跨国深谈

推理能力 (reasoning)

Properties1
aliasesreasoning

14分钟阅读

跨国深谈← 返回←推理能力 (reasoning)
推理

推理能力 (reasoning)

概念 · 又名 reasoning
本站收录 6 集 · 22 条金句 · 关联 10 个

集里怎么说它

① 提到它的金句

22 条

因为真正的推理能力并不存在。
Because the actual reasoning is not there.
—— Spiros · [30:56]

指向原始笔记的链接

我认为真正有价值的数据——现在人们正在意识到它有价值——是 RLHF 数据或者推理数据。
The data that I think is really valuable that I think now people are realizing is valuable is the RLHF data or the reasoning data.
—— Gabe · [23:17]

指向原始笔记的链接

因此,将这种推理嵌入到我们的搜索智能体中,实际上让我们能够将准确率从大约 50% 提高到了 90。
So embedding this sort of reasoning into our search agent is actually something that got us up from roughly like 50% accuracy all the way to 90.
—— 嘉宾 · [11:53]

指向原始笔记的链接

仿真扮演了一个非常重要的角色,这是现实世界的数据所没有的,那就是反事实推理,就是你在推演那些尚未发生或不可能发生的事件,或者你在现实世界中没有足够的数据让它发生。
There’s a very important role simulation plays that real world data doesn’t play, which is counterfactual reasoning, is that you play out events that hasn’t happened or cannot happen, or you don’t have enough data to make it happen in real world.
—— Fei-Fei Li · [20:06]

指向原始笔记的链接

因为我们发现的是，虽然我们通常可以越狱模型，很难让它在思维链中推理时，对自己即将做的坏事闭嘴。
Because what we found is that even though we can usually jailbreak the model, it’s really hard to get it to shut up about the evil thing that it’s about to do when it’s reasoning in the chain of thought.
—— Adam Gleave · [33:19]

指向原始笔记的链接

推理模型的「正常」，实际上相当于聊天模型的「故障」。
Reasoning models normal is actually a chat models outage.
—— Kanish Manuja · [08:06]

指向原始笔记的链接

与其写出一些长篇大论的推理文档，不如说是，我如何能尽快得到某种可以尝试并让用户测试的东西？
Rather than writing out some long reasoning doc, instead it’s like, how do I get to something I can try out and test with users as fast as possible?
—— Tara Seshan · [00:44]

指向原始笔记的链接

管理上下文的能力——不只是在两小时的会话里，而是跨越数周或数月的工作——我认为这将是关键的，而且需要比今天长得多的上下文推理能力。
The ability to manage context, not just through a two-hour session, but across weeks or months of work is something that I think will be critical and will also require much longer context reasoning than we have today.
—— Alexander Whedon · [32:19]

指向原始笔记的链接

我认为决策的严谨基础在质量上已经飙升了，也就是说，重新检查某件事的整个推理链变得非常容易。
I think that rigorous underpinning of decision-making has just skyrocketed in quality, which is that it’s super easy to recheck the entire chain of reasoning of something.
—— Tobi Lütke · [12:09]

指向原始笔记的链接

最大的问题是，模型根本没有足够的上下文去真正理解演讲者、听众、以及两者之间的关系，而再多的通用推理也无法让你跨过这道坎。
But I think the biggest issue was that the model just didn’t have enough context to really understand the presenter, the audience, the relationship between the presenter and the audience and no amount of general reasoning gets you past that.
—— Keith Parris · [04:52]

指向原始笔记的链接

神经网络的可监控性极其困难。而在这里我们处于一种神经网络就在直接进行推理、用自然语言把它们的思考过程摊开来供我们阅读的情形。这太方便了。这真的是安全方面的最佳情形。
Monitorability for neural nets is extremely hard. And here we have a situation where the neural nets are just flat out reasoning, laying out their thought process in natural language for us to read. That is so convenient. It was really the best case scenario for safety.
—— Noam Brown · [69:26]

指向原始笔记的链接

很多人不断在谈论模型，但问题是，像语言模型这样的模型，它们是推理中被冻结的部分，对吧？
Lots of people are talking about models constantly, but the thing is that models, like language models, they are the frozen part of the reasoning, right?
—— Ignacio Martinez · [03:56]

指向原始笔记的链接

所以 harness 工程的目标是一次又一次地创造可靠且可预测的输出，而推理模型是非常非确定性的。
So the goal of harness engineering is to create reliable and predictable outputs over and over, whereas a reasoning model is very non-deterministic.
—— Ignacio Martinez · [11:26]

指向原始笔记的链接

Dario 想说的也许是,我们将拥有数百万个具备爱因斯坦部分推理能力的实体,这一点我同意,但不是作为个体的爱因斯坦,不是能够在工作中学习的爱因斯坦。
Dario wanted to say that we’ll have millions of entities that will have some of the reasoning powers of Einstein, which I agree, but not Einstein’s as a person, not Einstein’s that are capable of learning on the job.
—— Daniel Dines · [05:53]

指向原始笔记的链接

离线使用一次昂贵的推理,然后把它蒸馏进可以廉价、快速地提供服务的模型中。
Use expensive reasoning once offline, then distill it into models that can be served cheaply and quickly.
—— Raghav Saboo · [06:10]

指向原始笔记的链接

在合适的地方,我们不是用 LLM 替换检索和排序系统,而是把推理和理解蒸馏进某种生产级排序架构。
Where it seems right, we’re not replacing the retrieval and ranking system with an LLM, but rather distilling the reasoning and understanding into some production ranking architecture.
—— Raghav Saboo · [09:06]

指向原始笔记的链接

把推理离线捕获为标签、语义 ID、记忆,然后让更小更快的模型来服务它。
Capture reasoning offline as labels, semantic IDs, memory, and then let smaller and faster models serve it.
—— Raghav Saboo · [21:16]

指向原始笔记的链接

一开始可能没问题，但如果你的 LLM 不理解位置和空间推理，到了某个时候就会崩掉。
It might be fine at the start, but if your LLM doesn’t understand the location and spatial reasoning, it falls apart at some point.
—— George He · [14:58]

指向原始笔记的链接

有时我觉得我们可能处于某种失忆循环中，对吧？我们的智能体经历推理阶段、行动，成功完成了一些任务，但随后就有点忘记了它们学到的东西。
Sometimes I feel like we’re in maybe like an amnesiac loop, right? Where our agents go through this reasoning phase, acting, they successfully complete some task, but then kind of forget what they’ve learned.
—— Will Lyon · [00:41]

指向原始笔记的链接

我思考推理的方式是：模型基本上是在从它们自己已经生成的内容中进行上下文学习。
And the way I think about reasoning is that models are basically in-context learning from things that they have generated already.
—— 嘉宾 · [12:25]

指向原始笔记的链接

我认为「推理」对于推理模型正在做的事情来说，几乎是个误称。
I think reasoning is almost like a misnomer for what reasoning models are doing.
—— 嘉宾 · [44:59]

指向原始笔记的链接

我还要说，以人类语言作为推理的起点有点成问题，因为它会带入我们所有关于我们认为智能体应该如何思考的偏见。
I would also say that starting from human language as like the approach for reasoning is somewhat problematic because it pulls in all of our biases about how we think agents should think.
—— David Silver · [26:30]

指向原始笔记的链接

② 出现在这些集

6 集

③ 关联

点进去有真内容 —— 本页主要出口

OpenAI · Anthropic · ChatGPT · 智能体 · RL · Cursor · Sarah Wang · a16z · Codex · Decagon

反向链接