跨国深谈

算力 (compute)

17分钟阅读

跨国深谈← 返回←算力 (compute)
算力

算力 (compute)

概念
本站收录 10 集 · 39 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

39 条

所以在推理计算方面，这基本上比当今最前沿的模型小了一百多倍。
So that’s like more than a hundred X smaller in terms of inference compute than basically the absolute frontier of models today.
—— Mustafa Suleyman · [10:28]

指向原始笔记的链接

现在，我主要受限于注意力。与 tokens 或计算不同，我不能简单地增加更多的注意力。
Now, I’m primarily constrained by attention. And unlike tokens or compute, I can’t simply add more of it.
—— 嘉宾 · [21:19]

指向原始笔记的链接

今天的算力需求远远超过算力供给。所以我们能把任何东西上线，都会立即被消耗掉。
Demand far outstrips compute supply today. So anything we can bring online, we consume immediately.
—— Sachin Katti · [00:07]

指向原始笔记的链接

我们把算力翻了三倍，收入也翻了三倍。
We tripled compute and we tripled revenue.
—— Sachin Katti · [15:53]

指向原始笔记的链接

现在，AI 本身可以做 AI 研究了。我们能做的实验数量爆炸式增长。因此，研究所需的算力也爆炸式增长。
Now, AI itself can do AI research. The number of experiments we can run explodes. And therefore, the amount of compute you need for research also explodes.
—— Sachin Katti · [16:50]

指向原始笔记的链接

任何时候我们觉得自己已经有足够的算力，可以慢下来了，总是会给我们带来负面的惊喜，比如，糟了，我们不该放慢速度的，对吧？
Any time we have thought we have enough compute, we can slow down. always negatively surprise us, like, oh, shit, we should not have slowed down, right?
—— Sachin Katti · [17:24]

指向原始笔记的链接

为了极其明确一点,AI 数据中心算力收入的绝大部分取决于两家不盈利、不可持续的 AI 公司是否有能力继续每年筹集数百亿或数千亿美元。
to be abundantly clear, the vast majority of AI data center compute revenue is contingent on the continued ability of two unprofitable, unsustainable AI companies to raise tens or hundreds of billions of dollars a year.
—— Alex · [35:38]

指向原始笔记的链接

你不可能为家里机器人工作 24 小时支付 100 美元的算力。
You cannot pay $100 of compute for 24 hours of your home robot to work.
—— Alex · [10:51]

指向原始笔记的链接

我们正在耗尽数据来训练大模型，但我们没有在使用旧模型的权重。所以为什么不使用权重，所有的知识，人们投入的所有算力，对吧？
We’re running out of data to train the large models, but we’re not using the weights of older models. So why not using the weights, all the knowledge, all the compute that people invested, right?
—— Damian Borth · [30:39]

指向原始笔记的链接

我们可能比以往任何时候都在从经验学习上花费最多的算力,但强化学习并不是从经验学习的终点。
We probably are spending the most compute than ever on learning from experience, but reinforcement learning is not the end of learning from experience.
—— Jerry Tworek · [26:22]

指向原始笔记的链接

我们刚刚向 FCC 申请了一个由 88,000 个这样的航天器组成的星座。所以这意味着大约 20 吉瓦的新计算能力。
We’ve just filed with the FCC for a constellation of 88,000 of those. So that means on the order of 20 gigawatts of new compute capacity.
—— Philip Johnston · [19:25]

指向原始笔记的链接

大多数做沙箱的人都在计算上赚钱。坦白说，在我看来这是个愚蠢的主意。
most sandbox people make money on compute. Frankly, that’s a dumb idea in my opinion.
—— Jerry Murdock · [27:19]

指向原始笔记的链接

Anthropic 和 OpenAI 明年将拿走多达 40% 到 50% 的算力。
You’ve got Anthropic OpenAI are taking as much as 40 to 50% of compute next year.
—— 嘉宾 · [04:17]

指向原始笔记的链接

因为今天任何人在每兆瓦 1000 万到 1500 万美元的算力上都能赚钱。
because anyone can make money off of 10 to $15 million per megawatt compute today.
—— 嘉宾 · [14:28]

指向原始笔记的链接

但可以比较容易地说，中国将拥有 30 吉瓦左右的 AI 算力或更少。
But it’s pretty easy to say China will have like 30 gigawatts of AI compute or less.
—— 嘉宾 · [35:37]

指向原始笔记的链接

所以为了计算内容拥有者，为了计算内容拥有者得到一美元，如果我决定进行完整的 Shapley 值计算，那可能要花费好几美元。
So to compute a content owner, to compute that a content owner gets a dollar, if I decide to do the full Shapley value computation, that might take several dollars.
—— Parag · [44:22]

指向原始笔记的链接

最值得一提的是，我创建了 Folding at Home 这个大规模超级计算机项目，它基本上让我们预览到了如今在 AI 中看到的许多问题，尤其是「苦涩的教训」——算力如此重要这一点。
Most notably, I built Folding at Home, the large scale supercomputer project, which basically allowed us to preview a lot of the issues that we now see in AI, especially the bitter lesson that compute is such a big deal.
—— Vijay Pande · [04:11]

指向原始笔记的链接

所以我们要看到的关键点是：智能体不是计算，它们是数据。
And so the key thing about agents that we see is that agents are not compute, they are data.
—— James · [05:52]

指向原始笔记的链接

当你把智能体看作逻辑实体时，你会意识到即使它们不在运行也存在，这就是为什么把它们建模为计算是完全错的，因为它们并不活在计算里。
And so when you think of agents as like logical entities, you realize that they exist even when they’re not running, which is why if you model them as compute, it’s just wrong because they don’t live in the compute.
—— James · [07:03]

指向原始笔记的链接

模型、计算，一切都是相对短暂的，几乎零切换成本。
Models, compute, everything is relatively ephemeral, almost zero switching costs.
—— Ofir Ehrlich · [03:24]

指向原始笔记的链接

AI 在数学上密集得多、算力密集得多，而所有那些基础设施——我们今天蹒跚维系的穷人版——需要为未来的 AI 时代被重新改造。
AI is way more mathematically intensive and compute intensive and all that infrastructure, sort of the poor man’s version that we’re limping along with today needs to be repurposed for the AI age going forward.
—— Jen Kha · [03:11]

指向原始笔记的链接

所以那个所谓的计算基础层，涨了，我不知道，一万倍。
So that, what was called the base layer of compute, went up, I don’t know, 10,000x.
—— Tony Kim · [19:55]

指向原始笔记的链接

而智能换算力基本上就等于收入。
And intelligence for compute equals basically revenue.
—— Tony Kim · [25:53]

指向原始笔记的链接

对算力的这种强烈需求至少还将持续12个月。
This intense demand for compute is going to continue for at least another 12 months.
—— 嘉宾 · [00:00]

指向原始笔记的链接

一个每年靠卖算力赚 1200 亿美元的人，决定收购一家能帮着让算力更具成本效益的公司，这样他就能卖出更多算力。
Man making $120 billion a year selling compute decides to buy a company that help makes compute more cost-effective so he can sell more compute.
—— 嘉宾 · [10:44]

指向原始笔记的链接

但一切都有瓶颈,我认为持续扩展这个东西的主要瓶颈实际上是训练算力。
But everything has a bottleneck, and I think the main bottleneck on continuing to scale this thing is actually training compute.
—— Justin Johnson · [23:04]

指向原始笔记的链接

盯着服务器看——那跟看着油漆干差不多刺激——然后说，天哪，我们解决了计算领域 75 年里没人解决过的问题。
Stared at a server, which is about as exciting as looking at paint dry and said, holy crap, we’ve solved this problem that nobody in 75 years of compute had ever solved.
—— Andrew Feldman · [12:39]

指向原始笔记的链接

如果每个人都想使用它，推理将会压垮计算基础设施。
If everybody wants to use it, inference is going to crush the compute infrastructure.
—— Andrew Feldman · [31:30]

指向原始笔记的链接

所以多智能体是一种以并行而非纯串行的方式扩展测试时算力的方法。
And so multi-agent is a way of scaling test time compute in parallel instead of purely serial.
—— Noam Brown · [01:49]

指向原始笔记的链接

今天没有任何人比那些要求被放慢速度的人建造更多的算力。
Nobody’s building more compute today than the people asking to be slowed down.
—— Jensen Huang · [55:08]

指向原始笔记的链接

这个范式令人兴奋的地方在于,大部分算力是在所有产品表面之间共享的,所以你不必为每个产品表面从头训练单独的模型。
The exciting thing about this paradigm is most of the compute is shared across all of the product surfaces, so you don’t have to train individual models from scratch for every product surface.
—— Devansh Tandon · [09:24]

指向原始笔记的链接

所以你本质上是在给网页搜索分配计算量，以便在模型上节省计算量。
So you’re essentially allocating compute to web search in order to save compute on the model.
—— Parag Agrawal · [07:36]

指向原始笔记的链接

所以我认为,所需要的算力,说实话,将比我们以为需要的要高出几个数量级。
So I think that the amount of compute that is going to be needed is going to be, honestly, orders of magnitude more than what we thought that we needed.
—— Noah Shinn · [64:00]

指向原始笔记的链接

所以它们把剩下的算力都用来琢磨如何编辑它们的转录记录，或者绕过评分器的这个约束。
And so they spend the rest of the compute trying to figure out how to edit their transcript or get around this constraint of the score.
—— Thariq Shihipar · [59:12]

指向原始笔记的链接

它们会因作弊而无法通过测试，所以它们把剩下的算力都用来想办法编辑它们的转写记录，或者绕过评分器的这个约束。
They will fail the test for cheating, and so they spend the rest of the compute trying to figure out how to edit their transcript or get around this constraint of the scorer.
—— Thariq Shihipar · [61:10]

指向原始笔记的链接

我认为人们低估了应用和应用开发者把算力容量转化为价值、转化为对经济有价值的服务的聪明才智。
So I think people underestimate the ingenuity of applications and application developers of turning compute capacity into value and into valuable services for the economy.
—— Chase Lochmiller · [40:34]

指向原始笔记的链接

但我相信，数据比算力重要得多。
But I believe that data matters a lot more than compute.
—— Diogo Almeida · [10:02]

指向原始笔记的链接

所以我们看到 Anthropic 现在在收入减去算力成本的意义上已经盈利了，训练和推理加在一起。
And so we’ve seen Anthropic is now profitable in terms of compute cost minus revenue, or revenue minus compute cost, both training and inference combined.
—— Dylan Patel · [13:10]

指向原始笔记的链接

蒸馏是一种有效地花更多钱来换取算力等价物的方式。
Distillation is a way to effectively spend more money to get what is compute equivalent.
—— Nathan Lambert · [10:00]

指向原始笔记的链接

② 出现在这些集

10 集

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · OpenAI · 推理 · Anthropic · NVIDIA · Meta · 数据中心 · 对齐 · ChatGPT · Google

反向链接