CHRISTY CHEONG

[CHRISTY CHEONG @christype2000]

another year ~

[Daniel Jeffries @Dan_Jeffries1]

I call on the California Assembly to vote no on SB-1047 and Governor Newsom (@GavinNewsom) to veto this pernicious bill if it does pass, or it's likely to destroy California's fantastic history of technological innovation.

This is not an AI safety bill. It is a Trojan horse.

It's designed to look simple and measured but its real purpose is to give a small, fringe group of anti-AI extremists a kill switch over advanced AI, while throwing a monkey wrench into California's tech industry.

The reason is clear. It was designed by a small group of non-profits who believe AI will cause the end of the world. Dan Hendrycks (https://t.co/oLMZAMfeH0), co-founder of the Center for AI Safety, one of the sponsors of the bill, believes with 80% certainty that AI will end all life on the planet. If you start with that dangerously delusional premise you can't possibly craft a sound, sane, innovation protecting bill. You can only craft a bill designed to dramatically slow down or destroy the growth of AI altogether.

California is faced with a clear choice:

Either accept the dystopian sci-fi hallucinations of this small fringe movement (https://t.co/gvcw0CuVz3), with some people in that movement beginning to openly call for violent terrorism against AI labs (https://t.co/EWxavFKjKl), or bet on California and its unique ability to bring together the best and brightest in the tech world to build the next generation of software.

It's really as simple as that.

Since any bill that starts from the delusional premise of an AI doomsday is fatally flawed by design, I urge lawmakers to start from scratch with a clean, clear bill designed to protect California's prosperity and its world leading tech industry. Consider starting with this draft, SB-1048 (https://t.co/rWqtYxA50q), written by pro-California and pro-innovation leaders in the community.

Passing the fatally flawed SB-1047 means passing the torch of the future to other states and other countries.

This is not a prediction. It's just basic cause and effect.

If a model company can choose between two states, one with a high and unreasonable legal burden and one without it, which will they choose as business owners?

Governor Newsom has outright said “I don’t want to cede this space to other states or other countries." (https://t.co/vno8BMSaYD)

That is exactly what the bill will do, Governor, unless you veto it or push back on it swiftly.

It's worse than that too. Not only will you cede this territory to other states, you risk ceding it to non-western dictatorships that do not share our values and who are racing ahead with AI.

China already has advanced advanced video generation models to rival OpenAI (https://t.co/7dJAs0S22u), a self-driving train (https://t.co/6vueYgVwPa) and robo-taxis (https://t.co/mueSXQOTr2) in production, Alibaba City Brain (https://t.co/Lta19jF0l7) deployed to dozens of cities to control traffic and prevent accidents, as well as advanced military AI (https://t.co/0psVpp8rqu).

We cannot afford to cede the future of AI to anti-democratic regimes.

Unfortunately, this bill has continued to move forward, despite the best efforts of the AI business community and more rational voices who have pushed for much needed changes to the bill but been largely ignored. Senator Scott Weiner says he is open to changes (https://t.co/jrlW246bP8). But the changes his team have made are mostly insubstantial.

The core of the bill is still utterly broken. When the core is broken it doesn't matter what else you do. Minor changes to the rest of it don't effect the essence.

The bill makes model manufacturers certify that nothing bad can ever be done with their models in the future. It holds them responsible for everything that goes wrong forever. This is like making Ford responsible for every drunk driver.

Regulate driving. Hold the drunk driver responsible, not the car maker.

Anti-AI advocates like to say that opponents of this bill hate any and all legislation but that's utter nonsense they know it. Senator Weiner even said "Silicon Valley doesn't like regulation." No, we don't like this bill in particular. We are all for speed limits on the road. That regulates the user of the technology, the car.

We are saying very clearly:

Regulate the use case.

Don't hold the models responsible for everyone else's crimes. Don't punish Microsoft because someone at Enron used Excel to defraud investors. Punish the cheats at Enron.

This is not very complicated. It works in all aspects of law and it has for thousands of years.

So how did we get here? Why has this bill gotten so far?

One of the worst reasons that I've heard from advocates of premature legislation is this persistent catch phrase:

"Look at social media. It's out of control. We should have gotten ahead of it. We've got to get ahead of AI. We've got to do something now."

This is broken reasoning. No regulator could have predicted social media's impact on the world three decades ago. When Representatives Cox and Wyden were crafting the section 230 amendment to the Communications Decency Act in 1996, the number one question of many representatives at the time was "What is the Internet?"

This idea that we can "get ahead of it" is an insidiously seductive message. We all like to believe in our own ability to perfectly predict the future. The problem is we're all just hopelessly bad at it. And the farther out that we go the more impossible it gets.

With all do respect to the Assembly members and Governor Newsom and their respective accomplishments, there is no chance that you or anyone else on the planet can accurately predict what will happen in ten or twenty or thirty years with AI. This is not an indictment of your intelligence or skills in other areas of life. It's just a simple fact.

The future is trillions and trillions of variables all changing at the same time. We might be able to predict small, close up events but the farther you get into the future, the more things change.

It's an illusion to believe that the Senators and Congresspeople in the 1990s could have foreseen remote teleworking, trillions of dollars in e-commerce transactions, the end of book stores, virtual reality, cryptocurrency, the darknet, massively multiplayer games, and social media for all its goods and ills. In the Wright Brothers era nobody on Earth, in Congress or anywhere else, is capable of writing the safety requirements of the Boeing 787 Dreamliner. You cannot design meaningful safety standards for how a commercial jet will function in 2024 back in 1904 when Wilbur is flying a glorified kite of fabric and wood for a few circles around Kitty Hawk.

If we can't predict the future, what do we do? Are we doomed?

No. We do what we've always done. We watch and observe and we address concrete, real world problems as they happen in the real world. For instance, where life and death is involved we should hold technology creators to a higher standard. If you are making a self-driving car, which can kill or injure people when things go wrong, you should have a higher standard.

The only way to properly regulate is to watch as the complex evolution of society and technology play out over time. You see it develop for good and bad, because it is always both, and you try to reduce the ills and incentivize the good. That is society and law in a nutshell. Punish bad, incentivize good.

As a community we feel frontier model companies should be publicly transparent on safety measures and realistic safety efforts like red teaming, which is designed to address actual real world problems with models now, not fantasy sci-fi problems. But we do not agree that model makers, under penalty of perjury, sign a statement saying their model can never be used for anything bad. This is asking for an impossibility.

It's asking people to perfectly predict the future and certify that their model can never be used for any ill purpose. Nobody can do this and it is likely already unworkable under existing contract law. You cannot ask someone to sign a document that says they have to flap their arms real fast and fly. If you do ask someone to sign it, they can easily argue in court that it is "impossible to comply" and that provision is now unenforceable.

As the California legal team at Stimmel law writes on their blog (https://t.co/NDWgWB41Iy):

"A party can invoke impossibility and argue that it did not perform its contractual obligations because it was impossible for it to do so."

It is absolutely impossible for a model creator to certify that there is no possible way anything can go wrong with a model in the future. As we already noted, nobody can predict the future perfectly. Nobody is an all knowing oracle, including me and anyone else who tries to predict the future regularly. It will be impossible for them to testify to future events that haven't happened yet.

If the law does pass, every model manufacturer will likely immediately take the state to court on these grounds and win.

It's time for a step back. Please look carefully at this bill and understand its true goals before you make your choice. This bill is designed to look like an AI "safety" bill but its real goal is to make sure that no advanced AI ever develops at all and that of it does, a few self appointed overseers from a fringe movement can kill it.

If your goal is to ensure AI doesn't develop in California and America, or other democratic countries, then you're on the right track because the sponsors of the bill wrote it to do just that.

But if that's not your goal, then I urge you to go back to the drawing board as quickly as possible.

Nobody can predict the future perfectly but we can make one prediction with confidence.

If this bill passes, the next technology revolution will be outside of California.

If that's what you want, Governor and Assembly members, you're making great strides.

If not, there's still time to change course.

[大葱 Fred @fredli25]

最近来到香港后，和一些投资人同行聊了聊，发现大家对于AI和DePIN赛道的关注度还是蛮高的。分享下最近关于AI和DePIN的思考，和大家一起交流讨论。

核心讨论四个问题：

1、为什么大部分去中心化算力项目都选择做AI推理而不是AI训练？
2、英伟达究竟牛在哪里？去中心化算力训练难做的原因在哪？
3、去中心化算力（Render、Akash等）的终局会是什么样？
4、去中心化算法（Bittensor）的终局会是什么样？

接下来咱们层层抽丝剥茧：

1）纵观这个赛道，除了Gensyn之外，大部分的去中心化算力项目都选择做AI推理而不是训练，核心在于对于算力和带宽的要求不同。

为了帮助大家的简单理解，我们来把AI比作一个学生：

- AI训练：如果我们把人工智能比作一个学生，那么训练就类似于给人工智能提供了大量知识、示例也可以理解为我们常说的数据，人工智能从这些知识示例中进行学习。由于学习的本质需要理解和记忆大量的信息，这个过程需要大量的计算能力和时间。

- AI推理：那什么是推理呢？可以理解为利用所学的知识去解决问题或者参加考试，在推理的阶段中人工智能是利用学到的知识去解答，而不是活动新知识，所以在推理过程所需要的计算量是较小的。

容易发现，两者的难度差本质上在于大模型AI训练需要极大的数据量，以及对数据高速通信需要的带宽需求极高，所以目前去中心化算力用作训练的实现难度极大。而推理对于数据和带宽的需求小的多，实现可能性更大。

2）那么数据和带宽的卡点究竟在哪？为什么去中心化训练难以实现？

这就涉及到大模型训练的两个关键要素：单卡算力和多卡并联。

- 单卡算力：目前所有的需要训练大模型的中心，我们把它叫做超算中心。为了方便大家的理解，我们可以以人体作为比喻，超算中心就是人体的组织，底层单元GPU就是细胞。如果单个细胞（GPU）的算力很强，那么整体的算力（单个细胞×数量）也可能会很强。

- 多卡并联：而一个大模型的训练动辄就是千亿GB，对于训练大模型的超算中心来说，至少需要万级别个A100打底。所以就需要调动这几万张卡来进行训练，然而大模型的训练并不是简单的串联，并不是在第一个A100卡上训练完之后再在第二张卡上训练，而是模型的不同部分，在不同的显卡上训练，训练A的时候可能需要B的结果，所以涉及到多卡并行。

为什么英伟达这么厉害，市值一路腾飞，而AMD和国内的华为、地平线目前很难追上。核心并不是单卡算力本身，而是在于两个方面：CUDA软件环境和NVLink多卡通信。

- 一方面，有没有软件生态能适配硬件是非常重要的，如英伟达的CUDA系统，而构建一个新的系统是很难的，就像构建了一个新的语言，替换成本非常高。

- 另一方面，就是多卡通信，本质上多卡之间的传输就是信息的输入和输出，怎么并联，怎么传输。因为NVLink的存在，没有办法把英伟达和AMD卡打通；另外，NVLink会限制显卡之间的物理距离，需要显卡在同一个超算中心中，这就导致去中心化的算力如果分布在世界各地就比较难实现。

第一点解释了为什么AMD和国内的华为、地平线目前很难追上；第二点解释了为什么去中心化训练很难实现。

3）去中心化算力的终局会是什么样？

- 去中心化算力目前难以进行大模型训练，核心在于大模型训练最看重的是稳定性，如果训练中断，需要重新训练，沉没成本很高。其对于多卡并联的要求是很高的，而带宽是受物理距离的限制的。英伟达通过NVLink来实现多卡通信，然而在一个超算中心里面，NVLink会限制显卡之间的物理距离，因此分散的算力无法形成一个算力集群去进行大模型训练。

- 但另一方面，对于算力要求相对较低的需求是可以实现的，例如AI推理，或者是一些特定场景的垂类中小型的模型训练是有可能实现的，在去中心化算力网络中有一些相对大的节点服务商时，是有潜力为这些相对大的算力需求服务的。以及像渲染这类边缘计算的场景也是相对容易实现的。

4）去中心化算法模型的终局会是什么样？

去中心化算法模型的终局取决于对未来AI的终局，我认为未来AI之战可能会是有1-2个闭源模型巨头（如ChatGPT），再加上百花齐放的模型，在这种背景下，应用层产品没有必要和一个大模型绑定，而是与多个大模型合作，在这种背景下来看，Bittensor的这种模式潜力还是非常大的。

[Kinza Finance @kinzafinance]

We’re excited! Kinza Finance was selected for a @binancelabs investment following MVB Season 6 🎉 Proud to be building the future of DeFi on @BNBCHAIN 💪

[CHRISTY CHEONG @christype2000]

Having a fantastic time at Token2049, where I've had the pleasure of meeting incredibly insightful individuals from various corners of the crypto world. The energy here is electrifying! ⚡️
@Davincij15 @gigiz_eth @Krystal_Eth
#Token2049
#singapore

[THUBA DAO @THUBA_DAO]

走进THUBA系列3⃣
带你认识THUBA优秀骨干丨办公室 X BD X 活动（2023Spring）

@NeilpyChen @PedroG9527 @EudemoniaCC @JellyZhouishere @christype2000 @cestfreya1

[CHRISTY CHEONG @christype2000]

把我的first 毕业pyq献给@THUBA_DAO 的小伙伴们
THUBA best association of all time!
真的很开心也很感谢能够在这最后的学期里认识到那么多THUBA的朋友们💓
@chenbeixugongzi @bufanthu123 @LunaLau8 @EudemoniaCC @The_Gus_God @Sean_miraclexh @wyeeeh @JellyZhouishere @AlbertXZZZZZ @cestfreya1

[CHRISTY CHEONG @christype2000]

A day with @THUBA_DAO @Edu_DAO #Hackathon
@LunaLau8 @EudemoniaCC @NeilpyChen @The_Gus_God @AlbertXZZZZZ

[CHRISTY CHEONG @christype2000]

Thrilled to be in the office and event track of @THUBA_DAO in my last semester. Looking forward to make contributions for the team and explore the boundless world of blockchain X Web3.0

[Tsinghua University @Tsinghua_Uni]

Writing #SpringFestival couplets, eating delicious food, attending family gatherings are all fun and festive traditions. What is your #ChineseNewYear like? Describe it using one word below to share #MyChinaStory.