即刻App年轻人的同好社区
下载
感谢嘻嘻酱来DC拯救我很久没碰酒精的final week
是嘻嘻酱: 是shei，仔细看了地铁站名还是坐反了方向？！是我！和@Scarlett-事业脑版 面即严重迟到中🧏🏽♀️
4 00
我感觉我每次听Sam Altman的访谈，我处理信息的multi-head attention机制最多的注意力都是花在学习Sam的diplomatic回应 转化成like“大概怎样的置信区间内他可能意思是xxx”，以及学习host们的访谈（审讯）技巧，然后再是Sam的防御式回应🤧 这是我能从他的公开发言中学到最多的部分
All in这期是host这边有四个人怼着他问，问题都挺sharp的，Sam确实是有点东西 www.youtube.com
看了一下高赞评论Top3 🤣
“Sam Altman - The master of talking without saying anything”
“I think Sam speaks more to confuse than to clarify”
“Why can’t this guy just answer a question directly? The panel needs to press him more to his face, they criticized him so much in past podcasts.”
All in这期是host这边有四个人怼着他问，问题都挺sharp的，Sam确实是有点东西 www.youtube.com
看了一下高赞评论Top3 🤣
“Sam Altman - The master of talking without saying anything”
“I think Sam speaks more to confuse than to clarify”
“Why can’t this guy just answer a question directly? The panel needs to press him more to his face, they criticized him so much in past podcasts.”
25 11
赶紧结束final week坐大牢🙏
然后18-19号去Brooklyn 听两天Jazz
🥰22号去湾区找男朋友呆到下个月3号，其间穿插25-27一起去San Diego，1-2号再随机去个national park
美东美西谈出了异国恋的感觉💀
然后18-19号去Brooklyn 听两天Jazz
🥰22号去湾区找男朋友呆到下个月3号，其间穿插25-27一起去San Diego，1-2号再随机去个national park
美东美西谈出了异国恋的感觉💀
16 10
很有意思，人类简史那个Yuval Harari 在Lex Fridman访谈上提到类似的观点，虽然表述和展开没有即友这么具体深入。真正渗透到人类社会交往范式中应该是需要这么一套系统的
王半城: 补看了《her》，能猜到这部电影为啥引起伪科学狂欢了，他们看不见剧情跟现实AIGC机制的差别，只觉得堆砌「智力和多模态」就能骗过人的感官。 但实际剧情表现出的自我意识，不是足够聪明，而是「为了自我生存权益而建立整套反馈策略」。拥有独立策略会让一个事物拥有「权」，我们就会把这种「权」赋魅为「意识」。斯嘉丽约翰逊的台词全是围绕这一点设计的。她有自己的主张，她不是谁的附庸。猫狗也是拥有这套完整策略的，尽管他们不会说话。但权责的主观主体性，让人愿意在乎、尊重，并触发共情羁绊机制。 如果找个里程碑标志，就是AI能否出于自我利益而主动向人发起权责征询请求行为。这跟智能度无关，只跟博弈策略设计有关。
8 00
这本书真不错，a16z 管理合伙人教创业者怎么work with VC，今天读了20%，对于创业者或者想了解美国创投生态来说都很值得一读。不过似乎还没有中文版
33 35
择偶标准这件事，遇到之前是前馈神经网络，遇到之后是这个人作为ground truth 通过反向传播算法来调整原先预设的权重。
之前：adventurous & smart 第一位
现在：adventurous & smart依然很重要，但是绝对值够用就好。communicative & emotionally available 以及 physical compatibility也都很重要。依然不喜欢家庭观念太重的人 但是能和身边的人vibing以及对家人朋友的commitment程度远比冷漠的本人高 似乎也是个不错的plus。
之前：adventurous & smart 第一位
现在：adventurous & smart依然很重要，但是绝对值够用就好。communicative & emotionally available 以及 physical compatibility也都很重要。依然不喜欢家庭观念太重的人 但是能和身边的人vibing以及对家人朋友的commitment程度远比冷漠的本人高 似乎也是个不错的plus。
17 32
目前 robotics 尤其是通用机器人到底研究进展如何？可以从Google DeepMind 在机器人方面在去年 PaLM-E 之后的六篇重要的工作中get到大概，www.youtube.com 请DeepMind researcher 详细walk through了每一篇。high-level 来看，除了最大的bottleneck依然是数据，算法架构这边虽然还有一些challenges，但是离能够scale不远了。
👇六篇research
1️⃣ RT-2 (28 Jul 2023): how Internet scale vision language models allow robots to understand and manipulate objects that they have never seen in training. arxiv.org
2️⃣ RT-X (13 Oct 2023): a collaboration with academic labs across the country that demonstrates how a single model trained to control a diverse range of robot embodiments can outperform specialist models trained for individual robots. arxiv.org
3️⃣ RT-Trajectory (3 Nov 2023): a project that shows how robots can learn new skills in context from a single human demonstration represented by a simple line drawing. arxiv.org
4️⃣ AutoRT (23 Jan 2024): a system that scales human oversight of robots even in previously unseen environments using a combination of large language models and a robot constitution to power firstling ethical and safety checks. arxiv.org
5️⃣ Learning to learn faster (18 Feb 2024): an approach that enables robots to learn more efficiently from human verbal feedback. arxiv.org
6️⃣ PIVOT (12 Feb 2024): a project that shows how vision language models can be used to guide robot action, this time with no special fine-tuning required. arxiv.org
👇六篇research
1️⃣ RT-2 (28 Jul 2023): how Internet scale vision language models allow robots to understand and manipulate objects that they have never seen in training. arxiv.org
2️⃣ RT-X (13 Oct 2023): a collaboration with academic labs across the country that demonstrates how a single model trained to control a diverse range of robot embodiments can outperform specialist models trained for individual robots. arxiv.org
3️⃣ RT-Trajectory (3 Nov 2023): a project that shows how robots can learn new skills in context from a single human demonstration represented by a simple line drawing. arxiv.org
4️⃣ AutoRT (23 Jan 2024): a system that scales human oversight of robots even in previously unseen environments using a combination of large language models and a robot constitution to power firstling ethical and safety checks. arxiv.org
5️⃣ Learning to learn faster (18 Feb 2024): an approach that enables robots to learn more efficiently from human verbal feedback. arxiv.org
6️⃣ PIVOT (12 Feb 2024): a project that shows how vision language models can be used to guide robot action, this time with no special fine-tuning required. arxiv.org
4 01
Yesss有意思的是No Priors的访谈在最开始就讨论了团队在ioi （很牛的一个programming competition）的背景和经历，我在这里get到的点是所以他们对于人类顶尖程序员是怎么解决通用的programming & algo 问题的理解和实践是top tier的 //@High寧: 听下来 devin 在已有模型 reasoning 能力的基础上采取的是模仿程序员 debug 的方式，至少从这个路径开始迭代？
Scarlett-事业脑版: 通勤路上听了一下新鲜出炉的Cognition Co-founder & CEO Scott Wu在No Priors上的访谈，关于我最关心的问题How does Devin work不出意外地只给了一个high-level的设计思想，不算有增量信息。 虽然披露的信息都有限，和前两个月被Nat Fridman & Daniel Gross 投了100M的 Magic.dev 应该在技术路线上还是有差异化的。 Magic 自己训long context（先进的分片，Nat认为不输Google Gemini）和拥有active reasoning能力（这部分估计也是在on top of base model用agent、tooluse的角度来做，预训练部分肯定也会着重加强coding generation能力）的模型+拥有几千张卡（除了训练模型用，对于推理也有一定竞争力，虽然微软卡多，GitHub CoPilot能分配到的算力量不清楚）。 Devin这部分去掉filling words再rephrase一下： So one way that Scott frame it is that suppose you were given an GitHub issue need to be solved like bug fixing. Theoretically some perfect intelligence can just take the entirely of the code base and future out exactly what’s going on and what needs to be fixed. But practically for humans and today’s AI, the cleaner path would involve running the code yourself, reproducing the bug, adding debugging and print statements, reviewing the logs, asking Stack Overflow etc. So a lot of efforts is about figuring out how to get depth and think and make decisions in that way, involving planning and evaluation.
5 10
通勤路上听了一下新鲜出炉的Cognition Co-founder & CEO Scott Wu在No Priors上的访谈，关于我最关心的问题How does Devin work不出意外地只给了一个high-level的设计思想，不算有增量信息。
虽然披露的信息都有限，和前两个月被Nat Fridman & Daniel Gross 投了100M的 Magic.dev 应该在技术路线上还是有差异化的。
Magic 自己训long context（先进的分片，Nat认为不输Google Gemini）和拥有active reasoning能力（这部分估计也是在on top of base model用agent、tooluse的角度来做，预训练部分肯定也会着重加强coding generation能力）的模型+拥有几千张卡（除了训练模型用，对于推理也有一定竞争力，虽然微软卡多，GitHub CoPilot能分配到的算力量不清楚）。
Devin这部分去掉filling words再rephrase一下：
So one way that Scott frame it is that suppose you were given an GitHub issue need to be solved like bug fixing. Theoretically some perfect intelligence can just take the entirely of the code base and future out exactly what’s going on and what needs to be fixed.
But practically for humans and today’s AI, the cleaner path would involve running the code yourself, reproducing the bug, adding debugging and print statements, reviewing the logs, asking Stack Overflow etc.
So a lot of efforts is about figuring out how to get depth and think and make decisions in that way, involving planning and evaluation.
虽然披露的信息都有限，和前两个月被Nat Fridman & Daniel Gross 投了100M的 Magic.dev 应该在技术路线上还是有差异化的。
Magic 自己训long context（先进的分片，Nat认为不输Google Gemini）和拥有active reasoning能力（这部分估计也是在on top of base model用agent、tooluse的角度来做，预训练部分肯定也会着重加强coding generation能力）的模型+拥有几千张卡（除了训练模型用，对于推理也有一定竞争力，虽然微软卡多，GitHub CoPilot能分配到的算力量不清楚）。
Devin这部分去掉filling words再rephrase一下：
So one way that Scott frame it is that suppose you were given an GitHub issue need to be solved like bug fixing. Theoretically some perfect intelligence can just take the entirely of the code base and future out exactly what’s going on and what needs to be fixed.
But practically for humans and today’s AI, the cleaner path would involve running the code yourself, reproducing the bug, adding debugging and print statements, reviewing the logs, asking Stack Overflow etc.
So a lot of efforts is about figuring out how to get depth and think and make decisions in that way, involving planning and evaluation.
8 34