集里怎么说它

• 《当系统出故障时，让 AI 代替作战室里的 50 个人——Traversal 谈如何造 AI SRE》(03:49起):被描述为选大模型时不得不考虑的现实经济约束;以前只看性能,随着智能体跑一次动辄 500 万 token,成本变得越来越重要。
• 《AI没有消灭客服，反而让他们更累了？》(10:49起):Dan 总结模型选型的框架：一切都过三道约束——控制成本、交付卓越体验（质量）、控制延迟，三者通常只能取其二；所以确定性规则和传统 ML 干简单活，前沿模型只留给复杂任务。

① 提到它的金句

100 条

它成为了墨西哥和许多其他国家的最畅销应用。它不是一个免费应用，而且它完全是本地的，所以没有后端，没有成本，没有广告。所以这是一个非常盈利的尝试。
It became the top-selling app in Mexico and in a lot of other different countries. It was not a free app, and it was also completely local, so there was no back-end, no cost, no advertising. So it was a very profitable endeavor.
—— Sebastian Barrios · [51:23]

指向原始笔记的链接

我们在行业内看到，通过算法数据和效率改进的结合，给定智能水平的成本降低了 10 倍。
We’ve seen in the industry a 10X decrease in cost for a given amount of intelligence through a combination of algorithmic data and efficiency improvements.
—— Benjamin Mann · [58:33]

指向原始笔记的链接

什么仍然困难到如果你改进它三到五倍，也许10倍，或者你从中拿走成本，比如10倍成本，十分之一，他们会说，“这太令人信服了，你现在应该拿走我的钱。“
What is still difficult to the point that if you improve it three to five times, maybe 10X, or you take cost out of it, like 10X costs, one 10th, they’ll be like, “This is so compelling, you should take my money now.”
—— Oji · [11:57]

指向原始笔记的链接

在 AI 时代，生产软件的成本正在趋近于零。
In the AI era, the cost of producing software is going to zero.
—— John · [07:04]

指向原始笔记的链接

这可能看起来不直观，但追求一个极其雄心勃勃的创业想法的成本和追求一个温和的想法的成本大致相同。
It may seem unintuitive, but the cost of pursuing a wildly ambitious startup idea and the cost of pursuing a modest one are roughly the same.
—— John · [08:27]

指向原始笔记的链接

他们从未如此兴奋于不用管理一个团队，因为他们意识到其中很多只是开销。其中很多只是协调成本。
They’ve never been more excited to not have a team to have to manage because they realize that a lot of that is just overhead. A lot of that’s just coordination cost.
—— Jiaona Zhang · [56:47]

指向原始笔记的链接

但我们在这个行业有着非常糟糕的沉没成本心态。
But we have such a bad sunk cost mindset in this industry.
—— Theo Browne · [08:44]

指向原始笔记的链接

如果你让我在多 20% 的质量和少 20% 的成本之间选，我绝对会选质量。
If you let me have 20% more quality, 20% less cost, I’ll definitely go for the quality.
—— Avishai Abrahami · [17:58]

指向原始笔记的链接

嗯，我认为这对我们来说是一个领先指标，过去六七个月我们的 Anthropic 成本每月增长 6 倍。
So, I mean, this is one of the, I think, the kind of leading indicators for us is that our Anthropic cost over the last six or seven months has 6X’d every month
—— Fred Turner · [50:02]

指向原始笔记的链接

我确实认为 token 的成本会大幅下降，未来三年内成本降低 10 倍，而这 10 倍的成本降低将驱动 100 倍的使用量。
I do think the cost of token will go down drastically, 10x cost reduction in the next three years, and this 10x cost reduction will drive 100x usage.
—— Lin Qiao · [00:12]

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

我还想论证的是，对于很多人和很多公司来说，即使模型是免费的，它仍然会让你花费更多。
I would also argue that for a lot of people and for a lot of companies, even if the model is free, it’ll still cost you more.
—— Jensen Wong · [12:00]

指向原始笔记的链接

即使你可以，一次调查可能会让你花费一个小国的 GDP。
Even if you could, a single investigation is probably going to cost you the GDP of a small country.
—— Anish · [03:49]

指向原始笔记的链接

大致来说，同样的智能水平，成本每年下降约 10 倍。
Like it’s a 10x cost reduction per year, roughly speaking, for the same intelligence.
—— 嘉宾 · [00:45]

指向原始笔记的链接

其实，我昨天跟团队说，这 12 亿的真正成本不是股权稀释，真正的成本是期望。
Actually, I told the team yesterday that the real cost of this 1.2 billion was not dilution. The real cost is expectations.
—— Alex · [22:23]

指向原始笔记的链接

它将把创建内容的边际成本驱动到零。
It’s going to drive the marginal cost of creating content to zero.
—— Victor Riparbelli · [03:50]

指向原始笔记的链接

关键在于让硬件开发看起来更像软件开发，无论在比特的世界还是原子的世界里，都要降低迭代的成本。
The key thing is to make hardware development look more like software development, to reduce the cost of iteration, both in the world of bits and in the world of atoms.
—— Blake Scholl · [12:34]

指向原始笔记的链接

所以随着软件工程成本的下降，我们对软件工程师的需求实际上上升了。
So our need for software engineers has actually gone up as the cost of software engineering has gone down.
—— Blake Scholl · [41:51]

指向原始笔记的链接

关键是让硬件开发看起来更像软件开发,降低迭代成本,无论是在比特的世界还是原子的世界里。
The key thing is to make hardware development look more like software development, to reduce the cost of iteration, both in the world of bits and in the world of atoms.
—— Blake Scholl · [12:34]

指向原始笔记的链接

AI 所做的是降低构建定制软件的成本,我认为这允许更多的企业拥有完美适合他们的软件。
What AI does is reduce the cost of building custom software, which I think allows more businesses to have software that fits them perfectly.
—— Blake Scholl · [25:23]

指向原始笔记的链接

将组织转变为使用 AI 的现代工作方式的转型成本是，你可能会花费相当于每个人两年薪水的总和。
The transformation cost to transform an organization into modern ways of working with AI is you are likely going to spend the equivalent of everyone’s salary over a two-year period.
—— Chris Blackburn · [00:34]

指向原始笔记的链接

但回到这个问题，粗略估计就是把你组织的全部薪酬总成本拿来，你要把它除以二，这就是你接下来几年肯定要花的。
But going back to the question, wet finger in the air is take the entire total cost of compensation for your organization and you’re going to run that, divide that in half, and that’s what you’re going to spend for a couple of years in a row for sure.
—— Chris Blackburn · [55:29]

指向原始笔记的链接

如果我们只把 AI 用于削减成本，那将非常令人难过。
It would be pretty sad if we used AI only for cost-cutting.
—— Melisa Tokmak · [31:09]

指向原始笔记的链接

这个模型以前的运行成本比现在高大约 100 倍。
This model used to cost about 100 times more to run than it does now.
—— Joon Sung Park · [21:31]

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

因为我们实际上从一个大型机构那里得到了关于做 StarCloud 1 成本的报价，他们说7500万到1亿美元。而我们做了整个 StarCloud 2，包括发射，只花了200万美元。
Because we actually had a quote from one of the primes on what it would cost to do StarCloud 1, and they said $75millionto$100 million. And we did the whole of StarCloud 2, including the launch, for $2 million.
—— Philip Johnston · [06:07]

指向原始笔记的链接

所以我们的散热器每瓦散热的重量至少比国际空间站的散热器少 10 倍。每瓦散热的成本比国际空间站的散热器少 500 倍。
So our radiator is at least 10 times less mass per watt of dissipation than the ISS radiator. and 500 times less cost per watt of dissipation than the ISS radiator.
—— Philip Johnston · [11:47]

指向原始笔记的链接

我认为还因为像 Cursor 所做的事情，人们开始说，哦，等等，我可以坐下来使用基础模型，但随着这些人试图 IPO 并拥有更好的现金流，它们可能只会增加成本。
I think also because of things like what Cursor did, people are starting to say, oh, wait, I could sit back and use the foundation models, but they’re probably only going to go up in cost as these guys try to IPO and have better cash flows.
—— Ian · [06:27]

指向原始笔记的链接

如果你在 Opus 和 Haiku 上运行 terminal bench,Opus 的表现会好大约三倍,成本却是 Haiku 的十分之一,尽管 Haiku 每个 token 的价格要便宜得多。
If you run terminal bench on Opus and Haiku, Opus will do about three times better at one-tenth the cost of Haiku, even though Haiku is significantly cheaper per token.
—— 嘉宾 · [15:26]

指向原始笔记的链接

即使你是一个拥有十亿代币的人,即使你有 10 个终端没日没夜地运行 Fable,机会成本仍然存在,它就是一切。
Even if you’re a token billionaire, even if you have 10 terminals running Fable night and day, then opportunity cost is still there, it’s everything.
—— Idan Gazit · [01:34]

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

所以 200 美元的订阅成本将从 1 美元到 5,000 美元。这就是现实。
So a $200subscriptionwillcostfromoneto$5,000. That’s the reality.
—— Paolo Ardoino · [43:20]

指向原始笔记的链接

你某种程度上是在押注失败，而押注失败与押注乐观的成本相比，那里的差别真的很大。
You’re sort of betting on failure and the cost of betting on failure versus betting on optimism is a real big difference there.
—— Kevin Scott · [25:21]

指向原始笔记的链接

这就是为什么履约是极少数随着规模扩大、单位成本实际上变得更贵的事情之一。
And that’s why fulfillment is one of the only things that as you scale, it actually gets more expensive on a per unit cost.
—— Kevin Gibbon · [17:08]

指向原始笔记的链接

所以我们不仅想让人们检查眼睛、拿到眼镜处方变得难以置信地容易，还想把成本降得足够低，以至于你真的没有任何理由不做这件事。
So we want to not only make it unbelievably easy for people to get their eyes checked and to get an eyeglass prescription, but also we want to lower the cost so much that you really have no reason to not do this.
—— 嘉宾 · [37:22]

指向原始笔记的链接

当你成为一家上市公司，你已经出去征服了你的市场，没有人会回头看说，天哪，六年前你的销售成本真的很糟糕。
When you’re a public company and you’ve gone out and you’ve conquered your market, nobody looks back and says, gosh, six years ago, your cost of sales was really terrible.
—— Andy McCall · [42:44]

指向原始笔记的链接

是的，很多人都在谈论围绕这些东西的护栏，但我认为从基础设施角度，成本方面也需要护栏，因为你不会想要不小心启动了什么东西，结果突然收到一笔巨额账单，只因为某个智能体做了一个不一定正确的决定。
Yeah, I mean, a lot of people talk about guardrails around these things, but I think from infrastructure, guardrails around the cost as well, because you don’t want to accidentally spin up something that suddenly you have a huge bill for because some agent made a decision that was not necessarily the right decision.
—— Sean Falconer · [22:56]

指向原始笔记的链接

我们模式的美妙之处在于，我们主要是一个可变成本模式。这意味着当需求下降时，很好，我们的成本随之下降。但这也意味着我们确实没有固定的产能可以送出去。
The beauty of our model is we’re primarily a variable cost model. It means when demand drops, great, our costs scale down with it. But it also means that we just don’t have a fixed capacity to give away.
—— Andrew MacDonald · [11:46]

指向原始笔记的链接

但如果你看看这背后的代价，那是社会里数万亿美元的无谓损失。
But if you were to look at the cost of that, that’s trillions of dollars of dead weight lost in society.
—— Farooq Malik · [18:23]

指向原始笔记的链接

一切都必须是掠夺性的吗？不。你能不能提供一个有竞争力、成本更低、还能赚更多钱的产品？能。
Does everything have to be predatory? No. Can you provide a product that is competitive, lower cost, and make more money? Yes.
—— Farooq Malik · [15:23]

指向原始笔记的链接

因为你会意识到，任何拆解人形机器人的人，几乎所有的成本都在执行器上。
Because you do realize anybody that’s breaking down a humanoid robot, almost all the cost is in the actuators.
—— Mike LeBlanc · [18:19]

指向原始笔记的链接

每个 AI 都能到达生产环境，只是它没能跑起来，或者没能证明自身的成本合理。
Every AI reaches production, but it just fails to work or justify its own cost.
—— Ayush Bhardwaj · [16:37]

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

傲慢可能不是你喜欢的东西，但它可能会让他们在所做的事情上非常出色。这可能是他们尖刺的代价。
And arrogance might not be something you like, but it might make them very good at what they do. It’s maybe the cost of their spike.
—— Julien Bek · [35:40]

指向原始笔记的链接

对我们来说,在成本、竞争、供应链、制造专业能力、车辆知识产权上,业界最好的真的是 BYD。
The best in business for us, cost-wise and competition-wise, supply chain, manufacturing expertise, the IP in the vehicle was really BYD.
—— Jim Farley · [12:25]

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

你可能能做到质量,但成本和延迟你会吃亏,只有建立自己的研究团队、训练自己的模型才行。
You can probably get the quality, but the cost and latency you’re going to lose out on, and only by building your own research team and training your own models.
—— Cliff Obrecht · [16:39]

指向原始笔记的链接

所以一切对我来说，当你对信息和各种工作的相关信息有无限的需求时，其核心是一个围绕质量、成本和延迟的优化问题。
So everything, to me, when you have an infinite appetite for information and relevant information for all kinds of work, it’s at its core an optimization problem around quality, cost, and latency.
—— Parag · [14:22]

指向原始笔记的链接

另一件我要反复强调的事是：不惜一切代价避免试点。
And the other thing I’ll hammer home is to avoid pilots at all cost.
—— Aliisa Rosenthal · [06:48]

指向原始笔记的链接

盲目重试只会成倍增加你的成本和你的尾延迟。
So blind retries just multiply your cost and your tail latencies.
—— Kanish Manuja · [02:54]

指向原始笔记的链接

我从未完成我的硕士论文，因为不去构建的机会成本已经变得太大了。
I never finished my master thesis because the opportunity cost of not building had become too large.
—— Max Junestrand · [05:18]

指向原始笔记的链接

在 Base 上的平均结算时间是 200 毫秒，每笔交易的成本约为十分之一美分。
The average settlement time is 200 milliseconds on base with about a tenth of a cent as cost per transaction.
—— Anil Nadiminti · [19:13]

指向原始笔记的链接

而我认为，直到今天 Affirm 最酷的事情之一就是它拥有负的客户获取成本。我是说，它是被付钱来获取客户的。
And I think, to me, one of the coolest things about a firm to this day is it has negative customer acquisition cost. I mean, it is paid to acquire a customer.
—— Alex Rampell · [44:52]

指向原始笔记的链接

随着时间推移，这实际上会让你花钱，因为每一轮都在喂入所有这些来自失败尝试和出错轮次的额外上下文。
That is actually going to cost you money over time because every turn is feeding all this extra context from failed attempts, turns gone wrong.
—— Santi Garza · [28:12]

指向原始笔记的链接

这些都是你只能在 Cursor 得到的模型，这些云端智能体能以非常快速高效的方式使用它们，而且一次完整的迁移可能只花你三四美元。
These are all models you can only get in Cursor because these cloud agents are able to use them in a really fast, efficient way and only cost you maybe like three or four dollars for an entire migration.
—— Amrita · [31:00]

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

成本是到目前为止最大的痛点，开放模型可以介入并解决。
Cost is by far the largest pain point that open models can jump in and solve.
—— Jeffrey Morgan · [02:18]

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

如今提供一项真正的、医学上可信的服务的成本结构，比过去低了大约一百倍。
The cost structure to deliver an actual like medically credible service is like a hundred times lower than it used to be.
—— Julie Yoo · [21:26]

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

而如果你在现有范式中投入了大量沉没成本，从零开始创造一个新东西是非常痛苦的。
And if you have a lot of sunk cost in that existing paradigm, it’s very painful to create a new thing from scratch.
—— Roman Ugarte · [35:05]

指向原始笔记的链接

那将花费 1.2 万亿美元——前提是我们国家首先还执行得了那个计划。
That’ll cost $1.2 trillion if we can even execute as a country in the first place on that plan.
—— 嘉宾 · [04:05]

指向原始笔记的链接

它带来的后果是，生成的成本现在基本上趋于零。
It comes with consequences of suddenly now the cost of generation is basically going to zero.
—— Thais Castello Branco · [03:49]

指向原始笔记的链接

用量会增加，所以 5% 会变成 50%，但成本会下降。
The usage will increase, and so the 5% will become 50%, but the cost will go down.
—— Matteo Franceschetti · [21:00]

指向原始笔记的链接

我认为机会在于,造东西的成本变得如此之低,以至于我们过去需要花费多得多的成本才能想象出来的东西,现在做起来太容易了。
I think the opportunity is like the cost of making things become so low that we can imagine something that cost a lot more in the past now.
—— Peng Zheng · [17:10]

指向原始笔记的链接

我觉得我们现在的新建议是：为你的智能体进行 7×24 小时不间断的红队测试，尤其是当你认为不良交互的代价可能相当大的时候。
And I think our new recommendation now is to run 24-7 red teaming for your agents, especially if you believe the cost of bad interactions can be rather large.
—— Sumanyu Sharma · [14:26]

指向原始笔记的链接

我们的主要原则是,我们基于利害关系而非成本来路由。
And our main principle is that we route based on stakes, not on cost.
—— Paula Dozsa · [06:08]

指向原始笔记的链接

你会获得显著的成本节约,77%、90% 的 token 减少很常见。
You get significant cost savings, 77, 90% token reduction is pretty common.
—— Edo Liberty · [18:44]

指向原始笔记的链接

问题是那样每次搜索要花，比如，1000 万美元。
The problem is that would cost, like, $10 million per search.
—— Will Bryk · [06:24]

指向原始笔记的链接

比如用同一个模型但不同的 harness,成本差异几乎有 2 倍。
Like if you use the same model but different harnesses, there’s almost 2x different cost difference.
—— 嘉宾 · [56:13]

指向原始笔记的链接

开放权重模型让你以几分之一的成本获得前沿质量的智能体。
Openweight models give you frontier quality agents at a fraction of the cost.
—— Byung-Gon (Gon) Chun · [03:05]

指向原始笔记的链接

处理一个缓存命中的 token 基本是免费的。它的成本是真正重新计算并生成那个 token 的千分之一量级
both processing a cash token is essentially free. It’s one one thousandth of the cost, you know, order of magnitude of actually having to recompute and generate that token.
—— Thomas Sohmers · [11:25]

指向原始笔记的链接

对于企业工作,我的预测是 90% 的流量会流向非常划算的模型。
For enterprise work, my prediction is that 90% of the flow will go to very cost efficient models.
—— Daniel Dines · [46:38]

指向原始笔记的链接

因为它的 PR 多得多,所以花的钱也多得多,所谓多得多,我是指整整 9 美分。
Because it has a lot more PRs, it cost me a lot more money, and by a lot more money, I mean nine whole cents.
—— 嘉宾 · [09:33]

指向原始笔记的链接

要是在 20 年前,这东西通常得花 50 万美元才能开发出来。
Now, this would normally have cost 20 years ago, $500,000 to build.
—— 嘉宾 · [01:16]

指向原始笔记的链接

这不是关于降低成本，而是关于拓展可能性。
It’s not about cost reduction. It’s about possibility expansion.
—— Anish Acharya · [38:19]

指向原始笔记的链接

而从长远来看，真正会被商品化的反而是这些增长极快的 AI 公司和机器人技术，因为在公司之间、甚至在个人层面复制它们的内部成本会变得如此之低。
And long term, it’s actually these super fast growth AI companies and robotics that end up commoditized by how low the cost becomes to replicate them internally across companies and for individuals even.
—— Sean · [13:47]

指向原始笔记的链接

而这个业务的飞轮在于，你服务的量越大，成本就越低、速度就越快、信任度就越高。
And the flywheel of this business is that the more volume you serve, the lower cost, the faster speed, and the greater trust.
—— Sean · [09:04]

指向原始笔记的链接

而如果你认为今天你的钱的最好和最高用途是优化你的成本结构，那这对你的收入机会说明了什么？
And if you think that the best and highest use of your dollars today is to optimize your cost structure, what does that say about the revenue opportunity for you?
—— David George · [25:46]

指向原始笔记的链接

当你从一个模型切换到另一个模型时,这可以把你的成本削减 10 倍、一个数量级。
And when you switch from one model to another, that can cut your cost by 10, by a factor of 10.
—— Christian Klein · [54:00]

指向原始笔记的链接

所以如果你只关注成本端，就像是在做空自己的未来；而如果你关注前端，你就是在以更广阔的视野看待自己能成为什么。
So it’s like shorting your own future if you are just focused on the cost side, whereas you’re taking an expansive view of what you can become if you’re focused on the front side.
—— David George · [09:42]

指向原始笔记的链接

因为成本只有几千美元的拦截机,能够有效地摧毁成本 5 万美元的 Shahid。
Because investors that cost a few thousands effectively destroy shahids that cost 50,000.
—— Andrii Hrytseniuk · [32:00]

指向原始笔记的链接

如果你有这样的规定,至少会有三家公司去提供这项服务。政府不用花一分钱。
And if you had that, at least three companies would go provide that. It would cost the government nothing.
—— Reid Hoffman · [16:09]

指向原始笔记的链接

你要让你的解决方案具备与问题相同的特性。也就是在规模、成本、速度上,你要做到阻抗匹配。
You want your solution to have the same characteristics as the problem. Namely in scale, cost, speed, you wanna have an impedance match.
—— Reid Hoffman · [27:43]

指向原始笔记的链接

经济性实际上坍缩为本质上一个东西：你所使用的内存的每 GB 成本。
the economics actually collapsed down to essentially a kind of cost per gigabyte of the memory that you’re employing
—— Walter Goodwin · [14:42]

指向原始笔记的链接

支撑 AI 的基础设施不会是集中式的,它会分布在能源低成本且充沛的地方。
The infrastructure to support AI wasn’t going to be centralized. It was going to be distributed where energy was low cost and abundant.
—— Chase Lochmiller · [00:09]

指向原始笔记的链接

而他们得以消除大约 98% 的数据中心总成本。
And they were able to eliminate something like 98% of the total data center cost.
—— Chase Lochmiller · [15:45]

指向原始笔记的链接

超过某个点之后，你会有点陷入收益递减：成本和延迟在增加，但输出的质量并没有显著变好。
Past a certain point, you kind of hit diminishing returns where your cost and latency are increasing, but the quality of the outputs are not significantly better.
—— Sydney · [00:57]

指向原始笔记的链接

提前剧透一下实验结果：我们实现了每线程中位成本降低 64%，且质量没有可测量的变化。
And to kind of spoil the results of the experiment here, we were able to see a 64% reduction in median cost per thread with no measurable change in quality.
—— Sydney · [02:00]

指向原始笔记的链接

我认为，这里是工程工作如此关键的落点、作为工程师在这里工作如此酷的根本原因是：归根结底，训练是成本中心，推理是收入中心。
I think the fundamental reason why this is such a critical place for engineering work to get done and why it’s such a cool place to work as an engineer is that in the end, training is a cost center and inference is a revenue center.
—— Charles Frye · [04:52]

指向原始笔记的链接

前沿模型很了不起，但我没想到人们会如此在意真正地把成本优化好，为正确的任务使用正确的模型。
Frontier models are amazing, but I didn’t expect how much people would care about really cost-optimizing and using the right model for the right task.
—— Yash · [43:57]

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

实际上客户每天支付的是几美元几美分，但在 Anthropic 的成本端，补贴高达 150 甚至 750，这种经济模型是没法规模化的，即便是 Anthropic 这个体量的公司也不行。
So effectively what the customers were paying daily was dollars and cents, but on the cost side for Anthropic, it was like 150 to even 750 subsidized, which is effectively not an economics that you can scale, even for the company at the size, at the scale of Anthropic.
—— Dor Sasson · [03:15]

指向原始笔记的链接

结果发现 Trainium 从绝对性能和性价比的角度来看，可能是目前市场上最好的推理芯片。
It turns out that Terranium is actually maybe the best inference chip on the market right now from an absolute performance and cost performance point of view.
—— Matt Garman · [38:03]

指向原始笔记的链接

我不在乎它是不是只有十分之一的每 token 成本。这就是不值得的花费。
I don’t care if it’s a tenth of the cost per token. It’s just not worth the expense.
—— Thomas Somers · [46:05]

指向原始笔记的链接

创造软件的成本降低了,因此需求大幅增加,但这是针对一种非常特定的需求——能够高效管理编码智能体、而不是产出一大堆垃圾的人。
The sort of cost of software, creating software has gone down, therefore the demand has increased a lot, but for a very specific kind of demand for people who can manage coding agents productively instead of producing a whole bunch of slop.
—— Sean Wang · [48:55]

指向原始笔记的链接

归根结底，持久性是入场的代价。
At the end of the day, durability is the cost of admission.
—— Viren Baraiya · [07:08]

指向原始笔记的链接

创建代码的成本在下降，但理解一个软件程序的成本并没有。
The cost of creating code is falling, but the cost of understanding a software program is not.
—— Gabriel Martinez · [03:13]

指向原始笔记的链接

我们有个东西叫放血榜。在上面我们记录每个跑实验的人的实验成本。
We have something called a bleederboard. Where we keep track of the cost of experiments of everybody who’s running experiments.
—— Daval Ajada · [11:53]

指向原始笔记的链接

所以我总是对产品团队说，不要在短期内过度优化成本。先构建惊艳的、有魔力的功能。
So I say to product teams all the time, don’t over optimize for cost in the short term. Build amazing magical features.
—— Mike Cannon-Brookes · [29:37]

指向原始笔记的链接

② 出现在这些集

2 集

• 《当系统出故障时，让 AI 代替作战室里的 50 个人——Traversal 谈如何造 AI SRE》 — 作为概念
• 《AI没有消灭客服，反而让他们更累了？》 — 作为概念

③ 关联

点进去有真内容 —— 本页主要出口

智能体 · Anish · Dan O’Connell · Raj · Craig · Harrison · Front · Traversal · TalkIQ · 生产世界模型