understanding of Trace data. GLM-5.2 produced much more effective Reflections, allowing each iteration to build on the previous one with far higher quality.

That’s seriously impressive. Damn, this thing is strong. 🔥

@Zai_org Thank you!

[Z.ai @Zai_org]

Intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere.

GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans.
https://t.co/AedZACyzej

As our new flagship model, GLM-5.2 delivers powerful coding capabilities, usable 1M-context support, and continued strengths in long-horizon tasks.

API and Chatbot services will launch next week. The model will also be officially open-sourced next week under the MIT License.

The future of AI is open, and it belongs to the people.