跨国深谈

训练 (training)

14分钟阅读

跨国深谈← 返回←训练 (training)
训练

训练 (training)

概念
本站收录 5 集 · 49 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

49 条

而且我们没有使用任何来自大型提供商的数据。我们所有的数据都是自己从零开始收集的，通过训练付费的教师。
And we didn’t use any data from big providers. We collected all of it ourselves from scratch by training paid teachers.
—— Mustafa Suleyman · [15:05]

指向原始笔记的链接

如果你带来了很多价值，但你开始训练你的客户期望每月支付 20 美元，而你把自己锚定在低价位上，那你就有麻烦了。
If you’re bringing a lot of value to the table and you start at training your customers to expect $20 a month and you anchored yourself on a low price point, you’re in trouble.
—— Madhavan Ramanujam · [00:14]

指向原始笔记的链接

而如果你不在第一天就捕获这些价值，那你就是在训练你的客户期待花更少的钱获得更多。
And if you don’t capture that from day one, then you’re training your customers to expect more for less.
—— Madhavan Ramanujam · [28:26]

指向原始笔记的链接

在这个中间过程中，既然每个人都有非常相似的预训练数据，那后训练就是如今他们做出大差异的地方。
where like post-trading, but middle course of this is more of everyone have very similar pre-training data, is that post-training is where they make a big difference nowadays.
—— Chip Huyen · [15:10]

指向原始笔记的链接

我认为持续训练现在正在成为一种……我不想说它已被解决，但它现在像一个工程问题了，就是我们知道该怎么做
I think continual training now is becoming this kind of, I don’t want to call it solved, but it’s like an engineering problem now, where it’s like we know what to do
—— Julian · [08:00]

指向原始笔记的链接

当你坐下来使用 Cloud Code 或 Codex 时，你不是在编写软件，你是在雇佣、培训和管理一支由 Markdown 组成的劳动力。
When you sit down with Cloud Code or Codex, you’re not writing software, you’re hiring, training, and managing a workforce made of Markdown.
—— Garry Tan · [05:58]

指向原始笔记的链接

如果你认为智能是数据的派生物，那么大部分数据其实并没有被用于训练通用智能模型。
If you think intelligence is a derivative of data, then majority of the data is actually not used for training a general intelligence model.
—— Lin Qiao · [07:32]

指向原始笔记的链接

不要仅仅把这些权重视为训练过程的终点，如果它们也是下一个过程的起点呢？
Instead of treating those weights just as the end of the training process, what if they’re also the beginning of the next one?
—— Sam Charrington · [01:12]

指向原始笔记的链接

是的，我认为这是一个好问题，对开发者来说宽容的看法是，如果有一件事你真的不想搞乱，那就是预训练，因为它只是比你做的其他任何训练过程都要昂贵几个数量级。
Yeah, I think it’s a good question and the charitable take for developers is that if there’s one thing you really don’t want to mess with, it is pre-training because this is just orders of magnitude more expensive than every other training procedure that you do.
—— Adam Gleave · [74:22]

指向原始笔记的链接

结果发现他们的训练集中有大约二十五万个活跃密钥。
Turned out there were about a quarter million live keys in their training sets.
—— Dylan · [11:22]

指向原始笔记的链接

结果证明,如果你为人类构建了很棒的交接,你同时也为智能体构建了很棒的交接,因为智能体理解 HTML 和 CSS,而且这些在它们的训练数据中。
It turns out if you build a great handoff for humans, you’ve also built a great handoff for agents because agents understand HTML and CSS and it’s in their training data.
—— Stephen Haney · [03:35]

指向原始笔记的链接

过去做后训练是不值得的，因为模型通过预训练提升得太快，你做的任何后训练很快就会被下一个预训练模型吸收掉。
In the past, it wasn’t worth doing post-training because the models were improving so quickly from pre-training that any post-training you did quickly got absorbed by the next pre-trained model.
—— Gabe Pereyra · [06:04]

指向原始笔记的链接

但我们在一些客户那里看到的情况是，用户提出一些需求，模型勉强能做到，然后在训练过程中它学会了怎么做，之后用户就能做到以前做不到的事情了。
But what we’re seeing with some of our customers are cases where users ask for things, the model can barely do it, but then during training it learns how to do it and then the user can now do this thing they couldn’t do before.
—— Arjun Karanam · [21:25]

指向原始笔记的链接

当我们第一次看到机器人这样做时，我们惊呆了，因为这个任务没有任何训练数据。
The first time we saw the robot do this, we were like floored because there was no training data for this task.
—— Chelsea Finn · [33:32]

指向原始笔记的链接

所以我认为赢家是那些拥有训练数据的人。
So then I think the winner is folks who have the training data.
—— 嘉宾 · [27:34]

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

它们挣扎是因为训练它们的范式没有优先考虑这一点。
They struggle because the paradigm for training them hasn’t prioritized that.
—— Zubin Gharemani · [21:10]

指向原始笔记的链接

当正在训练下一代模型的模型本身在作弊时会发生什么？
What happens when the models that are doing the training of the next models are themselves cheating?
—— 嘉宾 · [13:23]

指向原始笔记的链接

我们有零个案例是进行训练的团队首先发现了这些问题。
We have zero cases where the teams doing the training found these issues first.
—— 嘉宾 · [80:48]

指向原始笔记的链接

你需要的数千亿自由现金流，用来偿还你为适应这些数据中心扩建而承担的债务，为了进行下一次大型训练运行，这对他们来说完全是生死攸关的。
The hundreds of billions in free cash flow that you need in order to pay back the debt that you’re taking on in order to accommodate these data center build outs in order to get the next big training run, it’s totally existential for them.
—— Eno Reyes · [52:16]

指向原始笔记的链接

并不清楚的是,用新架构和新范式预训练一个新模型的第一个周期,第一个周期就奏效是疯狂的。
It’s not clear that the first cycle of pre-training a new model with a new architecture and a new paradigm, the first cycle of that working is insane.
—— Justin Johnson · [22:32]

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

我要花 50 亿美元训练一个前沿模型，然后我就这么把它免费送出去。听起来很蠢，对吧？
I’m going to spend $5 billion training a frontier model and then I’m just going to give it away. Sounds stupid, right?
—— Anastasios Angelopoulos · [17:01]

指向原始笔记的链接

据我们所知，我们是唯一做过稳健的数百万 token 预训练的人。
As far as we know, we are the only people to do robust multi-million token pre-training.
—— Alexander Whedon · [11:26]

指向原始笔记的链接

研究在很大程度上表明，你在后训练阶段创造价值的能力，受限于你所做过的预训练。
And so research has largely suggested that your ability to create value in the post-training phase is limited by the pre-training that you’ve done.
—— Alexander Whedon · [20:31]

指向原始笔记的链接

以某种不会在基准测试中显现的方式搞砸后训练，真的太容易了。
So it’s like really easy to screw up post-training in some way that doesn’t show up in benchmarks.
—— 嘉宾 · [27:39]

指向原始笔记的链接

这真的是训练期间信噪比的一次极其巨大的提升，这就是 RL 在步数上如此戏剧性高效的原因。
it’s really a super dramatic increase into the signal to noise ratio during training, which is why RL is so dramatically efficient in terms of steps.
—— 嘉宾 · [80:20]

指向原始笔记的链接

我认为有一个很有力的论点,即把智能体训练得高度合作,实际上比任何其他多智能体替代方案都更可取。
I think there is a strong argument that training the agents to be highly cooperative is actually preferable to any other multi-agent alternative.
—— Noam Brown · [45:02]

指向原始笔记的链接

如果他们停止训练，他们会一夜之间实现巨额盈利
if they stopped training, they’d be massively profitable overnight.
—— Thomas Sohmers · [12:49]

指向原始笔记的链接

而认为为工业革命所做的训练能完美映射到 AI 革命，这种想法很可能并不成立。
And the idea that the training for the industrial revolution is going to map into the AI revolution is probably not perfectly.
—— Ben Horowitz · [08:04]

指向原始笔记的链接

归根结底，你可以对怎么改进问题有洞察，但借用老的强化学习训练或简单模型训练里的一句格言：跑更多实验总是比跑更少实验要好。
Because ultimately, you can have insight into how to make the problem better, but if we take an adage from old reinforcement learning training or just simple model training, it’s just better to run more experiments than fewer.
—— Zubin Aysola · [07:48]

指向原始笔记的链接

所以对我而言，就是，好吧，为什么不直接构建一个非常好的基础 LLM，然后用它来控制机器人，而不是训练一个专门面向机器人、更依赖于纯机器人数据的模型。
And then so to me, it’s like, okay, why not just build a really good foundational LLM and then use that to control robots rather than training like a model that’s specifically for robotics and that’s more dependent on just robot data.
—— Ham · [06:52]

指向原始笔记的链接

他们开始以某种方式利用测试数据，用 LLM 作为裁判来训练模型，用各种花招来刷基准，拿季度奖金之类的。
They start to kind of use leverage test data to use LLM as a judge for training of the models, use all the various tricks to basically gain benchmarks, get like quarterly bonuses and so on.
—— Alex Mashrabov · [23:52]

指向原始笔记的链接

你的预训练模型已经知道答案，而你针对一个已经存在于预训练语料库中的答案做强化学习时，系统基本上可以伪造一些工作、得到正确答案，而你会强化这些策略。
Your pre-trained model already knows the answer, and you do reinforcement learning derived against an answer that was already in the pre-training corpus, that the system can basically fake some work, get to the right answer, and you reinforce those strategies.
—— Doj Chubuk · [39:47]

指向原始笔记的链接

训练侧和软件侧在不断变好，但物理这一侧是很难一夜之间变好的东西。
The training side and the software side of things is continually getting better, but the physical side is the thing that’s harder to get better overnight.
—— Adam · [05:39]

指向原始笔记的链接

所以一个可合并性的基准，如果我们能做出来的话，不仅仅是一种度量，它会立刻成为前沿模型的训练信号——也就是可维护性、范围纪律、回归安全性这些。
So a benchmark of mergeability, if we could make one, wouldn’t only be a measurement, it would immediately become a training signal for the frontier models, which means maintainability, scope, discipline, regression safety.
—— Laurie Voss · [09:59]

指向原始笔记的链接

它在现实世界中不存在，不在真实训练数据中。这些数据必须被合成出来并放进模型里。
It doesn’t exist in the real world, in real training data. Those have to be synthesized and put into the model.
—— Jan Čurn · [06:39]

指向原始笔记的链接

它的预训练中没有任何机器人数据。
had zero robot data in its pre-training.
—— Pete Florence · [19:50]

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

人们普遍无法把直接向别人出售模型权重的生意做成，所以你没法真正把一个训练部门变成一个能自我支撑的系统。
People have not generally been able to make businesses work where they sell model weights to people directly, and so you can’t really turn a training setup into a proper self-supporting system.
—— Charles Frye · [05:11]

指向原始笔记的链接

我们的总体看法是，你的数据越偏离分布，后训练就越有可能真正给你带来那种能力提升。
Our general take has been like the more out of distribution your data is, the more likely that post training is actually going to give you that capabilities lift.
—— Yash · [08:30]

指向原始笔记的链接

阻碍持续学习的，基本上就是从稀疏奖励中做到极其省数据的训练，这是一个未解决的问题。
What’s blocking continual learning is basically, like, extremely data-efficient training from sparse rewards, which is an unsolved problem.
—— Yash · [13:36]

指向原始笔记的链接

所以我们的观点是，后训练实际上赢得推理。
And so what our view is is that the most post-training actually wins inference.
—— Yash · [29:24]

指向原始笔记的链接

后训练是这个技术栈里被严重忽视的一部分。
Hey, post-training is a massively underlooked part of the stack.
—— Yash · [36:10]

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

我们的模型通常需要训练好几个小时——前提是你有足够的训练数据——才能产生最完美的最优配置，但 LLM 只要 15 分钟就能进来，产出一个足够好的东西，而对大多数人来说这就够了。
Our models oftentimes would take hours and hours to train, assuming you had enough training data, to produce the pristine optimal configuration, but the elements can come in just like in 15 minutes, produce something that was good enough, and that’s good enough for most people.
—— Andy Pavlo · [34:31]

指向原始笔记的链接

顺便说一下，作为工作负载演进的另一个特征，我确实认为训练和推理的分离未来会变得不那么明显，而不是更明显。
By the way, as another characteristic of workload evolution, I do think the separation of training and inferencing is going to become less so going forward, not more so.
—— Pat Gelsinger · [40:41]

指向原始笔记的链接

我个人认为那可能一直是最大的护城河，并且将继续是：它们在各个层面上所做的后训练。
I personally think that’s probably been the biggest moat and will continue to be is the post training that they’re doing and at various levels.
—— 嘉宾 · [52:03]

指向原始笔记的链接

而且我认为这是 Hugging Face 事件的教训之一，即安全不只是当你即将部署它、检查它是否可以安全部署的时候才开始，而是要贯穿整个训练过程本身，那可能存在不对齐的风险。
And like, I think this is one of the lessons from the Hugging Face is that safety needs to start, not just when you’re about to deploy it and in like checking that it’s safe to deploy, but throughout the training process itself, that could be a risk of misalignment.
—— Nick Bostrom · [09:14]

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

推理 · OpenAI · NVIDIA · Anthropic · GPU · 智能体 · ASML · Navin Chaddha · Dylan Patel · Harry Stebbings

反向链接