2024年的一些总结

这篇总结躺在草稿箱里一个多月断断续续地编辑今天终于能发出来
2024年经历了一个重要的事业转折点这一年大部分的时间是与自己相处跳出了以前的框架想明白了很多事情也许明白只是假象以下是部分罗列

  • 人生是一场游戏人们为自己的人生旅途中赋予了很多意义但其实人生的意义只在于体验游戏让人上瘾因为它们都有一个共同点快速的正反馈可以把人生大小小的困难设计成游戏分解目标建立正反馈系统

    Read More

强化学习和人工智能的奇点

Andre Karpathy近日在X平台上也对R1作了评价评价本身不重要重要的是他引申出了一个更加深刻的观点即AI的自我进化可能远超我们想象这个观点让人细思极恐

他提出合成数据和强化学习是等价的在强化学习的试错过程中每一次试验本质上都是模型生成的合成数据而它随后根据奖励函数来进行学习反过来说当你对合成数据进行筛选和排名时这个过程实际上就是一个0-1奖励函数可见数据很大程度上是算力下游的产物

Read More

互联网极简史

最早畅想互联网的人是谁大部分美国人同意是万尼瓦尔他在诚如所思中的描述是人类对互联网的最早的想象其它版本众说风云但也都指向了20世纪初

未来的作者是否不再使用手写或打字的方式而是直接对着记录机讲话就可以了呢他会通过速记员或者一个大圆桶间接地去做这件事如果他想直接录入一条键入的记录所需的东西都是现成的他需要做的就是充分利用现有装置并转换他的语言

—— 万尼瓦尔.布什 诚如所思1945

Read More

我所理解的创始人模式

前一段时间YC创始人Paul Graham在他的博客上写了一篇文章关于一家公司的创始人应该如何管理公司文章引发了LinkedIn上的热烈讨论甚至有人还拍了恶搞段子

对于我来说创始人应该具备以下三个特点

第一敢为天下先在方向还不明确的时候趟出一条路来 创始就是从0到1无中生有每一个创始人内心都有一个灯塔指引着他有路就走没有路就开路吸引其它人来跟随这个过程往往是非常曲折第一个产品往往是失败的产品别人用起来经常一头雾水第一次融资投资人也会不悄一顾

第二把自己变成销售 创始人应该像销售一样不断地向投资人讲自己的故事但是讲故事只是其中一个能力一个销售最核心的能力是在面对拒绝时候的韧性融资是创业者都要经历的过程被投资人拒绝也是常态创业者内心往往有一股笃定的力量这种力量让他们认为任何困难都是需要克服的任何拒绝都只是暂时的

即使是天才如乔布斯在融资之初也遭遇了很多投资人的拒绝以下的投资者拒绝了Steve Jobs矛盾的是他们作为一个群体最终”资助”了他…

  • Tom Perkins和Eugene KleinerKleiner Perkins拒绝与Jobs会面
  • Bill Draper在一位同事拜访后认为Jobs和Wozniak”傲慢自大”
  • Pitch Johnson对家用电脑表示怀疑问道“你们打算在上面放食谱吗
  • Stan Veit因不信任Jobs的外表拒绝了他以1万美元换取Apple 10%股份的提议
  • Nolan Bushnell拒绝以5万美元购买Apple三分之一的股份但将Jobs介绍给了Sequoia的创始人Don Valentine
  • Regis McKenna拒绝以20%的股份为Apple设计广告“20%的零等于零”但也将Jobs介绍给了Valentine
  • Don Valentine虽然持怀疑态度但在管理和营销方面指导了Jobs

Valentine将Jobs介绍给了三个联系人第三位Mike Markkula看到了Wozniak设计的潜力投资了91,000美元换取Apple 26%的股份成为其第一位天使投资人Markkula说服McKenna帮助Apple做宣传这最终导致了Apple标志性logo的诞生Markkula还说服了Venrock的Hank Smith投资300,000美元换取Apple 10%的股份最终Don Valentine投资了一小部分认为他在Apple董事会的存在会有益处

没有人单独资助了Steve是整个网络做到了他的坚持和人脉关系是Apple早期成功的关键

最后一个创始人的特质是在不确定性中不翻车 硅谷著名孵化器YC的共同创始人Ben Horowitz把他创立第一家公司Loudcloud的经历写进了这本书中The Hard Thing About Hard Things彼时正值互联网泡沫破灭市场的萧条也把公司一次又一次代入了绝境几乎每天都在破产的边缘

2000年互联网泡沫破裂严重影响了Loudcloud的客户基础和融资环境公司不得不大幅降低IPO发行价从每股10美元降至6美元

2001年公司面临持续的客户流失恶化的宏观经济环境和销售前景下滑在首次财报电话会议中不得不下调收入预测911事件后公司最大的交易占预订量三分之一又险些被取消

到2002年Loudcloud的商业模式被证明是不可持续的,Ben决定将公司转型为软件公司Opsware并秘密寻求出售Loudcloud业务, 并最终以6350万美元的现金价格将Loudcloud出售给EDS随后进行了艰难的裁员和重组

回到Paul的文章他讲的其实是在管理公司的时候创始人自主权的边界在哪里在一家公司中哪些事情应该创始人亲力亲为哪些事情可以交给管理者哪些事情可以听管理者的意见哪些事情必须听从自己的内心在Ben Horowitz的故事中虽然他的公司已经达到几百人的规模他手下有着一帮经理人打理公司但他的主要职责还是要带领公司应对外部的不确定性在很多重大决策上面他也不得不听从自己的判断

Paul在文章中提到讨论创始人模式的书籍很少商学院也没有它的存在我觉得这个世界上到处都是这方面的书籍他们就是各个创始人的自传某种程度上The Hard Thing About Hard Things 也是Ben Horowitz的自传这些书也许没有在讨论研究创始人如何运行一家公司因为大概不可能有一个标准的模板

**最后创始人的稀缺性注定他们不能被标准化地定义**市场上不需要那么多的创始人却需要大量的管理者一家公司创始人就那么几个管理者却有数十上百倍的数量需求量大导致了管理都需要规模化的培养而如果创始人可以规模化的培养那似乎就脱离的创始的本质创始人之间以及创始人与外部环境之间的竞争是残酷的我们看到的都是1%成功者没有看到死掉的另外的那99%总得来讲创始人与管理者之间最大的不同就是创始人的任务是开路管理者的任务是把一条开好的小路修成高速公路

OpenAI-o1真的会思考吗

在OpenAI的眼中实现通用人工智能大致要经历5个阶段对话(conversation)推理(reasoning)代理(agent)创新(innovation)组织(organization)OpenAI的o1模型已经达到了推理这一阶段相信第三阶段也指日可待在它到达第四阶段—创新—之前个人认为还需要达到直觉这个阶段

OpenAI在上周推出了最新的模型o1-preview名字的含义是Orion猎户座一代有人说计数器重置为1因此很可能不会有GPT-5o1将开启一个全新的时代
很多人对o1模型能力进行了解读也有大量的新奇的应用如雨后春笋一般涌现出来总得来说o1是思维链(Chain of Thought, COT)的集大成者它很可能推动LLM新一轮的范式改变侧重点从训练到推理转移然后以推理时间的增长来继续scaling law但是它真的会思考

o1模型本质上就是把思维链内化

o1本质上就是把思维链COT内化之前需要依赖人工写COT来指导LLM接近正确答案现在OpenAI的把COT隐匿地训练到了o1的内部就像AlphaGo下棋形成了巨大的树形搜索空间相当于一条条COT构成的空间这里COT的具体步骤的组会空间是巨大的人写的COT未必是最优的问题越复杂这个树的搜索空间就越大搜索复杂度就越高找到正确答案涉及到的COT步聚越多
这一突破意味着复杂的提示词工程(Prompt Engineering)可能会逐渐淡出舞台未来,用户无需再构造繁琐的提示词,AI系统将能够自主生成最适合的思维过程,这无疑是向着更高度智能化迈进的一大步

从训练到推理的转变

简单的来说OpenAI o1 系列模型在复杂推理上的性能提升模式与传统 LLM 预训练式的性能提升不同主要通过强化学习的方式让模型不断完善思考过程包括对不同策略进行尝试认识到错误等这个过程是推理阶段完成的
因此我们应该有两种方式来增加能力在训练期间以及在推理期间这实际上对Nvidia的竞争对手如Groq和SambaNova Systems来说是一个利好信号因为他们有更多机会在推理计算方面竞争而在训练计算方面几乎没有机会竞争

o1会思考吗

人类的物理数学大师一定程度上是不是因为他们计算能力强解决具体问题的逻辑能力强而是他们的强大的直觉物理学中有很多奠基性的理论都是物理学家先有了答案然后再回头去证明和推理心中的那个答案他们的答案是如何来的呢这就是大师们的直觉

A mathematician is a person who can find analogies between theorems; a better mathematician is one who can see analogies between proofs and the best mathematician can notice analogies between theories. One can imagine that the ultimate mathematician is one who can see analogies between analogies.
-—- Stefan Banach

LLM的预训练所得的是对一种简单规则的直觉以前我们通过大量的数据让LLM抽象出了数据背后最一般的映射也就是规则比如语言里的语法一些简单的数学规则基于统计意义上的但是很多复杂问题是多个简单规则的嵌套这时候LLM就力不从心了因为它无法通过死记硬背的方法来学会这些规则组合现在我们有o1模型可以借助COT来解决这些问题这时COT有点像人类的逻辑思考或者Reasoning近一步想如果通过大量的COT作为数据是不是可以训练出对复杂问题的直觉这个时候的LLM是不是就更像一个科学家那样思考了

李沐:语言模型的现状与未来

AI大牛李沐最近回到上海交通大学做了一次演讲聊了聊语言模型的现状和未来趋势李沐的观察很有启发不愧是这个站在这个领域最前沿的人他说现在的AI基本可以完成文科白领的大部分工作但是要完成复杂任务和与真实世界交互还有很长的路要走而制约大模型发展的瓶颈主要是内存电力数据其中数据的重要性在演讲中反复被强调

李沐最后还给交大学子们指点人生迷津他说创业就像是当海盗天天看着市面上什么好的机会一旦发现就all-in抢到了就爽一把没有抢到就死掉了相比起在大公司里上班和读博深造创业对动机motivation的要求是最高的你内心深处要有那么一件十分想去做不去做就抓耳挠腮就会后悔的事情这个事情要满足你最深层次的欲望你还要能抗得住一次次的打击在别人都不看好的时候坚持走你选择的路创业就是要笃定就是要让自己的内心强大到混蛋

以下是李沐演讲中关于技术方面的总结


算力内存和电力将是瓶颈

当前算力发展面临着几个关键问题首先带宽被认为是最重要且最难解决的问题随着模型规模的不断扩大数据传输速度成为制约因素目前每根光纤可以提供400Gbps的带宽未来有望达到800Gbps然而带宽的提升对于大规模分布式训练至关重要其次内存大小直接限制了模型的上限目前单个芯片可以封装192GB内存但这个数字在未来几年内可能难以突破这意味着模型大小可能会在一定程度上受到限制除非出现重大技术突破

虽然计算能力仍在遵循摩尔定律增长但NVIDIA在市场上的垄断地位影响了价格下降的速度这导致了高性能计算设备价格居高不下增加了研究和应用的成本随着计算规模的扩大供电成为一个日益突出的问题大规模AI训练中心的用电量已经达到惊人的程度甚至出现了自建发电厂的想法以降低运营成本

尽管面临这些挑战长期来看算力仍将变得越来越便宜这一趋势将持续推动AI技术的发展和普及然而如何有效利用和优化现有算力资源将成为未来AI研究和应用的重要课题

模型与算法不存在完全的垂直模型

目前语言模型的规模已经达到一个相对稳定的区间预训练数据量通常在10T到50T之间模型参数在100B到500B之间这个规模已经能够捕捉到大量的语言知识和世界知识但进一步增大可能会面临收益递减的问题

语言模型正在向多模态方向发展特别是在音频和视频领域取得了显著进展这使得模型能够理解和生成更丰富的内容如语音合成图像生成和视频处理等随着语音识别和合成技术的进步预计语音交互将在未来变得更加普及和自然这可能会改变人机交互的方式使得与AI系统的交流更加直观和便捷

预训练已经逐渐变成了一个工程问题而后训练则成为了技术难点如何有效地利用预训练模型并通过后训练使其适应特定任务成为了研究的重点实践表明并不存在真正的”垂直模型”即使是针对特定领域优化的模型也需要强大的通用理解和推理能力作为基础这意味着提升模型的整体智能水平仍然是核心任务

李沐也指出了评估的重要性目前的的评估方法往往无法反映模型在实际应用中的真实表现评估的复杂性源于自然语言的多义性,以及需要考虑准确性语言风格和逻辑连贯性等多个因素设定合适的评估标准是一个关键挑战,这些标准需要能够准确反映模型在实际应用中的表现,而不仅仅是在特定测试集上的表现李沐还指出了评估与数据之间的密切关系,好的评估方法本身可以成为有价值的数据来源,不仅是检验模型的手段,也是改进模型的重要途径

数据数据数据

李沐在演讲中多次强调了数据的重要性只要能采集到足够的高质量数据任何领域都有可能被自动化这意味着未来的竞争可能更多地集中在数据获取和处理能力上高质量数据对模型效果有巨大影响相比简单地增加数据量提高数据质量往往能带来更显著的性能提升

尽管模型训练技术不断进步但大部分时间仍然需要花在数据处理上如何高效地收集清洗和标注数据成为了AI项目成功的关键因素之一随着数据在AI发展中的重要性日益凸显数据伦理和隐私保护也成为了不容忽视的问题如何在充分利用数据的同时保护个人隐私和维护社会公平是整个行业面临的重大挑战

现在的AI还只是一个文科白领

文科白领工作是AI目前最容易替代的领域这包括了写作个人助理教育游戏策划等需要使用自然语言与人和世界打交道的工作李沐认为在这些领域语言模型已经能够完成80%到90%的工作例如AI可以生成各种文章报告处理文本甚至进行教学任务这种高度的替代能力主要归功于大型语言模型在处理和生成自然语言方面的卓越表现

工科白领工作如编程和问题解决目前还难以被AI完全替代但AI已经能够提供很大的辅助作用比如帮助程序员完成一些基础的编码任务如搜索和修改代码片段然而对于更复杂的编程任务如系统设计或解决复杂问题AI还无法完全取代人类

对于大多数需要与复杂物理世界互动的蓝领工作如搬运货物或服务行业AI替代还面临着巨大挑战这是因为这些工作需要理解和适应高度复杂和变化的环境这对AI来说仍然是一个难题当然也有例外自动驾驶是一个突出的例子它在特定的封闭环境中取得了显著进展这主要是因为交通环境相对稳定且可以收集大量数据李沐估计要让AI在这些领域取得突破可能还需要至少五年的时间来建立必要的基础设施和收集足够的数据

AI is electricity in late 1800s

Formal Google CEO Eric Schmidt has stirred quite a bit of controversy with his comments on Google’s work life balance, the Ukraine War, and startups strategy on pursuing growth. Putting those aside, what resonated with me was a concept brought up by the host Erik Brynjolfsson, that is general purpose technology. He said general purpose technologies are powerful because they ignite other complimentary innovations, but it usually takes process innovation or organizational innovation to fully unleash that power.

He used electricity as an example. When electricity was first introduced into factories, they didn’t become significantly more productive than the factories that were powered by steam engines.

The way steam engines worked was that there was this big steam engine sitting in the middle of the room, and power from that machine got distributed through crankshafts and pulleys to all the equipment. When they introduced electricity into factories, they would pull out the steam engine and got the biggest electric motor they could find and put it where the steam engine used to be and fired it up. But It didn’t change the production a whole lot.

It wasn’t until 30 years later that they started seeing a fundamentally different kind of factory where the giant central steam engine was replaced by many much smaller electric motors. When they started doing that, they started to have a new layout of factories. The layout was typically on a single story where the machinery was not based on how much power they needed but based on something else, like the flow of materials. And people started having these assembly line systems, which led to huge improvement in productivity. Henry Ford was one of the best examples in this revolution.

The lesson is that electricity is a fundamentally valuable technology. But it isn’t until you have that process innovation, or organizational innovation of rethinking of how to do production that you got the big payoff.

Electricity is just one example. General purpose technologies also include steam engine, railroad, mobile and information technology, etc. And in these other technologies, people had similar generational lags before they realized that this technology allowed them to do something completely different than they sued to do.

AI is bit like that in some way, there is going to be a lot of organizational innovations, new business models, new ways of organizing an economy that we haven’t thought before. Tools like ChatGPT are already making existing organizations a lot more productive. But right now we are just doing retrofitting.

It is not just technical skills that is important, it is also about rethinking other stuff. And there is a lot of opportunity for us to rethink our areas now that we’ve been given this amazing set of technologies. We have to be creative enough to think about where the gap is. And it will be even bigger once people figure out these complimentary innovations.

Patrick Collison's latest interview on Jensen Huang

“I’d rather torture you into greatness than firing you.” —- Jensen Huang

On April 24th, Patrick Collison, CEO of Stripe, interviewed Jensen Huang, who is the CEO of Nvidia. Below are the main takeaways:

No One-on-One Meetings. Jensen has 60 directors who report to him, but he prefers not to have one-on-one meetings. By having all 60 directors together at once, he has eliminated at least 7 layers from the company’s structure. He believes it is more effective to communicate with all stakeholders simultaneously, allowing information to flow more efficiently among people. Everyone can contribute to and share in the information flow. Feedback is learning. Not only do you work to solve the problem at hand, but you also create conditions for others to learn from your situation. On the other hand, isn’t it a great opportunity to learn from other people’s mistakes, disasters, and strategies?

No Reports or Operational Meetings. Jensen does not have regular meetings on his schedule. At heart, he is an engineer who prefers attending meetings focused on specific problems or brainstorming sessions.

Tenure Over Hiring. Despite Nvidia’s $2 trillion valuation, it has only around 28,000 employees, whereas Microsoft, valued at $3 trillion, has over 200,000 employees. When asked about Nvidia’s efficiency, Jensen said he prioritizes tenure over hiring new people. He would rather torture existing employees into greatness. He has embodied joy and suffering with his employees, building trust along the way.

CUDA’s Persistence. When CUDA first launched, its lack of applications nearly crashed Nvidia’s valuation. However, Jensen deeply believed in its value and insisted on persisting. It took ten years for CUDA to become successful, without which Nvidia would not have achieved its current success in AI.

Valuing Entrepreneurs. Jensen values entrepreneurs based on three criteria:

  1. Gut check: Having faith in what you’re building.
  2. Reasoning: Being able to reason through your goals and convince others.
  3. Cleverness: Having the cleverness to survive challenges, risks, and downfalls.

Industrial Revolutions

  1. The first industrial revolution turned steam into power output.
  2. The second industrial revolution turned fossil fuels into electricity.
  3. The third industrial revolution turned electricity into information.
  4. Today’s AI is the fourth industrial revolution, turning electricity and data into tokens, which will become the new currency in this era.

Llama2’s Impact. When asked about Llama3, Jensen instead mentioned Llama2, which he considers revolutionary for activating various industries and research areas, allowing them to adopt language model techniques. He believes language encompasses more than human speech, extending to life, nature, physics, and other domains that language models can learn.

The Art of Strategy: Lessons learned from Go

Go, the ancient board game originating from China, is not merely a game of skill but a profound exploration of strategy and decision-making. Its core concepts—such as Big Points, Frameworks, Influence, and Initiative—not only guide players to victory but also serve as metaphors for navigating life’s complexities. These principles offer invaluable insights into making better choices, anticipating outcomes, and achieving long-term success.

In Go, Big Points refer to key strategic areas on the board that yield the greatest potential gains. These points are often the first moves in a game and set the tone for future developments. Ignoring these points can lead to missed opportunities or allow opponents to gain a decisive advantage. Similarly, in life, identifying and prioritizing high-impact opportunities is essential. For example, an entrepreneur might focus on entering a fast-growing market rather than investing resources in smaller ventures. Recognizing the big points in life helps allocate energy effectively, ensuring meaningful progress.

A Framework in Go involves establishing a loosely defined area with a few stones that suggest potential expansion. It is a move that declares territorial ambition without committing fully, allowing room for flexibility. In Go, frameworks are crucial because they create opportunities for future growth while staking a claim in open spaces. In real life, this could translate to laying the groundwork for long-term success, such as networking in a new industry, investing in emerging technologies, or developing foundational skills. While the immediate benefits may not be clear, frameworks prepare the way for larger successes down the road.

The expansion of a framework leads to Influence, a concept in Go where stones dominate a large area, shaping the game’s dynamics without necessarily forming concrete territory. Influence is powerful because it represents potential, deterring opponents and opening strategic possibilities. In life, influence is equally intangible yet impactful. It can be seen in the reputation of a leader, the promise of an innovative idea, or the networking power of a strong connection. Although not immediately measurable, influence shifts the balance of power and creates a significant edge in competition.

In both Go and life, Critical Points often determine the outcome. These are moments or moves that demand immediate attention because they significantly affect the game’s trajectory. Winning or losing a critical point can mean the difference between control and chaos on the board. In life, these moments arise during pivotal exams, career-defining decisions, or crucial business deals. Success depends on the ability to recognize these moments, focus all efforts, and act decisively. Hesitation in these scenarios can lead to irreversible losses, whereas timely and calculated action can secure lasting advantages.

Initiative, or Sente in Go, refers to moves that force the opponent to respond, allowing the player to dictate the flow of the game. Retaining initiative is crucial for maintaining control and staying ahead in the match. In real life, initiative is about taking the lead—setting the agenda in a meeting, framing the narrative in negotiations, or introducing innovative ideas that demand attention. It empowers individuals to shape outcomes proactively rather than reactively. While taking initiative sometimes involves calculated risks, its long-term rewards often justify the effort.

Go also teaches the importance of Residual Value—or Aji, an unresolved position left on the board that can later be exploited for strategic gain. These are latent opportunities that hold potential for future use. In life, residual value appears as untapped resources, unfinished projects, or dormant relationships. Learning to recognize and preserve these assets for later leverage can provide a competitive advantage when circumstances align.

Similarly, Ko Threats in Go are moves used in negotiations during a Ko fight, where players vie for dominance in contested areas. A ko threat forces the opponent to respond, creating leverage. In real life, these threats might be social capital, intellectual property, or financial resources strategically deployed during critical negotiations. Effectively utilizing these threats can sway outcomes and secure advantages.

Finally, Playing Elsewhere, or Tenuki, is a strategic move in Go where a player abandons a local skirmish to make a more impactful move elsewhere on the board. This reflects the importance of prioritizing high-value opportunities over minor distractions. In life, tenuki teaches us to focus on what truly matters, even if it means letting go of less important tasks. It requires the discipline to avoid becoming overly attached to sunk costs and the foresight to invest in areas with greater potential.

Go’s principles remind us that success is not about immediate gains but about long-term vision, resource management, and adaptability. By understanding the importance of big points, frameworks, influence, and initiative, we can approach life’s challenges with clarity and confidence. In both Go and life, the key to mastery lies in seeing the entire board and making purposeful moves toward a meaningful goal.

爆米花大脑和刷手机

最近听到一个词叫作爆米花大脑 用来形容由于长时间上网而导致过度刺激引起大脑神经电路”爆裂”的感觉爆米花大脑”这个概念是由心理学家Dr. Levy提出的 与简单的网络成瘾不同爆米花大脑在现代生活中无处不在尤其是当人们感到压力或焦虑时更为常见

正如哈佛大学医生Aditi Nerurkar所解释的那样人们不断滚动浏览信息的冲动本质上是一种由于进化的缘故而根植于人类基因中的扫视危险的本能这一本能主要由大脑中的杏仁核主导只不过现代人已将扫视的对象从物理环境转移到了手机社交媒体等数字内容上受这种原始的扫视危险的冲动驱使过度接触数字信息就会导致精神过度刺激和注意力分散造成所谓的”爆米花大脑”状态

最近休息了一段时间我才意识到之前我的大脑似乎总是处于一种紧绷状态收集的信息越来越多深度思考越来越少工作时间排得满满当当大多数时候只是在重复一些既定的行为模式谈不上高效率也没有很多创造性的思考注意力经常涣散如果在思考时产生短暂的卡顿就会控制不住地打开手机刷起短视频新闻等

和好友聊到这个话题他说我们用忙碌来填充无意义工作带来的空虚感这种忙碌的状态给我们一种高效的错觉实际上深度工作一书中提到的真正重要的是把有限的注意力集中在关键任务上而非简单地填满时间这本书提到了”固定日程生产力”的方法就是将工作内容与时间期限分离开避免简单地将工作时间等同于工作完成的多少有时候严格地限制工作时间反而会取得意想不到的效果

如何拯救被忙碌和信息洪流毁掉的爆米花大脑通过这段时间的休息我逐步找到了一些应对之策:

第一步是真正暂时抽离放下手头的一切工作和烦扰给自己一段安静独处的时光正如中国古语所说”知止而后有定定而后能静静而后能安安而后能虑虑而后能得只有先学会停下来远离纷扰才能开始审视内心找到真正重要的事物

第二步是做减法剔除生活中那些不太重要的人和事对我而言最重要的是自身的身体健康(其中包括大脑)与家人朋友的感情联系很多表面上维系的人际弱联接和害怕错过的信息源大部分的时间只是在消耗注意力可以适当的抛弃我尝试把邮箱里未读的几十封Newsletter统统删掉第一次感受到前所未有的清爽与轻松关于弱联接还有一句话十分受用不要想着通过强连接来提升自己或者通过弱连接来证明自己很多社交上的心累就是因为过于在意别人对你的看法

第三步是建立界限去掉导致不当行为的诱因(Prompt)心理学家BJ福格提出了影响行为的”福格行为模型”:B=MAP其中B是Behavior(行为)M是Motivation(动机)A是Ability(能力)P是Prompt(诱因)刷手机只需要动动手指而我们的动机往往以很高所以一个有效的方法就是去除诱因为了去掉那些可能诱发分心行为的诱因我清理了微信上的朋友圈只选择性关注少数人的朋友圈我静音了全部消息包括个人和群组并把大部分群组都移到了隐藏组里我给手机设置了晚上10点到早上8点之间是downtime没有消息提示并且所以的消息类应用在这期间都无法使用

然后我发现整个世界都清静了原来主动截断信息并不会让你错过重要的事情反而让你可以静下来仔细想想哪些信息和人际关系是重要的静而后能安安而后能虑我开始想如何去主动获取少数的重要的信息这种掌控感很棒