这一个半月搬来纽约之后做的app终于发布啦！ 适应了以年计的发布周期之后，跑到创业公司真的感到耳目一新。拜现在的模型所赐，经常上午还在跟小伙伴炒架构设计，下午就进build了。 我们在agent harness和UI设计上都有一些独创的想法。这只是开始，有些粗糙简陋在所难免，更多有趣的功能还在后面！

最近最喜欢的一期播客。化解了我对宇宙的本质是什么的一些执念。podcasts.apple.com/us/podcas… 可以跟之前关于龙树的一期一起消化：podcasts.apple.com/us/podcas…

Form and content. Program and data. Rhythm and melody. Harness and model.

如果iPhone Duo的两个摄像头是20mm和40mm（2×可以获得80mm）就太美了…… 13mm和26mm实在是不太好用。作为超广角和日常焦距都太广了。

一些模型让你感觉比之前更有智慧，比如GPT-4、某一个Opus、Fable。你会希望让模型做自己的thought partner，跟它探讨各种各样的问题，而不是完成任务。这次GPT-6 Astra完全没有给人这样的感觉，只是一个更强的over-fit的工作机器。感觉孩子的通识教育出了问题。

21世纪怪现象：People posting really grateful messages on LinkedIn about being laid off.

It alleviates my latent guilt as well. But I think there’s still an aesthetic angle. There’s something sad about dumping tons of garbage in even a well-run landfill. And there’s something beautiful about reusable bags and water bottles, even if they cost more CO2. For sure, the reusable objects have to be reused. The worst of the worst is dumping reusable objects in the trash. Similarly, in efficient capitalism, buying new goods can cost less in labor and emissions than repairing or reusing old ones. But not trying to prolong the lifetime of objects feels morally off. I’m probably not a pure utilitarian after all…

感觉现在AI训练完全没有教育学和心理学的介入。

Yes, we actually parse the shell commands and do capability analysis using the AST. It doesn’t solve the halting problem but it saves ~80% of token usage.

LM Studio Bionic spotted in the M5 Ultra press release! @lmstudio apple.com/newsroom/2026/08/a…

Observations on Qwen 3.8 27B in @lmstudio Bionic I had some fun doing research work locally with Qwen 3.8 27B. - The “extra high” setting overthinks a lot, but it’s actually useful for better results. Whatever its latest attention mechanism is must have some magic to be able to pick up the right parts reliably. - It’s very fun to see the full thinking trace. That’s a luxury with open models. - The model uses the context windows almost like a mechanical scratchpad. I see it write drafts in its mind, critique itself, and iterate on them. It’s extremely good at repeating large chunks of text. You can also see the behavior in @simonw’s Pelican transcript: tools.simonwillison.net/mark… - I asked it to critique my writing. Surprisingly, the quality is better than much larger models like Opus 5, GPT 5.6-Sol, and Kimi K3. Is it less overtrained or what? This seems to be the first legit local model that can do serious work.