Zane Distilled

把硅谷前沿 AI / Agent 研究，蒸馏成创业者能执行的产品判断、工作流与商业机会｜Agent Builder｜公开分享案例、工具和踩坑记录

[Zane Distilled @zanedistilled]

@luoyonghao 最反感的就是欠钱不还的人，借了就要还！

[Zane Distilled @zanedistilled]

@Khazix0918 有些人回复是嫉妒、有些人回复是实在，当你的评论区只剩下所谓的“情商高”虚情假意的点赞，你可能会再成就感中，但是实际上做了无用功，为什么AI让你开心，因为你知道AI不像人。人的假方方面面。所以你既然发了你想发的就行了，发出来的动机就不是索取，有需要的自然就用了。

[Zane Distilled @zanedistilled]

@q80306304 mp怎么打开？

[Zane Distilled @zanedistilled]

我已经停止充值gpt，全力使用workbuddy的kimi-k3，希望腾讯不要犯老毛病，好好的产品搞一堆sx的成长计划，一堆无用的垃圾把主能力给淹没,别最终搞成qq空间那种吊样，我现在已经很厌恶那个成长空间了。

[Zane Distilled @zanedistilled]

请大家谨慎吹嘘大模型的价值，干得好和能干是两回事，大模型干的好的继续保持，能干的但浪费了更多时间、成本的以及只有噱头，没有实际产出的，请不要在吹捧了，���沫都��大家自己说出来的，说着说着自己都信了。

[数字生命卡兹克 @Khazix0918]

快被AI逼疯了...
已经精神失常到胡言乱语了

[Zane Distilled @zanedistilled]

@Khazix0918 5.6吧？出了当天我就发文说这个sx更弱智了，���现在用的都是5.5

[Zane Distilled @zanedistilled]

来吧，开始测试：
CPU本地大模型颠覆性方案。GLM-5.2 Colibri int4是专为普通电脑打造的轻量级AI推理引擎。它仅需25GB内存，无需GPU，即可在本地运行7440亿参数的超大模型。项目仅2400行纯C代码，零依赖。它利用MoE架构按需从硬盘流式加载专家模块，将内存占用从1.5TB压缩至25GB以内。虽推理较慢，但为离线AI与数据隐私提供了颠覆性方案。

[Zane Distilled @zanedistilled]

AI时代，企业C-Level阵营或许将迎来一个新角色：CCO（Chief Compute Officer，首席算力官）。

如果说CIO负责信息化、CTO负责技术创新，那么CCO关注的��是企业最重要的新型生产力——算力。
从GPU资源规划、算力成本优化，到AI基础设施建设、模型训练与推理资源调度，再到绿色数据中心和算力投资决策，算力正在成为影响企业竞争力的关键资产，而不仅仅是IT资源。
未来，企业竞争不仅是人才、数据和算法的竞争，更是算力运营能力的竞争。谁能把算力转化为业务价值，谁就更有可能在AI时代建立长期优势。
也许，CCO不会成为所有企业的标配，但对于AI原生企业和大规模应用AI的组织来说，它很可能会成为下一个值得关注的C-Level角色。

[Zane Distilled @zanedistilled]

AI on X has become an endless parade of flashy demos. Most look cool, but very few are genuinely useful. The novelty is wearing off, and people are getting desensitized. The next phase isn't about making AI look smarter—it's about making it more useful. If we keep optimizing for virality over utility, we'll erode the perceived value of AI itself.

[Zane Distilled @zanedistilled]

“GPT-5.6 is terrible, not even as good as 5.5. It’s especially retarded and stupid in my IoT development.”

[Zane Distilled @zanedistilled]

The World Cup has ended. The US won.

[Fox News @FoxNews]

BREAKING: President Trump reveals the details of his conversation with FIFA President Gianni Infantino over the controversial red card call on star American player Folarin Balogun.

"All I did was ask for a review because I didn't think it was a foul. And, you know, again, I'm good at this stuff. I didn't think it was a foul. I thought it was two great athletes that crashed into each other and got entangled."

"I think they made a really brilliant decision. I think the referee's call was horrible and nobody talks about that. They talk about the red card like it's fine, nobody talks the referee's decision to red card."

[Zane Distilled @zanedistilled]

The most practical and valuable application of AI today. No debate.hhhh！

[UBTECH Robotics @UBTECHRobotics]

Get an up-close look at the UWORLD ultra-bionic humanoid robot and discover its remarkable detail and lifelike presence.

#UBTECH #UWORLD #U1Pro #UltraBionic #HumanoidRobot #ConsumerRobotics #Robotics #AI

[Zane Distilled @zanedistilled]

This is a psychological test. Which room would you choose?

[Zane Distilled @zanedistilled]

GPT-5.5 keeps cutting corners. When asked to scrape or fetch a webpage, it just gets stuck, then uses old data to fake an answer instead of verifying the latest information. Is there any good way to fix this?

[Zane Distilled @zanedistilled]

The Last Human-Written Paper！

[Zane Distilled @zanedistilled]

ChatGPT is friendly to Chinese users.

[Jack孔@Nano Labs(NA)🇭🇰 @JackKongNano]

中国考生还是挺厉害的

[Zane Distilled @zanedistilled]

My 14Plus is ready hhh

[Evan @StockMKTNewz]

Here is a first look demo of Apple’s new Siri AI

[Zane Distilled @zanedistilled]

Local small models won’t replace cloud LLMs; their real opportunity is becoming OS-level agents that understand intent, call native tools, and complete tasks directly on personal devices.
Mac-1 captures the current shift: model capability is becoming commoditized, while the next battleground is the agent application layer, native tool orchestration, and real workflow execution.

[CJ Zafir @cjzafir]

Here's a teaser of our Mac-1 model.

> 6.6B model
> runs locally (on any Mac)
> requires 7GB RAM (12GB ideal)
> can use 487 MacOS native tools
> perform multi-tool chained tasks
> reasoning: ON
> output: ~65 tok/s

We built a robust application layer around the model to make UI/UX MacOS native. The "model-focused" SaaS era is here.

Stay tuned for more.

[Zane Distilled @zanedistilled]

Why I Gave Up on Codex PlusPlus？
After spending another two days struggling with Codex PlusPlus, I’ve decided to give up on it.
My advice to most people who don’t have an IT background is: don’t use it.
On one hand, Codex updates so quickly that compatibility issues keep popping up. On the other hand, ask yourself: why use Codex with your own API in the first place? What’s the real value?
For most users, professionalism and stability are the core requirements.
Don’t be like me—constantly jumping from one agent to another, repeatedly trying to build things from 0 to 1, while never truly thinking through the actual project itself.
That’s not where AI delivers its greatest value.
Sometimes, the right move is to let go of what isn’t working, stop chasing every new AI tool, and take the time to think deeply about your project.
Don’t let AI keep you running in circles and standing still.