跨国深谈

令牌 (tokens)

15分钟阅读

跨国深谈← 返回←令牌 (tokens)
令牌

令牌 (tokens)

概念
本站收录 5 集 · 75 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

75 条

在 Resolve，我们现在正朝着每季度消耗一万亿 token 的方向前进。
So, here at Resolve, we are now on track to consume one trillion tokens per quarter.
—— Roshan · [17:34]

指向原始笔记的链接

所以他们只是挂了一个公开数字，像是我们每天服务 5.5 万亿个 token。
So they just had like the public number that was like, we serve 5.5 trillion tokens a day.
—— 嘉宾 · [07:01]

指向原始笔记的链接

市场正在给那些转售 token 的 SaaS 供应商增加溢价。是的。我认为那是不正确的。
The market is adding premium to SaaS vendors that are reselling tokens. Yes. And I think that’s incorrect.
—— Ivan Burazin · [61:44]

指向原始笔记的链接

我们不需要人员编制。我们需要 token 以便能够管理我们的运营。
We don’t need a headcount. We need tokens in order to be able to manage our operation.
—— Satya Nadella · [30:56]

指向原始笔记的链接

我让一个智能体去调试某个东西,它花了 500 美元,因为它决定读取大量日志文件并消耗大量 token。
I asked an agent to debug something and it spent $500 because it decided to read a lot of log files and burn a lot of tokens.
—— Matei Zaharia · [21:04]

指向原始笔记的链接

我通常寻找对这门学科的明显掌控，但还要有品味，能说：“嘿，你要拥有无限的 token，我们不能只是在制造垃圾。”
I generally look for obviously command over the discipline, but then the taste to say, “Hey, you’re going to have unlimited tokens and we can’t just be doing slop.”
—— Andrew Ambrosino · [31:20]

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

但下一层是关于，好，如果 token 并不是真正可互换的，你需要给它们分配不同的工作，比如这个 token 负责建议、那个负责执行，这个在「做梦」、那个在执行，诸如此类，你会想开始组合这些协同配合的编排式策略
But the next one is about, okay, if tokens aren’t really fungible and you need to give them different jobs, like maybe this token is advising versus this token is executing, this token is dreaming versus this token is executing, so on and so forth, you want to start composing these kind of orchestrated strategies that go together
—— 嘉宾 · [11:34]

指向原始笔记的链接

所以我认为把数据中心可视化最好的方式，就是巨大的工厂，把电子变成 token。
So I think the best way to visualize data centers is giant factories that are turning electrons into tokens.
—— Sachin Katti · [05:52]

指向原始笔记的链接

如今我们每天处理超过 40 万亿个 token。
Today we process more than 40 trillion tokens a day.
—— Lin Qiao · [37:01]

指向原始笔记的链接

这次重构几乎花了我 2 亿个 token。
This cost me almost 200 million tokens to refactor.
—— Heitor Lessa · [01:00]

指向原始笔记的链接

你可能不需要前沿的 token。你可能能够用前沿减一,或者你选择的开源权重 token 来应付。
you may not need frontier tokens. You may be able to get by with frontier minus one or your open weight token of choice.
—— Sriram Krishnan · [06:41]

指向原始笔记的链接

我认为，我的意思是，我们的一些轨迹可能是 500 万个 token，而整个《哈利·波特》系列是 200 万个 token。
I think, I mean, some of our trajectories can be 5 million tokens and the entire Harry Potter series is 2 million tokens.
—— Raj · [19:06]

指向原始笔记的链接

如果你现在愿意每年花 10 万美元在 token 上,你就是在过着 2028 年的人将要过的生活方式。
If you are willing to spend $100,000 a year right now in tokens, you are living the way somebody in 2028 is going to live.
—— Lenny · [20:11]

指向原始笔记的链接

就像 Jerry 说的,现在有公司在竞争发布周期,因为 tokens 不具粘性。
It’s like, as Jerry said, there are companies now competing for release cycles because tokens are not sticky.
—— 嘉宾 · [34:05]

指向原始笔记的链接

模型必须真正地推理并理解你的有害意图，并在数千个 token 中配合它。
The model has to really reason about and understand your harmful intention and go along with it for thousands of tokens.
—— Adam Gleave · [62:11]

指向原始笔记的链接

我认为在智能体循环和弄清楚如何开发系统方面仍然存在天文数字的机会,这些系统使你能够在持续反馈循环中花费 1000 倍或 100 万倍的 token 来推动结果。
I think there’s still just astronomical opportunity in agentic looping and figuring out how you develop systems that enable you to spend 1,000x more or 1,000,000x more on tokens to drive an outcome in a continuous feedback loop.
—— Alexandr Wang · [27:11]

指向原始笔记的链接

在物理世界，一个错误的代价可以用人类生命来衡量，而不是 tokens。
In the physical world, the cost of a mistake can be measured in human lives, not tokens.
—— Dmitri Dolgov · [04:11]

指向原始笔记的链接

但是他们在其上叠加的另一部分,也是开始变得真正有趣的地方,是他们开始奖励最少 token 路径。
But the other piece that they’ve layered on top, and this is where it starts to get really interesting, is they’ve started to reward the path of least tokens.
—— Dylan · [10:02]

指向原始笔记的链接

上下文溢出到主上下文，因为上下文意味着 tokens，tokens 意味着钱，而且你有越多的上下文，LLM 在给你答案时就会越困惑。
Context spill over into the main context because context means tokens, tokens mean money, and the more context you have, the more confused the LLM is gonna be in giving you an answer.
—— Frank Coyle · [13:11]

指向原始笔记的链接

所以走得更快意味着一个小小的错位可能滚雪球成大量浪费的工作,而且那工作消耗代币,代币现在要花真金白银。
So going faster means that a small misalignment can snowball into a ton of wasted work, and that work costs tokens, and tokens cost real money now.
—— Idan Gazit · [04:15]

指向原始笔记的链接

卖给你 token 的人的激励措施与你的并不真正一致。
The incentives of the people selling you tokens aren’t really aligned with yours.
—— Arjun Singh · [02:18]

指向原始笔记的链接

对于我们相对较小的团队,我们在过去一个月有 105 亿个 token
for our relatively small team, we had 10.5 billion tokens over the past month
—— Arjun Singh · [16:17]

指向原始笔记的链接

用于 AI 的 ASIC 能够以每秒 17,000 个 token 的速度运行 Llama 3.2。
The ASICs for AI were able to run Llama 3.2 at 17,000 tokens per second.
—— Paolo Ardoino · [24:06]

指向原始笔记的链接

这些代币是我们在现代经济中见过的贬值最快的资产之一,在性能恒定的情况下逐年持续下降 70% 到 80%,至少在过去四年里是这样,没有理由预期这种情况会改变。
These tokens are among the most rapidly depreciating assets we’ve ever seen in a modern economy that they’ve continually been falling 70 to 80 percent year over year on a constant performance basis for at least the last four years and there’s no reason to expect that to change
—— Paul Kedrosky · [13:13]

指向原始笔记的链接

在未来六年左右,你必须看到大约一亿倍的代币增长。
You have to see around 100 million-fold growth over the next six years in terms of tokens.
—— Paul Kedrosky · [20:38]

指向原始笔记的链接

它是一种公用事业。它将成为底层支撑一直发生的许多其他事情,我不再会知道谁提供我的代币,就像我不知道给我 MacBook 供电的电是来自哪个水电站一样。
It’s a utility. It will become underlying a host of other things that go on all the time, and I will no more know who provides my tokens than I do from which hydroelectric dam the power came from that’s powering my MacBook right now.
—— Paul Kedrosky · [66:37]

指向原始笔记的链接

这导致了一些非常令人惊讶的行为，因为实际上你离能够访问 10,000 个上下文桶只有不到两次 LLM 调用的距离，其中每一个都包含 200,000 个 token。
And that leads to some really surprising behaviors because you’re literally no more than two LLM calls away from being able to access 10,000 context buckets each one of which contains 200,000 tokens.
—— Flo Crivello · [32:49]

指向原始笔记的链接

你的智能体可以持有令牌，并在使用点上使用这些令牌访问数据，而不允许数据在系统中随意流动。
Your agents can bear tokens and use those tokens to access the data at the point of use and not allow data to flow around the system as it likes.
—— Saul Howard · [12:03]

指向原始笔记的链接

有图谱和没有图谱相比，我们在 token、轮次和延迟上看到了巨大的改进。
With and without graph, we see massive improvement in tokens, turns, and latency.
—— Uday Kiran Medisetty · [10:08]

指向原始笔记的链接

大多数的 token,压倒性多数的美国企业界的 token,不会跑在前沿模型上，而是会在那些企业内部。
The majority of tokens, the overwhelming majority of tokens in corporate America. will not be on the frontier models, it will be inside those enterprises.
—— 嘉宾 · [35:30]

指向原始笔记的链接

如果你看看一个每个 token 成本为双位数的美元的前沿模型，再看看一个每 token 成本为 10、11 美分的开源模型，虽然不是所有的 token 都生而平等，但这仍然是一个巨大的差异，足以带来大规模的采用。
if you’re looking at a frontier model with double dollar digit cost per token, and you’re looking at an open source model that’s 10, 11 cents per token, while all tokens aren’t created equal, it’s still enough of a difference that there’s going to be a massive adoption.
—— Jerry Murdock · [14:06]

指向原始笔记的链接

但是对于 token 和 GPU 的需求是如此无限，像字面上你可以决定投入多少钱来驱动漏斗顶部和增长。
But the demand is so unlimited for tokens and for GPUs, like literally you can just decide how much money you’re putting into it in order to drive top of funnel and growth.
—— Martin Casado · [47:51]

指向原始笔记的链接

我的工作就是把 token 做到人类可能做到的最低价。
My job is to make the tokens as cheap as humanly possible.
—— Neil Movva · [69:39]

指向原始笔记的链接

智能需要在所有这些事情背后运行,许多不同的系统需要被连接起来,所有那些 token 都要花钱。
Intelligence needs to be run behind all of these things, and many different systems need to be connected, and all those tokens cost money.
—— Cliff Obrecht · [04:13]

指向原始笔记的链接

如果你使用 parallel 搜索，你在很大程度上会在你的智能体中使用不到一半的 token。
If you use parallel search, you will, for the most part, use under half the tokens in your agent.
—— Parag · [13:49]

指向原始笔记的链接

当我们与我们的一些较大的客户交谈时，几乎有一个 token 的黑市。
When we speak to some of our larger customers, there’s almost a black market for tokens.
—— Elaina O’Mahoney · [22:57]

指向原始笔记的链接

如果他们使用自己的 token，他们将不得不去找其他三个员工说，“我的原型完成了 80%，我的 token 被切断了。Susie，我可以虹吸你的一些 token 吗？“
If they use their own tokens, they’ll have to go to three other employees to say, “I’ve got 80% of the way there with my prototype, and my tokens got cut off. Susie, can I siphon some of your tokens?”
—— Elaina O’Mahoney · [23:03]

指向原始笔记的链接

他们发现，在每月 200 美元的 ChatGPT 订阅上，你可以烧掉价值 14,000 美元的 token。
They found that on a 14,000 worth of tokens.
—— 嘉宾 · [12:40]

指向原始笔记的链接

OpenAI 有一项研究出来了，我想，大约一周前，说在 AI 令牌支出和每位员工收入之间没有联系。
OpenAI had a study that came out, I think, like a week ago, that said there was no connection between spending on AI tokens and revenue per employee.
—— 嘉宾 · [65:51]

指向原始笔记的链接

token 减少 50%,分类更快，答案更好。而这正是上下文引擎能做到的。
50% fewer tokens, faster triage, better answers. And that’s exactly what a context engine can do.
—— Peter Werry · [17:45]

指向原始笔记的链接

对消费者来说，这是一个糟糕的生意，以200美元出售价值10,000美元的tokens是我们这个时代最糟糕的商业模式之一，对吧？
It’s a crappy business that consumers, selling 200 is one of the worst business models of our lifetimes, right?
—— 嘉宾 · [39:01]

指向原始笔记的链接

这一切在很大程度上是由那些大型 AI 厂商驱动的，我认为他们的利益就在于让我们所有人都花更多的 token。
This is all driven by kind of the big AI vendors who I think their interest is for us all to spend more tokens.
—— Roberto Milev · [18:12]

指向原始笔记的链接

智能体使用的 token 量是人类的五倍。而且我们刚刚跨越到智能体的流量超过人类的时刻，也就是现在互联网上智能体的数量比人类还多。
Agents are using five times the amount of tokens as humans are. And we just crossed over to agents being more on the volume of agents are now more on the internet than humans are.
—— Jen Kha · [18:42]

指向原始笔记的链接

有几项学术研究表明，最开头的 7K 和最末尾的 7K token 是最重要的，而中间的那些东西最终只会在 LLM 试图想出答案时把它搅浑、让它困惑。
There’s a couple academic studies that show that like the first 7K and like the last 7K tokens are the most important and the stuff that’s in between can end up just muddying and confusing the LLM as it’s trying to come up with an answer.
—— 嘉宾 · [56:56]

指向原始笔记的链接

总得有什么东西来负责编排、仲裁、决定这些 token 去哪里。这就是 CPU 所做的事。
Something has to do the orchestration, arbitration, decision around where those tokens go. That’s what CPUs do.
—— Rene Haas · [00:09]

指向原始笔记的链接

绝大多数的 token——这是我们的看法——在企业内部会是开源模型，比如说 80%、90%。
The super majority of tokens, and this is our take, it will be open models within a business, call it 80, 90%.
—— Jeffrey Morgan · [18:19]

指向原始笔记的链接

你必须预期 token 的成本会随着时间越来越趋近于基础设施的成本。
you have to expect that the cost of tokens converge closer and closer to the cost of the infrastructure over time.
—— Aaron Levie · [12:57]

指向原始笔记的链接

我们希望随着时间的推移，一百万 token 在智能水平、成本、延迟等方面，感觉起来像五万 token。
We want a million tokens to feel like 50,000 tokens in terms of intelligence, cost, latency, et cetera, over time.
—— Alexander Whedon · [31:20]

指向原始笔记的链接

为此，我们真的希望人们觉得输入 token 是免费的——只管为你的问题考虑所需要的上下文就好。
To make it, we really want people to feel like the input tokens are free, like just consider the context that you need to for your problem.
—— Alexander Whedon · [45:32]

指向原始笔记的链接

你所有的创新代币都花在开源项目上了，而不是花在交付上。
All your innovation tokens are being spent on that versus on the delivery.
—— Jordan Tigani · [11:04]

指向原始笔记的链接

你所有的创新代币都花在构建开源项目上，而不是花在交付上。
All your innovation tokens are being spent on that versus on the delivery.
—— Jordan Tigani · [10:56]

指向原始笔记的链接

token 减少 50%,分诊更快,而且答案质量实际上更好,因为它知道业务内部正在发生什么。
50% fewer tokens, faster triage, and the answer quality is actually better because it knew what was going on inside of the business.
—— Brandon Waselnuk · [09:51]

指向原始笔记的链接

我认为提醒自己这一点很重要：不管间接还是直接，我们很可能都在售卖 token。
I think it’s important to remind ourselves that in some way or another, we probably are selling tokens, whether it’s indirectly or directly.
—— Maximillian Piras · [05:03]

指向原始笔记的链接

token 作为内部系统的度量当然是有用的，但归根结底，它们只是一种输出。
tokens are, of course, useful as a measurement of an internal system, but at the end of the day, they’re just an output.
—— Maximillian Piras · [10:49]

指向原始笔记的链接

是由 token 及等价物驱动的，而 RAMP 将成为人们管理 token 支出、优化支出、衡量 ROI 的主要方式，这大概会在很多方面让我们与 Anthropic 相撞。
of the world is driven through tokens and equivalent, and RAMP’s going to be the primary way people manage their token spend, how they optimize it, how they measure ROI, and that will lead us to collide probably with Anthropic in many ways.
—— Keith Rabois · [17:22]

指向原始笔记的链接

而且最终，比如五年后，我敢打赌企业里 90% 的 token 都会是用户从未启动过的东西，他们只是看到一个结果。
And ultimately, like in five years from now, I would bet like 90% of all tokens in the enterprise are things that a user never kicked off and they just see a result.
—— Aaron Levie · [47:46]

指向原始笔记的链接

我们从需要大约 150,000 个 token 来给我们的智能体提供所有需要的工具,降到了少于 1,000 个 token 来指定它们从 Nexus 获取信息所需的全部接口。
We went down from having something like 150,000 tokens to give our agents all the tooling they need to less than 1,000 tokens to specify all the interfaces they need to get information out of Nexus.
—— Edo Liberty · [17:07]

指向原始笔记的链接

如果你能影响或在那些 token 里留下印记，那价值比出现在单个上下文窗口里高出数量级。
if you can get influence or imprint in those tokens, it’s worth, like, orders of magnitude more than being in a in a in single context window.
—— Ben Moore · [52:00]

指向原始笔记的链接

因为它是 diffusion，因为你同时输出很多 token，我们能够比自回归模型快大约 10 倍地生成文本。
Because it’s diffusion, because you’re outputting many tokens at the same time, we were able to generate text like 10x faster compared to the autoregressive model.
—— Stefano Ermon · [05:36]

指向原始笔记的链接

他们字面意义上不收输出 token 的费用。
They literally don’t charge for output tokens.
—— Ranjan Roy · [63:33]

指向原始笔记的链接

贯穿这些整个会话的所有 token 里，大约有 96% 是被缓存的
about 96% of all the tokens that go through these entire sessions are cached.
—— Thomas Sohmers · [39:29]

指向原始笔记的链接

通过运行这些实验，你很快会看到在几万 token 的量级上出现极其陡峭的性能下滑。
And by running these experiments, what you quickly see is a super sharp drop off in the tens of thousands of tokens.
—— Hamad Bashir · [13:38]

指向原始笔记的链接

一旦你使用越来越多的工具，实际上省得就越多，它真的能扩展，你可以节省 50% 的 token 甚至更多。
Once you use more and more tools, the more actually you save, it really scales and you can save 50% of tokens or more.
—— Tereza Tížková · [17:12]

指向原始笔记的链接

我们不认为以今天这种训练方式训练出来的系统，会思考十亿或万亿个 token 后突然直观地领悟到：这里有一个新的高温超导体。
And we don’t think that a system trained in the way that they’re trained today will think for a billion or a trillion tokens and suddenly intuit, here’s a new high-temp superconductor.
—— Doj Chubuk · [05:16]

指向原始笔记的链接

有人让我衡量 token 来看我们用了多少 AI，我说不，因为它太容易被操纵了。
Someone asked me to measure tokens to see how much AI we do, and I said no, because it’s so easily hackable.
—— Rania Khalaf · [12:55]

指向原始笔记的链接

现在，人们愿意为前沿 token 支付 10 倍的价格。
Right now, people are willing to pay 10x more for Frontier tokens.
—— David George · [11:22]

指向原始笔记的链接

同样一个五个字符的文件，最贵的模型消耗了 70 倍的 token。
The same five character file took the most expensive model 70x more tokens.
—— Harald Kirschner · [17:18]

指向原始笔记的链接

我们最终卖给客户并从中赚钱的产品其实有三种:我们可以卖数据中心,可以卖 GPU,还可以卖 token。
There’s really like three products that we ultimately sell to customers where we’re making money. We can sell data centers, we can sell GPUs, and we can sell tokens.
—— Chase Lochmiller · [00:23]

指向原始笔记的链接

人们花在闭源前沿模型上的钱确实比开源多,但他们在开源上生成的 token 比闭源多。
I do think that people are spending more money on closed source frontier models than they are on open source, but they are generating more tokens on open source than closed source.
—— Chase Lochmiller · [49:15]

指向原始笔记的链接

所以这已经成为让你能够在 NVIDIA 硬件上以每秒约 1,000 个 token 运行相当大的模型的关键技术之一，而不必去用像 LPU 或 Cerebras 这样的新型加速器。
So this has emerged as one of the key techniques to allow you to operate, say, 1,000 tokens per second on NVIDIA hardware for actually fairly large models without having to go to a novel accelerator like an LPU or Cerebris.
—— Charles Frye · [53:06]

指向原始笔记的链接

你必须每六个月把那硬件生成 token 的能力翻一倍。而其中来自软件的部分会和来自硬件的部分一样多，甚至更多。
You have to double the capability of that hardware to generate tokens every six months. And as much or more of that is going to come from software as it is from hardware.
—— Amin Vahdat · [13:33]

指向原始笔记的链接

AI 所做的，就是它会超级加速劳动力份额与资本份额之间的这个长期趋势，因为 AI token 基本上只是以芯片和数据中心形式存在的资本，它们坐在那里创造着劳动力。
What AI does is it’s going to supercharge that long-term trend of labor share versus capital share because AI tokens are basically just capital in the form of chips and data centers that are sitting there and creating labor.
—— Dylan Patel · [17:13]

指向原始笔记的链接

基本上，你要衡量的是结果，而不是 token。
Basically, you want to measure outcomes, not tokens.
—— Davis Palmie · [16:31]

指向原始笔记的链接

所以一百万 token 或两百万 token 的上下文，在性能方面不会帮到你。
So one million tokens or two million tokens, context will not help you in terms of the performance.
—— Vlad Luzin · [03:56]

指向原始笔记的链接

② 出现在这些集

5 集

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · Stripe · OpenAI · Claude · Vercel · 推理 · Notion · Cursor · Claude Code · Dylan

反向链接