qhc

计算机专业学生｜服务器、网络、AI 与科技工具
记录真实测试、项目进展和踩坑经验

[qhc @qinhaochen1]

https://t.co/YOp9N4Epw1

[qhc @qinhaochen1]

@PennyJoly 没错哈哈哈，还有开放pro订阅

[qhc @qinhaochen1]

Paul Christiano 进 OpenAI 安全委员会这件事，我想继续看下去。

他刚加入 OpenAI Foundation 董事会和 SSC ，同一天又公开说，整个 AI 行业，包括 OpenAI，目前都没有把失控风险降到可接受水平。

OpenAI 过去明确写过，SSC 和董事会可以因为安全问题推迟模型发布。

如果后面真碰到模型发布和安全判断冲突，他能把这些担忧带进实际决策到什么程度？

[qhc @qinhaochen1]

@vu0tran A few years ago I probably would’ve laughed this off as sci-fi. Now OpenAI is calling Astra its first model to hit the Critical cyber threshold. Taking down a whole country in a day still feels like a huge leap, but I can’t really laugh the rest off anymore.

[qhc @qinhaochen1]

@matthias_mrc I think I accidentally outsourced the addictive part 😅 I built a product-manager agent around Codex, so now I just tell it what I want and roughly how it should look, then I lean back and scroll my phone while it keeps going.

[qhc @qinhaochen1]

I’ve been running into this while learning HarmonyOS. I’ll follow an older DevEco/ArkTS tutorial, then try basically the same code in a newer setup and start getting compiler errors. AI usually fixes it pretty quickly, but a lot of the time I’m just using it to get old tutorial code working again. So even with better coding tools, the rest of the development experience can still be pretty messy.

[qhc @qinhaochen1]

@JayHhdmg AI 确实能把一个人能做的项目撑大很多，但人的注意力还是那点。代码越来越多以后，最费劲的反而不是写，而是你得一直确认它到底改了什么。

[qhc @qinhaochen1]

@PandaTalk8 我反而觉得，先别盯着哪个功能最值钱，先看用户到底有什么事已经烦到想马上解决。真碰到这种问题，功能简单一点没关系，能解决就有人愿意付钱。

[qhc @qinhaochen1]

@thdxr If someone genuinely thinks the risk is that extreme, I’d expect it to change what they’re willing to give up. Staying long enough to vest and then sounding the alarm makes the warning a lot harder to take at face value.

[qhc @qinhaochen1]

@npashi I’m more curious about cutile-rs here, especially since NVIDIA itself recommends starting with Tile. I’d like to see how far you can get before you need the extra control SIMT gives you.

[qhc @qinhaochen1]

@poteto Yeah, I think it’s less about English and more about taking a fuzzy idea and breaking it into clear enough pieces that the model can actually do something useful with it.

[qhc @qinhaochen1]

@moshhamedani Yeah, when getting something out gets this fast, actually sitting down and testing the boring stuff is probably the first thing people start skipping.

[qhc @qinhaochen1]

鸿蒙开发踩坑 https://t.co/1kEB3GAf4F

[qhc @qinhaochen1]

@fankaishuoai 把活做快和决定裁谁还是两回事吧。工具省下来的时间，是让大家少加点班还是直接砍人，最后还是公司说了算。

[qhc @qinhaochen1]

这次约 1 万个 Agent 一起啃一个数学问题，我是真想多看几次这种玩法。

88 小时后能跑出一套证明，后面他们还会拿什么问题来试，我挺好奇。

[qhc @qinhaochen1]

@jackzhj 是这样的

[qhc @qinhaochen1]

@elshayib_ I've had posts disappear from my profile before, so seeing all these rejections makes me wonder what X is actually seeing when it reviews an account.

[qhc @qinhaochen1]

https://t.co/XxAhNiRydM

[qhc @qinhaochen1]

DaVinci Resolve Studio 21.1 这次我是真想试一下。

找段长视频丢进去，让Codex 先剪一版，看看能做到什么程度。要是能用个七七八八，后面自己再微调就省事多了。

[qhc @qinhaochen1]

@hilbertspaess Yeah, if every lab assumes the others will keep going anyway, I don’t see who’s supposed to slow down first.