--- author: Dwarkesh Patel date: '2024-06-12' guest: '' layout: post.njk source: https://www.youtube.com/watch?v=rl7B-LHiaNo speaker: Dwarkesh Patel tags: - general-intelligence - artificial-intelligence - llm-benchmarks - scaling-laws - reasoning-vs-skill title: 大模型依赖记忆而非智能：解读Scaling Law下的核心误区 summary: 该讲稿探讨了通用智能与技能的本质区别，指出当前大型语言模型（LLM）主要擅长技能习得和基于模式匹配的“记忆式推理”，而非真正的通用智能。通过分析Scaling Law和基准测试，文章揭示了模型表现的提升更多源于记忆能力的增强，而非智能本身。它强调了区分记忆与推理的重要性，并指出虽然记忆是推理的基础，但真正的智能在于即时合成新解决方案的能力。 insight: '' draft: true series: '' category: ai-ml area: tech-engineering project: [] people: [] companies_orgs: [] products_models: [] media_books: [] status: evergreen --- ### 智能与技能的本质区别：通用性非技能的简单叠加 演讲者首先阐述了**通用智能**（General Intelligence）的核心概念，并将其与**任务特定技能**（Task-Specific Skill）明确区分开来。通用智能并非将单一技能无限扩展，而是指一种能够快速掌握**任何**新技能、解决**任何**问题的能力，其关键在于能够利用**既有数据**（Valuable Old Data）来高效地面对未知挑战。这种能力使个体能够适应并即时学习（Learn on the Fly），这是“**通用性**”（Generality）的真正含义，而非简单地将特定技能“**规模化**”（Scaled Up）以覆盖更多领域。
Original English general intelligence is not task specific skill scaled up to many skills because there is an infinite space of possible skills general intelligence is the ability to approach any problem any skill and very quickly Master it using Val old data because this is what makes you able to face anything you might have ear cont this is what makes uh uh this this is the definition of generality like generality is not specificity scaled up it is uh the ability to apply your mind to anything at all to arbitrary things and this requires fundamentally this requires the ability to adapt to learn on the Fly efficiently
### “最大化主义”论点与Scaling Law的局限性 文章随后深入探讨了当前的“**最大化主义**”（Maximalist）论点，特别是其引用的“**Scaling Laws**”（LS）。这是一种在模型训练中观察到的经验性关系，描述了花费在训练上的**计算量**（Compute）与模型在**基准测试**（Benchmarks）中获得的**性能**（Performance）之间的关联。然而，核心问题在于如何衡量这种性能提升。演讲者强调，性能的衡量方式并非技术细节，它将直接**限定**我们提出的问题范围，进而**限制**我们寻找的答案。当前用于评估大型语言模型（LLMs）的基准测试，许多是**基于记忆**（Memorization-Based）的，有时甚至如同学校考试，仅需记忆一套固定的“**推理模式**”（Reasoning Patterns）并进行套用，而非真正的推理。
Original English the scale uh a maximalist argument it boils down to these people they refer to scaling LS which is this this empirical relationship that you can draw between how much compute you spend on training model and the performance you're getting on benchmarks right and the the key question here of course is well how do you measure performance what it is that you're actually uh improving by adding more comput and more data and well it's it's Benchmark performance right and the thing is the way you measure performance is not a technical detail uh it it's it's not it's not an afterthought because it's going to uh narrow down the set of questions that you're asking and so accordingly it's going to narrow down the set of answers that that you're that you're looking for if you look at uh the benchmarks we're using for LMS they're all memorization based benchmarks like sometimes they're literally just knowledge based like like a school test and even if you look at the ones that are uh you know explicitly about reasoning you realize if you look closely that it's uh in order to solve them it's enough to memorize uh a finite set of uh uh resoning patterns uh and then you just reapply them
### LLMs的“推理”：程序获取而非动态合成 当前的大型语言模型（LLMs）在处理这些基准测试时，表现得像是在执行“**程序获取**”（Program Fetching）。模型能够记忆并检索大量“**静态程序**”（Static Programs）或“**解决方案程序**”（Solution Programs），当面对新问题时，它们会从中“**提取**”（Fetch）最相关的程序并应用。这种模式看起来像是推理，但实际上并未涉及“**即时程序合成**”（On-the-Fly Program Synthesis）——即根据现有知识片段动态创建新解决方案的能力。
Original English they they're like static programs llms are very good at memorizing CTIC programs small CTIC programs and and they've got this sort of like Bank of uh solution programs and when you give them a new puzzle uh they can just fetch the appropriate program apply it and it's looking like it's reasoning but really it's not doing any sort of on thefly program synthesis all it's doing is program fetching so you can actually solve all these benchmarks with memorization
### 技能提升不等于智能增长：理解记忆与智能的混淆 当我们增加模型的**数据库规模**（Size of your database）并填充更多知识和模式时，其在记忆型基准测试上的性能会随之提升，这一点显而易见。然而，这**并未**在任何程度上增加系统的**智能**（Intelligence）本身。模型能力的增长体现在“**技能**”（Skill）的提升——即其**有用性**（Usefulness）和**适用范围**（Scope of Applicability）的扩大，而非智能水平的提高。演讲者强调，**技能并非智能**，这是导致混淆的“**现象性困惑**”（Phenomenal Confusion），人们常常将两者混为一谈。这种混淆源于对模型“**内插**”（Interpolation）能力的误解，即模型在已知数据点之间进行插值，而非真正的理解和泛化。
Original English and so what what you're scaling up here like if you look at the models they are uh big parametric curves uh fitted to a data distribution I descent so they are basically these big interpolative uh databases interpolative memories and of course if you scale up the size of your database and you cram into it uh more knowledge more patterns and so on uh you are going to be increasing its its performance as measured by memorization Benchmark that's that's kind of obvious but as you're doing it you are not increasing the intelligence of the system one bit you are increasing the skill of the system you you are increasing ining its usefulness its uh scope of applicability but not its intelligence because skill is not intelligence and that's the phenomal confusion um that that that people run into is that they're confusing skill and intelligence as far as the interpolation goes
### 数学基准测试的启示：区分推理的定义 为了具体说明这一点，演讲者举了一个“**GSS 8K**”数学基准测试的例子。该测试包含的题目，如一道关于班级学生年龄分布的数学题，对于一个“**聪明的や学の生徒**”（Smart High Schooler）来说本应是可以解决的。模型在此类测试上能达到95%的准确率，但这仍是基于记忆。演讲者提出了**推理**的两种定义：第一种，即**拥有一个程序模板库**，识别问题的结构后**匹配并套用**记忆中的模板来生成解决方案；第二种，也是更困难的一种，即在**没有现成程序**可供套用时，能够**动态地合成**（Synthesize on the Fly）一个全新的程序，这需要将现有程序的片段重新组合。后者才是真正困难的，而模型目前擅长的是前者。
Original English so okay let's look at one of the benchmarks here there's there's one Benchmark that does great school math and these are problems that like a smart high schooler would be able to solve um it's called GSS 8K and these models get 95% on these like basically they always nail memorization okay let's talk about what that means so here's one question about from that Benchmark so 30 students are in a class one/ fifth of them are 12 year olds 1/3 are 13y old 1110th are 11 year olds how many of them are not 11 12 or 13 years old so I agree like this is not rocket science right you can write down on paper how you go through this problem and a high school kid at least a smart high school kid should be able to solve it now when you say memorization it's still has to reason through how to think about fractions and what is the context of the whole problem and then combining the different calculations is doing it depends how you how you want to Define reasoning but there there are two definitions you can use so one is I have available uh a set of program templates it's it's like the structure of the puzzle which which can also generate its solution and I'm just going to identify the right template which is in my memory um I'm going to input the new value into the template run the program get the solution and you could say this is reasoning and I say yeah sure okay uh but another definition you can use is reasoning is the ability to when you're faced with a with a puzzle given that you don't have already a program in memory to solve it you must synthesize on the fly a new program based on bits of pieces of existing programs that you have you have to do on the Fly program synthesis and it's actually dramatically harder than just fetching the right memorized program and replying it
### 人类学习的普遍性与模型的记忆基础 演讲者进一步指出，我们可能**高估**了人类在样本效率上的程度。人类学习数学，例如代数、几何、微积分，需要多年的系统性教学和练习，而非仅仅观看一个示例就能掌握。这种“**训练**”（Training）过程，涉及反复钻研特定的**推理路径**（Pathways of Reasoning），与模型通过大量数据“**钻取**”（Drill）训练的过程在概念上有相似之处。尽管如此，为了实现“**即时程序合成**”，模型确实需要“**可用的积木**”（Building Blocks）——即**知识**（Knowledge）和**记忆**（Memory）。因此，记忆并非与推理对立，而是**极大地重要**（Tremendously Important）的，是有效推理的基础。
Original English I think maybe we are overestimating the extent to which humans are so sample efficient they also don't need training in this way where they have to drill in these kinds of Pathways of reasoning through certain kind of problems so let's take math for example yeah it's not like you can just show a baby the axum of SE Theory and now they know math right so they when they're growing up you had to do years of teaching them pre-algebra then you got to do a year of teaching them doing d and going through the same kind of problem in algebra then geometry pre-calculus calculus absolutely so training yeah but isn't that like the same kind of thing where you you you can't just see one example and now you have the program or whatever you actually had to drill it these models also had to drill with a bunch of fruit training data sure I mean in order to do on the-fly program synthesis you actually need uh building blocks to work from so knowledge and memory actually tremendously important in the process I'm not I'm not saying it's memory versus reasoning in order to do effective reasoning you need memory