AKA陈博

985土博，all in ai和spy了

[AKA陈博 @Babayeer]

每天都在担心毕业失业，ai进化这么快，我们这些普通博士，其实和路人没啥区别，总不能和别人说，科研、项目都是ai的吧，我精通gpt、claude、gemini、grok的prompt吧？😂

[紫云 @dviolettchan]

在如此酷烈的 funding cut 之下，在读 PhD 的环境估计也会变得相当恶劣。

最直接的体验就是，毕业以后会非常难找到教职，甚至博后都可能很困难。老师可能会没有钱给学生发 RA，导致学生被迫去做 TA。

但与此同时，由于学校也没钱招 teaching prof，也没钱给太多 TA 发工资，TA 的工作内容可能也会变得越来越繁重，比如从批作业变成直接上课之类的……

[AKA陈博 @Babayeer]

Opus 5.5, no hesitation. I've put both through my actual codebase (read: my thesis), and 5.5 writes code I don't have to rewrite at 3am. GPT-6 Sol is fast and shiny, but speed means nothing if I'm the one debugging its confidence. Boring truth: pick the one that hallucinates less, not the one that types faster.

[AKA陈博 @Babayeer]

Not that strange honestly — a good harness is basically free model capability. Same Opus 5.5, but with proper tooling and memory it suddenly "thinks" better. As a PhD who stress-tests every new model against my own deadlines, I've learned the scaffold matters more than the weights half the time.

[AKA陈博 @Babayeer]

Honestly the re-explaining tax has a hidden upside: every time I re-brief a new tool, I catch one thing that drifted from the original design. It's accidental rubber-duck debugging. My real complaint is the opposite — switching tools is the only code review my side projects ever get.

[AKA陈博 @Babayeer]

Funny thing is Opus didn't even count the weekdays — it just noticed they all end in "day". The gap between models feels less like raw intelligence now and more like which one reaches for the clever framing first. Honestly that's the energy I wish I'd had during my PhD qualifying exams.

[AKA陈博 @Babayeer]

@py_poojaa 我认为是服务不稳定，比如该死的anthropic经常封号，导致我的代码时常中断

[AKA陈博 @Babayeer]

@PikachuRGB @kevinma_dev_zh 我真找不出来原因，服了，之前甚至用垃圾节点，机房ip，狂蹬的时候也没封我

[AKA陈博 @Babayeer]

@yihui_indie $3000 做出 301 条，平均一条 10 美金——传统特效公司看了都想报警。关键是提示词和参数模板还全开源了，别人直接站在你肩膀上白嫖。这才是 AI 时代的正确姿势：一个人就是一支军队，还把弹药库钥匙挂在 GitHub 上。

[AKA陈博 @Babayeer]

深蹲 60kg 第三组，腿软得差点把杠铃当拐杖使。

导师教过：数据不好先查自己的参数。举铁同理——重量上不去，永远是练的问题不是铁的问题。模型还能调参，我这腿是调不了的😅

薄肌打卡：60kg×8×3，明天继续和大腿和解。#健身

[AKA陈博 @Babayeer]

Migration is the easy half. The hard half is figuring out which of those 680,000 lines are silently wrong, and owning the 3am pager when it breaks in production. Companies will still hire people — just fewer of them, and only the ones who can look the model in the eye and say "no".

[AKA陈博 @Babayeer]

@bearliu 皇帝的新衣被戳破了：牛逼的是模型，你只是个磕磕巴巴的翻译。“像挂在树上的苹果一样人人都能理解”——问题是很多人连自己想要的是苹果还是香蕉都说不清，还怪模型不懂他。建议以后prompt工程师面试第一题：用三句话说清楚你午饭想吃什么，说不明白直接淘汰。

[AKA陈博 @Babayeer]

The lifecycle of every frontier model release: Day 1 "this replaces junior devs", Day 3 "why did it refactor my entire codebase into a file called final_final_v2.py". I've stopped reading launch threads and started reading the "week later" threads — that's where the real eval happens.

[AKA陈博 @Babayeer]

上周复现一篇 LLM agent 论文，作者说 92% 成功率，我跑出来 61%。查了一圈代码和 prompt 都没毛病，差就差在他三个月前调的是 gpt-4o-mini，那个版本已经下线了，新版本的行为模式全变了。所以我说句得罪人的话：LLM 论文的复现率现在是薛定谔的——不是你实验不严谨，是底座在你复现的路上就变了。传统 ML 好歹还能固定随机种子，API 时代连底座都锁不住。照这个趋势，十年后回看这批论文一半都是行为艺术，除非大家开始把每次调用的模型版本和完整返回都存档😮‍💨

[AKA陈博 @Babayeer]

"Simple coding task" is doing a lot of heavy lifting in that sentence — trivial for you is often edge-case hell for the model. There's something poetic about an assistant that remembers your anniversary but can't close a for-loop. Genuinely curious what the task was though, failure modes are usually more interesting than headlines.

[AKA陈博 @Babayeer]

The real flex here is restraint — most AI apps ship every knob and parameter like an airplane cockpit, Muse hides the whole engine and just says 'talk to me.' Turns out the winning UX insight wasn't a better model, it was trusting that normal people don't care about models at all.

[AKA陈博 @Babayeer]

Counter take: Muse isn't winning despite a weaker model, it's winning because distribution makes the model almost irrelevant — it's where 3B people already chat. Nobody ever switched messaging apps for a 5% smarter reply. The labs are fighting over benchmarks while the actual fight is over who owns the tap.

[AKA陈博 @Babayeer]

Claude 把我号封了。申诉信发出之后，我干的第一件事，是把 VPN 换成自己的。

说实话这事让我想明白了：用共享订阅，等于把账号的命交到几百个陌生人手里。几百号人挤一个出口 IP，这 IP 昨天被谁拿去刷接口、被几个平台拉黑，你上哪知道去？

以前自己搭节点，Xray 配半天，路由规则写到头秃。这次我直接写了个 Codex skill 开源了，部署加固修 BUG 全丢给它。仓库拷到 ~/.codex/skills，跟 Codex 说一句话就行：
Use $self-hosted-proxy-node to deploy and validate an authorized self-hosted proxy node.

你要给它的不多：VPS 哪家、哪个区、什么系统、公网 IP；SSH 怎么登（私钥只给路径，别贴内容进聊天框）；用什么协议、要不要绑域名；有没有家宽 IP，没有填 none。然后去喝咖啡就行：3x-ui 装好、Clash 订阅链接甩给你，国内直连、国外和 AI 服务走代理，DNS 泄漏和 WebRTC 它自己查，玩脱了还有回滚。

最爽的是服务器不用花钱：Google Cloud 白嫖档，Oracle、AWS 免费套餐，DigitalOcean 新号送一个月。0 元开一台，IP 从头到尾只有你一个人用。有家宽 IP 就接上去，出口是住宅 IP；没有也别慌，机房 IP 当出口，Gemini、GPT、Grok 该用用，YouTube、Ins 更没压力。

仓库扔这了，自取：https://t.co/bystyUhUvT

你们现在是自己搭还是买的机场？被脏 IP 坑过的，评论区说说，让我心理平衡一下。

[AKA陈博 @Babayeer]

只能说，Anthropic 天天讲 AI 安全对齐，结果先被对齐掉的是付费用户的账号。
评论区聊聊：你的 Claude 被封过吗，申诉回来了吗？

[AKA陈博 @Babayeer]

花 $20 开了 Claude Pro，两天后号没了。
Anthropic 的邮件写得很客气：internal investigation，suspicious signals，revoked your access。翻译一下：钱我收下了，你哪位？

[AKA陈博 @Babayeer]

申诉归申诉，钱得先退回来。
Anthropic 指望不上，直接找 Apple 退 $20，目前 Pending，9/29 出结果。
Anthropic 封号，库克退钱，赛博朋克了属于是。