Has Generative AI Already Peaked? - Computerphile

Summary

本文探讨了使用生成式人工智能（AI）来理解图像和文本的CLIP嵌入技术。作者质疑了通过增加数据和模型规模就能实现通用智能的观点，指出要达到零样本学习性能，所需的数据量可能是巨大的。文章通过实验发现，对于困难问题，现有模型在数据量不足的情况下表现不佳。此外，数据集中类别的分布不均也影响了模型对稀有类别的识别能力。尽管大型科技公司可能通过增加GPU和使用人类反馈来改进模型，但作者认为，若要处理在互联网文本和搜索中不常见的难题，可能需要寻找新的方法。

Takeaways

• 🧠 讨论了使用生成式AI来生成新句子、新图像等，并理解图像和文本的能力。
• 🔍 通过大量图像和文本配对学习，可以提炼图像内容的语言表示。
• 🚀 有观点认为，随着训练数据和网络规模的增加，AI将发展出跨领域的通用智能。
• 🔬 科学方法强调实验验证而非假设，对于AI性能提升的乐观预测持谨慎态度。
• 📉 近期论文指出，实现通用零样本性能所需的数据量可能非常庞大，难以实现。
• 📈 论文通过实验数据展示了数据量与模型性能之间的关系，通常呈现对数增长而非线性。
• 📊 论文分析了约4000个概念在数据集中的分布，以及它们在下游任务中的表现。
• 🌐 讨论了CLIP嵌入（图像和文本的共享嵌入空间）及其在分类、推荐系统等任务中的应用。
• 🚧 论文指出，对于困难问题，现有数据量不足以有效训练模型，导致性能受限。
• 📚 强调了数据集中类别和概念的不均匀分布问题，常见类别（如猫）过度表示，而特定类别（如某些树种）则表示不足。
• 🔑 暗示了除了收集更多数据之外，可能需要新的数据表示方法或机器学习策略来提升对困难任务的性能。

Q & A

什么是CLIP嵌入（clip embeddings）？

CLIP嵌入是一种通过大量图像和文本对训练得到的表示方法，能够将图像和文本映射到共享的嵌入空间中，使得描述相同内容的图像和文本在嵌入空间中彼此接近。

为什么有人认为通过增加更多的数据和更大的模型就能实现通用智能（general intelligence）？

这种观点基于观察到的现象：随着数据和模型规模的增加，AI在图像识别等领域的表现逐渐提升。因此，一些人认为，只要继续增加数据量和模型规模，AI最终能够处理所有类型的任务。

为什么说实验验证比假设更重要？

在科学领域，实验验证是检验假设正确性的关键。仅仅提出假设而不通过实验来证明它们，无法确保这些假设在实际应用中的有效性。

这篇论文为什么反对仅通过增加数据量和模型规模就能解决所有问题的观点？

这篇论文通过实验发现，要实现在新任务上的零样本（zero-shot）表现，所需的数据量是极其庞大的，以至于实际上无法实现。这表明仅靠增加数据和模型规模并不能无限提升AI的性能。

什么是下游任务（downstream tasks）？

下游任务是指在训练好基本模型后，利用这些模型进行的特定应用任务，如分类、推荐系统等。

为什么说在困难问题上应用下游任务需要大量的数据支持？

因为困难问题往往涉及到更具体的概念，这些概念在数据集中可能非常少见，导致模型无法学习到足够的特征来进行有效识别或分类。

这篇论文是如何测试不同概念在数据集中的分布和模型性能的关系的？

论文中定义了约4000个不同的概念，并分析了这些概念在数据集中的分布情况，然后测试了在这些概念上的下游任务性能，并将其与对应概念的数据量进行对比。

为什么说数据集中的类别分布不均匀会影响模型性能？

如果数据集中某些类别（如猫）过度表示，而其他类别（如特定树种）表示不足，模型在常见类别上的性能会很好，但在不常见类别上的性能会较差，因为模型没有足够的数据来学习这些类别的特征。

为什么说仅靠增加数据量可能无法实现AI性能的大幅提升？

论文中的实验结果表明，随着数据量的增加，模型性能的提升会逐渐趋于平缓，即达到一个平台期，这意味着继续增加数据量可能无法带来预期的性能提升。

这篇论文对于未来AI发展的意义是什么？

这篇论文提供了对当前AI发展模式的批判性思考，提示我们可能需要寻找新的方法或策略来提升AI的性能，而不仅仅是依赖于数据量的增加。

Outlines

00:00

🤖 AI与图像文本嵌入的局限性

本段讨论了AI在图像和文本嵌入方面的应用及其局限性。提到了通过大量图像和文本对来训练AI，使其能够将图像内容转化为语言描述的尝试。然而，有观点认为，随着数据量和模型规模的增加，AI将实现通用智能，但最新研究显示，要达到新的未见任务的零样本性能，所需的数据量将是巨大的。论文通过实验表明，对于困难问题，没有足够的数据支持，这些模型无法有效应用。此外，还提到了下游任务，如分类和推荐系统，以及它们如何利用图像和文本的嵌入空间来提高性能。

05:00

📈 数据量与AI性能的关系

这段内容通过一个图表形象地展示了数据量与AI性能之间的关系。研究者定义了约4000个核心概念，并分析了这些概念在数据集中的普遍性，以及它们在下游任务中的表现。实验结果表明，随着数据量的增加，AI性能的提升并不是线性的，而是呈现出对数增长，最终趋于平缓。这意味着，尽管增加数据量和模型规模可以提升AI性能，但很快就会达到一个性能提升的瓶颈。此外，数据集中类别和概念的分布不均也是导致性能下降的一个因素，常见类别如猫和狗的数据量远大于特定树种等不常见类别。

10:01

🎮 特定任务下AI的挑战与未来展望

最后一段讨论了AI在处理特定任务时遇到的挑战，尤其是在数据量不足的情况下。举例说明了当AI被要求生成不常见物体的图像或解释复杂概念时，性能会下降。同时，提出了对于AI未来发展的一些思考，包括是否需要新的数据表示方法或机器学习策略来突破当前的性能瓶颈。此外，还提到了Jane Street公司提供的技术问题解决项目，鼓励对计算机和问题解决感兴趣的观众参与。

Keywords

💡CLIP嵌入

CLIP嵌入指的是一种将图像和文本映射到共享的嵌入空间的技术，使得图像和描述它的文本在该空间中具有相似的表示。在视频中，CLIP嵌入被用来讨论如何通过大量图像和文本对来训练模型，以实现对图像内容的语言描述。例如，视频提到通过训练，模型能够将图像和描述性文本映射到嵌入空间中的相同点。

💡生成式AI

生成式AI是一种人工智能技术，它能够生成新的数据实例，如句子、图像等。视频中提到使用生成式AI来生成新句子和图像，并且讨论了这种技术在理解图像等方面的应用。生成式AI的核心在于通过学习大量数据来创建新的、以前未见过的内容。

💡通用智能

通用智能指的是一种能够在多种任务和环境中有效工作的人工智能，与特定领域的智能相对。视频中提到了一种观点，即通过训练足够多的图像和文本对，AI最终能够发展出通用智能，处理各种不同的问题。

💡数据集

数据集是用于训练机器学习模型的数据集合。视频中讨论了数据集的大小对于训练有效AI模型的重要性，指出要实现零样本学习等高级功能，可能需要极其庞大的数据集。

💡零样本学习

零样本学习是一种机器学习范式，其中模型能够在没有直接针对特定类别的训练数据的情况下进行分类。视频中提到，实现这种性能可能需要天文数字级别的数据量。

💡下游任务

下游任务指的是在预训练模型的基础上进行的特定应用任务，如分类、推荐系统等。视频通过CLIP嵌入的例子，讨论了如何利用共享嵌入空间来进行图像分类和推荐系统等任务。

💡推荐系统

推荐系统是一种信息过滤系统，用于预测用户可能感兴趣的项目。在视频中，推荐系统被用作CLIP嵌入的一个应用实例，通过嵌入空间来推荐用户可能感兴趣的节目或内容。

💡概念普及度

概念普及度指的是在数据集中某个概念出现的频率。视频中通过对4000个不同概念的普及度进行分析，研究了数据量对于模型性能的影响。

💡性能提升

性能提升是指机器学习模型在特定任务上的表现改进。视频通过实验数据展示了随着数据量的增加，模型性能提升的趋势，并讨论了可能遇到的性能瓶颈。

💡模型泛化

模型泛化是指模型对新数据或未见过的数据的预测能力。视频中提到，尽管模型可能在常见类别上表现良好，但在少见或难于识别的类别上可能性能不足，这与模型的泛化能力有关。

💡数据分布

数据分布描述了数据集中各类别或概念的代表性。视频中指出，数据集中某些类别可能过度表示，而其他类别则可能严重不足，这影响了模型对少见类别的识别能力。

💡机器学习策略

机器学习策略包括用于提高模型性能的各种方法和技术。视频最后提出了一个问题，即是否需要新的机器学习策略来解决当前模型在处理难任务时遇到的瓶颈。

Highlights

讨论了使用生成性AI来生成新句子和图像，并理解图像和文本。

提出了通过足够多的图像和文本配对学习，可以提炼图像内容的语言表示。

提出了随着训练数据和网络规模的增加，AI将发展出跨领域的通用智能。

科学界通常通过实验来验证假设，而不是仅仅基于理论推测。

最近发表的一篇论文认为，要实现零样本学习，需要的数据量将非常庞大。

论文中的数据和图表显示，增加数据量和模型规模并不能无限提高性能。

介绍了CLIP嵌入的概念，包括图像和文本的Transformer编码器。

CLIP嵌入可以用于分类、图像召回和推荐系统等下游任务。

论文指出，对于困难问题，没有大量数据支持，CLIP嵌入的下游任务效果不佳。

论文通过定义核心概念，并分析这些概念在数据集中的普遍性，来测试下游任务的表现。

实验结果表明，对于某些概念，即使数据量增加，性能提升也有限。

论文提出了对于数据集中类别和概念分布不均的问题。

讨论了大型语言模型在处理训练集中不常见的问题时可能出现的准确性下降。

提出了对于困难任务，可能需要寻找除收集更多数据之外的其他方法。

论文的实验结果对于AI领域的未来发展提出了质疑和思考。

讨论了大型科技公司可能对AI发展过于乐观的宣传。

提出了对于AI模型训练成本和效率的考量。

提到了Jane Street公司提供的技术问题解决项目和赞助。

Transcripts

so we looked at clip embeddings right

and we've talked a lot about using

generative AI to produce new sentences

to produce new images and so on and so

to understand images all these kind of

different things and the idea was that

if we look at enough pairs of images and

text we will learn to distill what it is

in an image into that kind of language

so the idea is you have an image you

have some texts and you can find a

representation where they're both the

same the argument has gone that it's

only a matter of time before we have so

many images that we train on and so and

such a big Network and all this kind of

business that we get this kind of

general intelligence or we get some kind

of extremely effective AI that works

across all domains right that's the

implication right the argument is and

you see a lot in the sort of tech sector

from the from some of these sort of um

big tech companies who to be fair want

to sell products right that if you just

keep adding more and more data or bigger

and bigger models or a combination of

both ultimately you will move Beyond

just recognizing cats and you'll be able

to do anything right that's the idea you

show enough cats and dogs and eventually

the elephant just is

implied as someone who works in science

we don't hypothesize about what happens

we experimentally justify it right so I

would say if you're going to if you're

going to say to me that the only upward

trajectory is is going you know the only

trajectory is up it's going to be

amazing I would say go on and prove it

and do it right and then we'll see we'll

sit here for a couple of years and we'll

see what happens but in the meantime

let's look at this paper right which

came out just recently this

paper is saying that that is not true

right this paper is saying that the

amount of data you will need to get that

kind of General zero shot performance

that is to say performance on new tasks

that you've never

seen is going to be astronomically vast

to the point where we cannot do it right

that's the idea so it basically is

arguing against the idea that we can

just add more data and more models and

we we'll solve it right now this is only

one p

and of course you know your mileage may

vary if you have a bigger GPU than these

people and so on but I think that this

is actual numbers right which is what I

like because I want to see tables of

data that show a trend actually

happening or not happening I think

that's much more interesting than

someone's blog post that says I think

this is going what's going to happen so

let's talk about what this paper does

and why it's interesting we have clip

embeddings right so we have an image we

have a big Vision Transformer and we

have a big text encoder which is another

Transformer bit like the sort of you

would see in a large language model

right which takes text strings my text

string today and we have some shared

embedded space and that embedded space

is just a numerical fingerprint for the

meaning in these two items and they're

trained remember across many many images

such that when you put the same image

and the text that describes that image

in you get something in the middle that

matches and the idea then is you can use

that for other tasks like you can use

that for classification you can use it

for image recall if you use a streaming

service like Spotify or Netflix right

they have this thing called a recom

recommended system a recommended system

is where you've watched this program

this program this program what should

you watch next right and you you might

have noticed that your mileage may vary

on how effective that is but actually I

think they're pretty impressive what

they have to do but you could use this

for a recommender system because you

could say basically what programs have I

got that embed into the same space of

all the things I just watched and and

recommend them that way right so there

are Downstream tasks like classification

and recommendations that we could use

based on a system like this what this

paper is showing is that you cannot

apply these effectively these Downstream

tasks for difficult problems without

massive amounts of data to back it up

right and so and the idea that you can

apply you know this kind of

classification on hard things so not

just cats and dogs but specific cats and

specific dogs or subspecies of tree

right or difficult problems where the

the answer is more difficult than just

the broad category that there isn't

enough data on those things to train

these models and way I've got one of

those apps that tells you what specific

species a tree is so is it not just

similar to that no because they're just

doing classification right or some other

problem they're not using this kind of

generative giant AI right the argument

has been why do that silly little

problem where you can do a general

problem and solve all your problems

right and the response is because it

didn't work right that's that's that's

that's why we're doing it um so there

are pros and cons for both right I'm not

going to say that no generative AI is

useful or no or these these models are

incredibly effective for what they do

but I'm perhaps suggesting that it may

not be reasonable to expect them to do

very difficult medical diagnosis because

you haven't got the data set to back

that up right so how does this paper do

this well what they do is they def they

Define these Core Concepts right so some

of the concepts are going to be simple

ones like a cat or a person some of them

are going to be slightly more difficult

like a specific species of cat or a

specific disease in an image or

something like this and they they come

up about

4,000 different concepts right and these

are simple text Concepts right these are

not complicated philosophical ideas

right I don't know how well it embeds

those and and what they do is they look

at the prevalence of these Concepts in

these data sets and then they sh they

they test how well the downstream task

of let's say one zero shot

classification or recall recommended

systems works on all of these different

concepts and they plot that against the

amount of data that they had for that

specific concept right so let's draw a

graph and that will help me make it more

clear right so let's imagine we have a

graph here like this and this is the

number of

examples in our training set of a

specific concept right so let's say a

cat a dog something more difficult and

this is the performance on the actual

task of let's say recommend a system or

recall of an object or the ability to

actually classify as a cat right

remember we talked about how you could

use this for zero shck classification by

just seeing if it embeds to the same

place as a picture of a cat the text a

picture of a cat that kind of process so

this is performance right the best case

scenario if you want to have an all

powerful AI that can solve all the

world's problems is that this line goes

very steeply upwards right this is the

exciting case it goes like like this

right that's the exciting case this is

the kind of AI explosion argument that

basically says we're on the Custer

something that's about to happen

whatever that may be where the scale is

going to be such that this can just do

anything right okay then there the

perhaps slightly more reasonable should

we say pragmatic interpretation which is

like just call it balanced right which

is but there a sort of linear movement

right so the idea is that we have to add

a lot of examples but we are going to

get a decent performance Boost from it

right so we just keep adding examples

we'll keep getting better and that's

going to be great and remember that if

we ended up up here we have something

that could take any image and tell you

exactly what's in it under any

circumstance right that's that's kind of

what we're aiming for and similarly for

large language models this would be

something that could write with

Incredible accuracy on lots of different

topics or for image generation it would

be something that could take your prompt

and generate a photorealistic image of

that with almost no coercion at all

that's kind of the goal this paper has

done a lot of experiments on a lot of

these Concepts across a lot of models

across a lot of Downstream tasks and

let's call this the evidence what you're

going to call it pessimistic now it is

pessimistic also right it's logarithmic

so it basically goes like this right

flattens out it flattens out now this is

just one paper right it doesn't

necessarily mean that it will always

flatten out but the argument is I think

that and it's not an argument they

necessarily make in in the paper but you

know the paper's very reasonable I'm

being a bit more Cavalier with my

wording the suggestion is that you can

keep adding more examples you can keep

making your models bigger but we are

soon about to hit a plateau where we

don't get any better and it's costing

you millions and millions of dollars to

train this at what point do you go well

that's probably about as good as we're

going to get with technology right and

then the argument goes we need something

else we need something in the

Transformer or some other way of

representing data or some other machine

learning strategy or some other strategy

that's better than this in the long term

if we want to have this line G up here

or this line gar up here that's that's

kind of the argument and so this is

essentially

evidence I would argue against the kind

of

explosion you know possibility of but

just you just add a bit more data and we

were on the cusp of something we might

come back here in a couple of years you

know if you're still allow me on

computer file after this absolute

embarrassment of of these claims that I

made um and we say okay actually the

performan has improve improved massively

right or we might say we've doubled the

number of data sets to 10 billion images

and we've got 1% more right on the on on

the classification to which is good but

is it worth it I don't know this is a

really interesting paper because it's

very very fough right if there's a lot

of evidence there's a lot of Curves and

they all look exactly the same it

doesn't doesn't matter what method you

use it doesn't matter what data set you

train on it doesn't matter what your

Downstream task is the vast majority of

them show this kind of problem and the

other problem is that we don't have a a

nice even distribution of classes and

Concepts within our data set so for

example cats you can imagine are over um

emphasized or over represented over

represented yeah over represented in the

data set by an order of magnitude right

whereas specific planes or specific

trees are incredibly under represented

because you just have tree right so I

mean trees are probably going to be less

represented than cats anyway but then

specific species of tree very very

underrepresented which is why when you

ask one of these models what kind of cat

is this or what kind of tree is this it

performs worse than when you ask it what

animal is this because it's a much

easier problem and you see the same

thing in image generation if you ask it

to draw a picture of something really

obvious like a castle where that comes

up a lot in the training set it can draw

you a Fant fantastic castle in the style

of Monet and it can do all this other

stuff but if you ask it to draw some

obscure artifact from a video game

that's barely even made it into the

training set suddenly it's starting to

draw something a little bit less quality

and the same with large language models

this paper isn't about large language

models but the same process you can see

actually already happening if you talk

to something like chap GPT when you ask

it about a really important topic from

physics or something like this it will

usually give you a pretty good

explanation of that thing because that

in the training set but the question is

what happens when you ask it about

something more difficult right when you

ask it to write that code which is

actually quite difficult to write and it

starts to make things up it starts to

hallucinate and it starts to be less

accurate and that is essentially the

performance degrading because it's under

represented in the training set the

argument I think is at least it's the

argument that I'm starting to come

around to thinking if you want

performance on hard tasks tasks that are

under represented on just general

internet text and searches we have to

find some other way of doing it than

just is collecting more and more data

right particularly because it's

incredibly inefficient to do this right

on the other hand we they you know these

companies will they've got a lot more

gpus than me right they're going to

train on on bigger and bigger corpuses

better quality data they're going to use

human feedback to better train their

language models and things so they may

find ways to improve this you know up

this way a little bit as we go forward

but it's going to be really interesting

see what happens because you know will

it Plateau out will we see trap GPT 7

or8 or 9 be roughly the same as chat

dpt4 or will we see another

state-of-the-art performance boost every

time I'm kind of trending this way but

you know it'll be excited to see if it

goes this way take a look at this puzzle

devised by today's episode sponsor Jane

straight it's called bug bite inspired

by debugging code that world we're all

too familiar with where solving one

problem might lead to a whole chain of

others we'll link to the puzzle in the

video description let me know how you to

get on and speaking of Jane Street we're

also going to link to some programs that

they're running at the moment these

events are all expenses paid and give a

little taste of the tech and problem

solving used at trading firms like Jane

Street are you curious are you Problem

Solver are you into computers I think

maybe you are if so well you may well be

eligible to apply for one of these

programs check out the links below or

visit the Jane Street website and follow

the these links there are some deadlines

coming up for ones you might want to

look at and there are always more on the

horizon our thanks to Jane Street for

running great programs like this and

also supporting our Channel and don't

forget to check out that bug bite puzzle