Ben77

2.1K posts

Ben77 banner
Ben77

Ben77

@blapta

Designing incentives, not illusions.

UTXO Katılım Nisan 2017
1K Takip Edilen3.8K Takipçiler
Sabitlenmiş Tweet
Ben77
Ben77@blapta·
如果康德是用 Lisp 写《纯理性批判》,也许读起来会让人好受得多。 今天的哲学家,很多并不在哲学系里,而是隐藏在程序员、物理学家和 AI 研究者之中。 经历过真正的工程和研究训练的人都很容易学习哲学,因为这些工作和环境反馈其实是在不断训练自己,用更精准的语言来描述世界。 最近看过很多创新的案例都让我不断相信,把技术仅仅当作工具的人,潜意识里其实也是把自己当工具。 很难相信这样的态度能做出突破性的创新,而非落入一种盲目的跟风。
中文
7
2
61
7.6K
Ben77
Ben77@blapta·
@MiTiBennett Most organizations, especially companies, are machines for turning smart people into mediocre ones.
English
0
0
0
10
Michael Timothy Bennett
Michael Timothy Bennett@MiTiBennett·
I joined a company where 400 people were doing the work of 20… made the mistake of pointing this out. Got fired, company went bankrupt a year later. It isn’t that all the people are incompetent, but a few bad actors can cripple everyone else. I have discovered it is usually easier to work alone if you really want to get something specific done.
Roy@usr_bin_roygbiv

Every once in a while you work at a company that only hires people who know what they're doing, and suddenly its 20 people doing the same work as 400 somewhere else. There's zero meetings, everyone talks once a week on slack, and you go huh, how much garbage is there actually.

English
1
0
17
533
Ben77 retweetledi
Ryo Lu
Ryo Lu@ryolu_·
when the dream becomes the job i used to think the dream was to make my hobby my work. to spend my life close to the things i loved: computers, tools, interfaces, systems, tiny details, the feeling of making something from nothing. the things i used to do at night, after school, after work, in the margins of life, somehow became the center of my life. and it is a dream. i don’t want to pretend otherwise. there is a very specific happiness in getting paid to care. in meeting other people who care. in building a language around your obsession. in seeing something that lived only in your head become real enough for other people to touch. but there is also a quiet pain in it. because the thing that once belonged only to you no longer belongs only to you. it becomes attached to deadlines, teams, customers, strategy, money, reputation, momentum... curiosity becomes roadmap. taste becomes decision-making. play becomes output. the thing you once used to escape the world becomes the thing through which the world reaches you. and now there is another strange layer: the machines are getting good at the work too. not just the boring parts. the parts that used to feel close to the heart. writing, coding, designing, reasoning, making taste-like decisions, turning vague ideas into something real. the things we once used as proof that we had something special inside us are becoming things a model can approximate, remix, and sometimes do faster. that can feel scary in a very specific way. not only because of jobs or competition, but because it touches the myth you built around yourself. if the thing i became good at can be done by a machine, then what part of it was really me? i think that question is painful because it arrives at the same place. the original love. before the title, before the company, before the taste became a role, before the work became useful to other people, there was a small fire inside. curiosity. obsession. the feeling that the world became more alive when you understood a little more of it, or made a tiny piece of it bend to your will. AI can make the output faster. it can even raise the floor of the craft. but it cannot want on your behalf. it cannot decide what is worth loving. it cannot protect the strange private thread that made you care in the first place. maybe that is what becomes more important now. not whether you can still produce the artifact by hand. not whether you can outperform the machine at every task. that game is already changing. the real question is whether you still have a relationship with the source. the part with no audience. no roadmap. no deadline. the part that can wander, collect strange references, make useless things. follow beauty, be wrong, be slow, be you, be true. because that part is not separate from the work. it is the source of the work. if you lose it, you can still operate. you can still manage the machine. you can still prompt, review, decide, ship, and keep things moving. but the work becomes thinner. safer. more explainable. less alive. so maybe the real discipline is not to work harder, rest more, or care less. maybe it is to keep returning to the place where the love began – before it became useful, before it became legible, before anyone needed anything from it, before even the machine could mirror it back to you. not to escape the work. to make sure the fire inside is still alive.
English
19
14
234
6.5K
Ben77
Ben77@blapta·
@rwayne 但是你会发现很多人似乎都不需要人生的意义,这让我感到非常可怕。这种人也没有任何行动出发的原点,随时可以为利益出卖一切。但是今天很多人会认为这叫聪明。
中文
0
0
0
71
Roland.W
Roland.W@rwayne·
很多大厂的牛马觉得没有行动力,是我们不相信我们做的这件事情有意义 盲信的人其实挺幸福的,因为他不去质疑自己做的这件事情有何意义,他就会有非常强的行动力,很快乐地去把这件事情完成 当我发现我以前相信的那套东西崩塌了,你已经考到了一个重点学校,然后突然间就一片虚无 真正的主体性觉醒,是知道我没有人可以帮我,我只能自己去面对,找到我的那条路 哪怕这个目标是别人植入的,只要我认定了之后,我的行动力和主体性就会激发出来,他是正交的两件事
中文
3
0
5
1.5K
Ben77
Ben77@blapta·
@aiandcloud API只会让本就不富裕的家庭雪上加霜
中文
0
0
1
55
Ben77
Ben77@blapta·
订阅了Codex pro,你就会收获无尽的精神内耗:每天都活在reset的恐惧之中。 每天都在要不要开 Sol ultra fast猛干和是否自行reset中纠结。 特别是对于我们这种Asian bro,你永远不知道,白天自己 reset 之后,半夜官方会不会突然也 reset 一波。 只怕是官方夜里悄无声息给 reset 了,结果咱在梦里还在省着用。
Tibo@thsottiaux

Morning. The last 48 hours of Codex and ChatGPT Work have been intense! Three important updates: - Temporarily removing the 5 hour usage limit restriction for all Plus, Business and Pro plans - Rolling out changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage being used so that it can take you further. Exact impact to be quantified and shared - We hit 6M active users, and are landing a usage reset in the next hour Go do things

中文
52
17
436
45K
Ben77
Ben77@blapta·
ChatGPT Live 对话的效果让人惊艳,几乎可以做到和真人对话的效果了,这几天我每天都会花半个小时用它来练习英文。 还记得一年多前和老白 @Wuhuoqiu 闲聊的时候,他推荐了几个AI初创公司类似的产品给我。以今天的标准来看,那些公司当时已经具备今天ChatGPT Live的八成的效果了。 虽然不知道这些AI初创公司今天活得怎么样,但是似乎又是一个大模型厂商碾平垂直领域的故事。 按照业务来划分垂直赛道,似乎在这个大模型时代暂时失效了。看过几家Ai垂类公司,似乎都是在做另类中转站的生意,背后的逻辑都在于token补贴和工具支持。 从市场的结构来看,当大部分做互联网产品的公司都不再有私有信息,那么商业壁垒从何而来呢。 也许我们需要找到一种新的方式来区分和生产私有信息,重新塑造市场。
中文
24
0
8
762
Alex Fang
Alex Fang@ms_afang·
@Khazix0918 差不多3周前我第一次跑Grok build没有经验,直接在system32目录下启动了,当时只是想随便看看,当晚没有把进程关掉,于是精彩的来了第二天我的C盘爆了,原因就是Grok 把 C:\WINDOWS\system32 当成了项目/代码库目录,然后开始做 repo state / trace 上传准备,把大量文件复制进本地 upload_queue。
中文
1
1
1
601
JasonQQQQQ
JasonQQQQQ@weiqinghua2015·
@blapta grok拿你的数据啊,甚至.env啊,小心。
中文
1
0
1
43
Ben77
Ben77@blapta·
Grok4.5表现超出预期,有X Premium的朋友都可以试试。 这两天在使用grok build cli,发现干活儿逻辑很清晰,执行迅速同时还自我迭代能力,比Cursor提供的composer 2.5要强的。 感觉拥有X的Premium用户都可以在Grok的Cli登录然后试试,目前看起来没有五小时限制,只有按周刷新的limit。
中文
13
0
22
2.7K
Ben77
Ben77@blapta·
@nake13 感觉是从我做的东西持续撞上fable5 安全过滤器开始,疯狂给我降级到opus,然后没两天就给我ban了。
中文
0
0
1
54
Ben77
Ben77@blapta·
@capyminer 目前看起来premium+的额度也不是很多,除非单开heavy,否则不能作为主力开发工具。
中文
0
0
1
174
capyminer
capyminer@capyminer·
@blapta 确实很快 但是半天就把一周的额度用完了 这么看来完全不便宜啊 别的等5小时能继续用 这个要等一周??????
中文
1
0
1
188
Ben77
Ben77@blapta·
目前看起来最大的缺点,在于这个服务还不够稳定,可能会干活儿一半出现403错误,这种服务端问题应该可以很快得到解决。
中文
0
0
2
426
Ben77
Ben77@blapta·
@elonmusk @elonmusk Grok build is unusable for serious work. After login it works for a bit, then starts returning constant 403 errors midway. Re-login only gives temporary relief.
English
0
5
11
358
Ben77
Ben77@blapta·
@DeepKlee 我认为是好事,最终会把那些只会干体力活儿,做开源小玩具,刷存在感的人给退出市场。
中文
0
0
2
201
Klee Kawaii
Klee Kawaii@DeepKlee·
这下不是“Junior 先死、Senior 晚点死”,而是凡是靠执行稀缺性吃饭的岗位,都一起被拉进风险区,区别只剩下谁死得更有层次感一点。 学术代码并不是一种神秘代码;它只是以科研活动为业务领域的软件开发。业务系统里实现一个新的 payment strategy,科研系统里实现一个新的 attention module; 业务系统跑 integration tests,科研系统跑 benchmark 和 ablation。 差异主要在领域语义,而不是软件工程本体。 一旦这么看,很多过去被包装成“科研能力”的工作,确实会显露出它其实是: 领域需求→模块接口→实现→实验→指标 只要一位 Research Engineer 把中间这条流水线搭好,研究员每次真正需要提交的,可能就只剩下一个相对局部的 module。 很多论文代码之所以看起来“只有研究员本人才能改”,不是因为算法多么玄妙,而是因为它往往具有: 全局变量横飞; 参数靠 argparse 神秘传递; 模型、训练、数据、评估耦合; 随机种子不可控; checkpoint 格式不稳定; 实验名称编码在路径里; bash 脚本互相调用; 一个 notebook 承载整篇论文; 指标计算和训练逻辑混在一起。 这种代码的困难不是科学困难,而是偶然复杂度(accidental complexity)。 一旦把它转换成:稳定接口+显式类型+可复现实验协议+自动化评估。新增方法的实现就会从“阅读两周祖传代码”变成“实现一个受约束的模块”。Research Engineer 的核心价值,是把开放式科研活动编译成 Agent 可以执行的受约束任务。 大量论文方法实际上也是对一个有限设计空间的搜索。例如深度学习中经常是: Architecture × Objective × Regularization × Data × Optimization 每个维度都有若干候选: 替换 attention; 加一个 residual branch; 改 loss; 调 sampling; 加 auxiliary task; 改 normalization; 改训练 schedule。 如果 infra 已经把这些都做成插件,那么“提出新方法”可能退化成: 模型枚举候选组合; Agent 实现模块; 自动跑小规模筛选; 扩大有潜力的实验; 自动做消融; 研究员选择值得写成论文的叙事。 这本质上是程序化科学搜索(programmatic scientific search)。对很多增量型论文来说,最昂贵的环节过去不是思想,而是实现、试验和耐心。现在这三者同时被机器降价。 这里最黑色幽默的地方是:Research Engineer 把科研模块化之后,也顺手把自己的大量工作自动化了。于是组织结构可能从: 3 位 PI / Scientist 8 位 Research Scientist 10 位 Research Engineer 若干实习生 变成: 1–2 位方向负责人 2–3 位强研究员 1 位平台型 Research Engineer 大量 Agent 和算力 大家倒也未必“都得死”——更像是十个人的工作会收缩成一个人带一群电子幽灵完成。而那个留下来的人,既要懂领域、懂系统、懂评价,还得知道什么时候该叫 Agent 停手。不然它就会再烧 60% 周额度,给你重建一整个文明。
中文
3
0
13
1.2K
Ben77
Ben77@blapta·
@sixsixjay 你说得很对,真面对这样的选择之前,也是以为自己也能成为一代大割呢。后来自己真的碰上了,就是另一回事。经历过之后,会发现其实大家都知道是诈骗,但是利益交错之间,没人会主动提起戳破表演。
中文
1
0
1
79
Jay
Jay@sixsixjay·
另外和ben老师分享一点我的个人感受,我刚入行前几年,自认还算比较理想主义的类型吧,我在17年就开始写微博了,当时大牛市有很多人主动加我,让我建社区,如果我当时随便搞搞什么收费群,代投群,闭眼赚。 但我基本啥都没干,中途还有各种各样的行业机会,我也都没主动选择。 但后来,我们当时fund深度参与的几个项目,然后一些饭局的时候,他们founder或者co-founder和我说他们的规划,内幕,然后提醒我该建仓建仓,该出货出货。 你说是你的话,你能忍住吗?你觉得这种道德吗?还是觉得无伤大雅?反正大家都这么在干。 再往后,各路热钱涌过来,接还是不接?合作还是不合作? 经历了这些事后,我才真正意思到一件事,一个人,怎么看自己,怎么想象自己不重要。只有当真正利益放在面前的时候,你会怎么选,那才是你最真实的样子。 自己想象出来的底线不是真的。 再后来嘛,经历的就更多了,所以现在我基本处于一种精分状态,我已经接受和默认了这些行业现象,但我又不希望它们是这样的。 站在一个从业者的立场,我会去骂,会去呼吁,甚至偶尔有点痛心疾首。 但是站在一个乙方的立场,我给我的甲方提的又是最真实最中肯的操作方式,我不会和他们扯那些虚头巴脑的道德。
中文
1
0
3
134
Jay
Jay@sixsixjay·
老市场+运营人秒懂,这都是真实实操总结下来的行业经验和规律。 一些新入场的团队一天天听那些漂亮话,车轱辘话,那这辈子都有了,先慢慢原地打转个几年再说吧。 那些说漂亮话的人,他总不能在公开场合说,他们当初是怎么舔的,是怎么组局的,是怎么在背后进行利益分配的,是怎么做局请君入瓮的? 就像孙割今天教你创业,教你不忘初心,教你押注趋势。 漂亮小词一套一套的。 但他肯定不会和你聊18年他是怎么和各种玩盘的合作,让盘圈去波场上面玩。 这么多年下来吧,讲实话,很多创业team认知真没比散户高多少。
Jay tweet media
中文
1
0
10
1.3K