# OpenAI Overview
## 中文
歡迎回來，我們將概述OPEN AI公司。

如果您已經註冊了這門課程，我相信您對Openai這家公司應該有一定的了解，並且您可能已經聽說過GPT、GPT和Dolly等事情。但是，在我們深入探討Openai Python API所創建的驚人項目和事物之前，我想為您簡要介紹一下公司本身，讓您了解其在更大的技術和AI生態系統中的角色。

現在需要注意的是，在人工智能領域，事情發展非常迅速，因此，您在觀看此課程時，信息和歷史可能不會完全更新。為了讓您了解事物發展的速度，即使在創建此課程期間，Openai公司的額外API和資金也已經宣布。因此，我們的歷史將帶您一直到2023年1月，當時他們與微軟進行了最新一輪的融資。

但讓我們從開始時介紹Openai創建時的主要人物開始著手。

回到2015年12月，在蒙特利爾的神經信息處理系統會議結束時，埃隆·馬斯克和Sam Altman宣布創建了一個新的非營利組織，他們將這個組織稱為Openai，即以開放的方式開展人工智能研究，讓人們可以像訪問開源代碼一樣輕鬆地訪問這些研究成果。

現在讓我們談談這兩個人，Sam Altman和Elon Musk。Sam Altman在2005年輟學於斯坦福大學，創建了一個基於位置的社交網絡Looped，並於2012年被收購，收購價為4340萬美元。值得注意的是，Looped實際上籌集了約3000萬美元。因此，這並不是一個巨大的退出，但很明顯，Sam能夠在19歲輟學斯坦福大學後建立這家公司是非常令人印象深刻的事情。

然後，在2011年，Altman實際上開始成為創業加速器Y Combinator的兼職合夥人。幾年後，他成為了Y Combinator的總裁，而Y Combinator是矽谷著名的加速器。許多著名的公司都來自這裡，比如Airbnb和Dropbox。
因此，即使Openai是在2015年成立的，但實際上直到2019年，Altman才退出了Y Combinator的總裁職務，全身心投入到Openai中。在這個拍攝時，Garry Tan現在是Y Combinator的總裁。

現在談談Elon Musk。

如果你已經參加了這門課程，如果你不知道誰是埃隆·馬斯克，我會非常驚訝。
但讓我們快速介紹一下埃隆，因為他參與了許多公司。

1995年，他和他的兩個兄弟Kimbal Musk和Greg Curry創立了ZIP。他們於1999年以3.07億美元的價格出售給康柏公司，當時埃隆的股份價值約為700萬美元。然後在1999年，Musk創立了一家在線金融服務公司X dot com，x dotcom的願景實際上是成為全球的網上銀行。任何你可以在銀行裡做的事情，你都希望能夠在線上做到。當然，這是在矽谷的.com熱潮期間，後來由於X dotcom和Infinity試圖吸引用戶而燒錢，所以在2000年它們合併成立了PayPal。然後在2002年，PayPal實際上被eBay收購，價值15億美元。此時，埃隆·馬斯克的股份價值約為1.76億美元。

對於埃隆來說，這是一筆相當可觀的交易。

現在，有點瘋狂的是，他繼續用之前公司的全部收益創立更多的公司。所以在2002年，他還創立了Space X這家公司，如果你沒聽說過，實際上是一家火箭製造商和發射器。他們發射衛星進入太空。Space X有點不同的是，他們開發的火箭實際上可以回來，而不只是在大氣層中被摧毀。後來，Space X實際上會通過StarLink開發自己的衛星互聯網。接著在2004年，馬斯克實際上投資了特斯拉。

與普遍的觀念相反，他並非特斯拉的創始人。他是最早的投資者之一，並且成為特斯拉最大的股東，這使他在2005年能夠在特斯拉中扮演更積極的角色，最終成為了CEO。從那時起到2022年，發生了許多事情。馬斯克實際上以430億美元收購了Twitter。如果您對這個過程有興趣，可以進行一次谷歌搜索。
他還參與了許多其他公司。

他參與了SolarCity，後來被特斯拉收購。他還參與了The Boring Company，這是一家隧道公司，以及Neuralink。他是Neuralink的創始人之一和主要投資者，該公司致力於開發可以從大腦中提取電信號，然後與身體的其他部分進行通信的植入物。StarLink是他的衛星網路業務，實際上是SpaceX的子公司。
當然，還有OpenAI，正如我們之前提到的，這家公司於2015年成立。
這位男子參與的公司真的很多。

現在，讓我們回到 Openai 並把焦點轉回去。
在開始時，有一些真正的主要投資者。其中包括 LinkedIn 的創始人 Reid Hoffman、Y Combinator 的創始人之一 Jessica Livingston、Peter Thiel、Infosys、Khosla Ventures 和 YC Research。所以，在矽谷也有很多大玩家在 Openai 的開始時就參與其中。

事實上，Openai 在創立時有大約 10 億美元的資金和資本。還有很多早期的員工，你可能認識他們的名字，也可能不認識。你應該注意的是，這些人基本上都是人工智能和機器學習研究的 A+ 級人才，而 Openai 成立於 2015 年，正是人工智能開始的時候。因此，他們聘請了 AI 和機器學習研究的權威人士，還有很多其他受人尊敬的顧問，他們雖然沒有參與 Openai 的工作，但是為公司提供了戰略上的建議。
現在，我提到 Openai 最初是作為非營利組織成立的，主要是為了安全地開發人工智能。

伊隆·馬斯克和 Sam Altman 成立 Openai 的主要原因之一是，到 2014 年（Openai 成立的前一年），Google 收購了一家名為 DeepMind 的公司，如果你曾經看過 DeepMind 在早期開發中培養機器學習模型以打敗 Atari 遊戲的影片，就會對這家公司有所了解。然後，他們後來開發了 AlphaGo 和其他 Alpha Fold，基本上是機器學習模型，用於征服象棋、圍棋，然後是蛋白質折疊。

現在，他們擔心在 2014 年，如果 Google 收購了這家公司，他們將對人工智能產生壟斷地位。他們擔心全球只有幾家最大的技術公司將控制人工智能生態系統。這就是他們在 2015 年創立 Openai 的主要原因。就在一年後的 2016 年，Openai 推出了 Jim Library。

一開始，Openai 專注於一個叫做強化學習的領域，而 Jim Library 則提供了一個易於使用的環境，用於強化學習。但幾年後，Openai 實際上宣布了 GPT 的第一個版本，GPT 代表生成式預訓練轉換器或預訓練轉換器，視您的時態而定。2018 年由於特斯拉當時正在開發自己的人工智慧系統，埃隆·馬斯克辭去了 Openai 的董事會席位，以免未來出現利益衝突。2018 年主要是為了自駕車，但現在特斯拉實際上正在一個叫 Optimus 的人形機器人開發中使用相同的人工智慧系統。

一年後，即 2019 年，Openai 實際上從非營利組織轉變為有限營利組織，以接受 10 億美元的投資。在此過程中與 Microsoft 合作，後者也是這個 10 億美元輪的首席投資者。現在，我應該指出，這個有限營利組織的想法基本上是，在獲得一定的利潤後，Openai 將不得不將這些錢捐贈或以一種不只是記錄在帳簿上的方式使用它。

同樣在 2019 年，Openai 宣布了一個名為 GPT-2 的新模型，實際上是 GPT 的下一階段，而 GPT-2 最初並沒有向公眾發布，原因是關於它可能在大規模上創造虛假的錯誤信息的安全擔憂，特別是因為即將到來的 2020 年選舉。

2020年，Openai 宣布推出 GPT-3，並於 2020 年 6 月宣布建立 API 以存取新的 AI 模型，申請者需通過審核才能存取該 API。Openai 非常注重安全和倫理，因此在釋出模型時比較謹慎，會設置相應的保障措施。2021年，Openai 宣布開發出 Dolly 模型，該模型能夠通過文本生成圖像。需要注意的是，最初的版本並不是開放源碼或通過 API 可用，主要是一篇關於如何從文本中生成圖像的研究論文。2022年，Openai 推出了 Dolly2，能夠通過文本提示生成更高保真度的圖像。同年，他們還宣布了 Chat 模型，這是一個針對人類反饋進行對話訓練的 GPT 或 GPT 3.5 的優化版本，我們稍後再談。這些模型和技術一直發展至今。2023年初，Openai 宣布 Microsoft 為 OPENAI 投資了新的 100 億美元。

在這份投資中，Microsoft和Openai的關係變得更加緊密。
這意味著Azure，Microsoft的雲服務，實質上是他們版本的亞馬遜Web服務或Google Cloud平台，他們實際上將成為Openai的獨家雲服務提供商。模型API調用。我應該注意到，API仍然可以直接從Openai獲取，這就是我們使用Azure上的Openai模型API調用。基本上，他們所說的獨家是指Openai模型不會很快出現在AWS或GCP上。我應該指出，那些其他雲服務提供商很可能正在開發自己的大型語言模型，未來也許會開發API，特別是如果Google使用它的Lambda模型。

從2015年到2023年，發生了很多事情，發展的速度是指數級的。

在這門課程中，我們盡力使討論盡可能具有未來性，特別是著重於Openai DirectX API。因此，我們將專注於直接連接到Open AI，創建一個帳戶並獲取密鑰，然後直接使用Open AI的API。因為在這次拍攝時，你實際上仍需要在Azure上申請才能通過Azure Cloud獲取訪問權限。我們還專注於將方法應用於未來模型。因此，當下一個版本的GPT推出或者可能是Dolly的下一個版本，我應該指出三個版本並未宣布或保證將作為獨立模型宣布。
也許Dolly Too只會變得更好。GPT四已經大致地被暗示。

然而，在本課程中教授的方法應適用於這些模型。因此，我們將教授你未來通用的思想和方法。希望這些模型能更好。

好的，我們根據 OpenAI 的歷史涵蓋了很多內容。如果你對學習 AI 的近期歷史感興趣，可以回溯到 Google、Facebook 和 OpenAI 自己的發展，以及企業競爭和在這個新領域發展的重要性。有一本非常好的書叫做《天才製造者》，我非常喜歡，從 Google 收購 DeepMind 的想法開始，探討了這一切如何從商業角度發展。

好的。所以我應該指出，我們在這門課程中使用的主要模型類型基於 GPT 和 Dolly。基本上，我們使用人工智慧來生成文本或圖像。
我們還會涵蓋其他模型，例如文本嵌入或代碼生成模型，或使用像 GPT 這樣的模型進行翻譯。因此，我們不僅僅涵蓋 GPT 和 Dolly，還包括其他模型。
但顯然，很多這些模型和想法都是基於 GPT 和 Dolly 生成文本和圖像。
因此，我們將首先對 GPT 進行高級概述，然後在稍後對 Dolly 進行高級概述，從根本上回答 GPT 如何運作的問題。如果您想了解細節，我們還將鏈接到解釋 GPT 和 Dolly 背後數學原理的論文。

但現在讓我們進入下一節課。

我們將進行一次 GPT 的高級概述課程，了解 GPT 是如何運作的。

## 英文

Welcome back to this lecture where we're going to give an overview of open AI, the company.
So if you've enrolled in this course, I'm sure you're probably familiar to some degree with Openai
the company, and you've probably heard about things like GPT, GPT and Dolly.
But before we dive into those amazing projects and things that we're going to be building with the Openai
Python APIs, I want to give you a brief history, an overview of the company itself so you can understand
its role in the larger technology and AI ecosystem.
Now something to note is things move super fast in the field of artificial intelligence, so the information
and history likely won't be completely up to date by the time you're viewing this.
To give you an idea of how fast things are moving, even additional APIs and funding for Openai, the
company were announced during the creation of this course.
So our history is going to take you all the way up to January 2023 after they closed their latest funding
round with Microsoft.
But let's start at the very beginning by taking a look at the major personas involved at the start of
Openai's creation.
So back in December 2015, at the end of the Neural Information Processing Systems Conference in Montreal,
Elon Musk and Sam Altman announced the creation of a new non-profit organization, and they called this
organization Open AI, as in developing research for artificial intelligence in an open way that will
be accessible to people, just like open source code is accessible.
Now let's talk about these two people, Sam Altman and Elon Musk.
Sam Altman dropped out of Stanford University to create a location based social network called Looped
in 2005, and later on it was acquired in 2012 for $43.4 million.
I should note here that looped actually raised somewhere around $30 million.
So it wasn't a huge exit, but clearly very impressive that Sam was able to build this company, especially
after dropping out of Stanford at 19 years old.
So then in 2011, Altman actually starts off being a part time partner at the Startup accelerator Y
Combinator.
And even just a few years later, he became the president of Y C, which is the shorthand way of saying
Y Combinator and Y Combinator super famous accelerator in Silicon Valley.
Lots of famous companies have come from this, like Airbnb and Dropbox.
So even though Openai was started in 2015, it actually wasn't until 2019 that Altman transitioned away
from being president of OIC to focus on Openai essentially full time.
And at the time of this filming, Garry Tan is now currently the president of Y Combinator.
And now Elon Musk.
So if you're enrolled in this course, I'm going to be very surprised if you don't know who Elon Musk
is.
But let's give you kind of a rapid background on Elon since he has so many companies that he's been
part of.
In 1995, he started ZIP two of his brother Kimbal Musk and a Greg Curry.
They sold to Compaq in 1999 for $307 million and Elon Stake was worth about 7 million at the time.
Then in 1999, Musk started an online financial services company, X dot com, where the vision for
x dotcom was actually to be kind of the world's online bank.
And anything you could do in a bank you want it to be able to do online.
Of course, this is during the dotcom boom in Silicon Valley and later in 2000 due to the amount of
money that X dotcom was burning as well as infinity was burning trying to sign up users, they merged
to form PayPal.
And then in 2002, PayPal was actually acquired by eBay for $1.5 billion.
And at this point, Elon Musk's stake was approximately worth $176 million.
So this is quite a large exit for Elon.
Now, what's a little crazy is he continues to kind of double down with all his earnings of these previous
companies to create more companies.
So in 2002, he also founded the company Space X, which, if you haven't heard of, is essentially
a rocket manufacturer and rocket launcher.
They launch satellites in space.
What's a little different about space is they develop rockets that can actually come back down instead
of just being destroyed in the atmosphere And later on Space X would actually develop its own satellite
internet via StarLink.
Then in 2004, Musk actually invested in Tesla.
Contrary to common belief, he's not technically a founder of Tesla.
He was one of the very first investors, and he actually became its largest shareholder, which allowed
him in 2005 to take a much more active role in Tesla and end up being the CEO.
Now, a lot of stuff has happened between then in 2022.
Must actually acquire Twitter for $43 billion.
I'll let you do a Google search on that if you're interested in kind of the process.
That was a little crazy and how he acquired that company.
And then I should point out there's many other companies he's been involved in.
He's been involved with SolarCity, which was actually later on acquired by Tesla.
He's been involved in the Boring company, which is a tunneling company, Neuralink as well.
He's kind of one of the founders, major investors of Neuralink, which works on essentially implants
that grab electrical signals from the brain and then can communicate that to other parts of the body.
StarLink, which is that satellite Internet arm, it's technically a subsidiary of SpaceX.
And then, of course, Openai, which as we mentioned, was started in 2015.
So quite a lot of companies that this man has been involved in.
Now, let's go back and shift our focus to open AI.
So there's also really major players that are major investors in open A at the start.
Reid Hoffman, founder of LinkedIn, Jessica Livingston, one of the founders of Y, C, Peter Thiel,
Infosys, Khosla Ventures and Y C Research.
So there's a lot of big players in Silicon Valley that were also at the start and open AI.
In fact, it started with about $1,000,000,000 in funding and capital.
There's also lots of early employees that names you may recognize or may not recognize.
You should note that these are basically all like A-plus players at the very start of artificial intelligence
in 2015 of open AI.
So they essentially hired kind of the who's who of AI and machine learning researchers, and there's
lots of other well regarded advisors that maybe weren't employed to open AI but were advising the company
on strategy.
Now, I did mention it was initially formed as a non-profit, mainly to safely develop artificial intelligence.
And one of the main reasons that Elon Musk and Sam Altman developed Openai was that in to 2014, which
was one year before Openai, Google actually acquired a company called DeepMind, which you may be familiar
with if you've ever seen things like the early development of DeepMind training and a machine learning
model to beat an Atari game.
And then later on they developed AlphaGo and other like Alpha Fold, basically machine learning models
to kind of conquer the chess game, the go game, and then later protein folding.
Now, they were concerned all the way back in 2014 that if Google acquired this company, they would
kind of have a stranglehold on artificial intelligence.
And they were worried that just a few of the world's largest technology companies would have control
over the AI ecosystem.
So that's mainly the reason they started opening AI in 2015.
Just a year later.
Now in 2016, Openai releases the Jim Library.
At first, Openai was highly focused on a field of study called Reinforcement Learning and Jim Library
allowed for an easy to use environment for reinforcement learning.
But then a couple of years later, open A actually announced the first version of GPT, which stands
for Generative Pre Training Transformer or Pre-trained Transformer, depending on your tense.
Now, in 2018, Elon Musk resigned his board seat at Openai due to potential future conflict of interest
because Tesla was actually developing its own AI systems at the time.
In 2018 it was mainly for self driving cars, but now Tesla actually operates the same AI systems in
their development on a humanoid robot, which they call Optimus.
A year later in 2019.
And this is the interesting part open.
I actually transitions from being a nonprofit organization to a capped for profit organization in order
to accept an investment of $1 billion.
Partnering with Microsoft in the process, who is also the lead investor of that billion dollar round.
Now, I should point out that this capped for profit organization, the idea behind it is basically
after a certain amount of profit, Openai will then have to either donate that money or do something
with it in a way that isn't just recording it as profit in the books.
Also, in 2019, Open AI announces a new model called GPT two, Essentially the next phase of GPT and
GPT two is actually not initially released to the public due to safety concerns regarding its ability
to possibly create false misinformation at a large scale.
Especially they were concerned because of the upcoming 2020 election.
In 2020, GPT three was announced and in June of 2020 Openai announced that it would have a creation
of an API to actually access as new AI models that it's been creating and initial users had to apply
to be accepted in order to have access to the API.
So a big part of Openai's kind of process and development is they take safety and ethics quite seriously,
so they're a little slower to release models due to this safeguards that they put in place.
In 2021.
A year later, Openai announces the creation of Dolly, which is a model capable of producing images
from text.
I should point out that the very first version of Dolly wasn't actually open sourced or available via
an API.
It was mainly just a research paper that kind of developed the idea of how can I take text and produce
an image from it?
And then in 2022, Dolly two is announced, which is able to create much higher fidelity images from
text prompts.
And also in 2022, chat is announced.
It's announced right at the end of December 2022, and it was actually an optimised version of GPT or
GPT 3.5 for dialogue trained on human feedback and we'll talk about that later on.
And that brings us to essentially the current timestamp of when we're filming this.
And at the start of 2023, Openai announced that Microsoft made a new $10 Billion investment for OPENAI.
And as part of this investment, the link between Microsoft and Openai has become a lot stronger.
And that means Azure, which is Microsoft's cloud service, essentially their version of Amazon Web
Services or Google Cloud platform, they are actually going to become the exclusive cloud provider for
openai.
Model API calls.
And I should note that the API is still available directly from Openai, which is what we use The Openai
model API calls on Azure.
Basically what they mean by exclusivity is you're not going to see the Openai models on AWS or GCP anytime
soon.
I should point out that it's very likely that those other cloud providers are working on their own large
language models and maybe they'll develop APIs for them in the future, especially if Google with its
Lambda model.
So clearly a lot has happened in just a short time span from 2015 to 2023, and the pace of development
is exponential.
So we've tried our best to make the discussions in this course as futureproof as possible, especially
by focusing on the Openai DirectX API.
So instead of using Azure, we're going to focus on linking directly to open AI, creating an account
there and getting the keys and then just using the API directly with open AI.
Because at the time of this filming, you technically still need to apply on Azure to get access via
Azure Cloud.
We've also focused on methodologies that will apply for future models.
So when the next version of GPT comes out or there's maybe a next version of Dolly, I should point
out three hasn't been announced or anything, and it's not guaranteed to be announced as a separate
model.
Maybe Dolly too just keeps getting better.
GPT four has been kind of roughly hinted at.
The methodologies we teach in this course, however, should apply to those models.
So we're going to teach you ideas and methodologies that are essentially futureproof in that manner
as well.
It's just that the models will be hopefully better.
Okay, So we covered a lot just based off the history of Openai.
If you're actually interested in learning about the recent history of AI, going back to maybe Google
and Facebook's and Openai's own developments and kind of also the corporate competition and developing
in this new space and how important it is.
There's a really good book called Genius Makers that I really enjoyed that kind of starts from this
idea of Google acquiring DeepMind.
And it's a great read in case you're interested in how this is all developing from kind of a business
standpoint.
All right.
So I should point out, the main types of models we use in this course are based on GPT and Dolly.
Essentially, we're either going to generate text or generate an image using AI.
We also do cover other models like text embedding or code generation models or using something like
GPT for translation.
So we don't just cover GPT and Dolly, we also cover other things.
But clearly a lot of these models and ideas are based on just GPT and Dolly generating texts and generating
images.
So what we're going to do is we're going to get a very high level crash course in GPT and then later
on a very high level crash course in Dolly, essentially answering the question how does GPT work kind
of at a higher level.
If you want the kind of nitty gritty details, we will also link to the papers that explain the mathematics
behind GPT and Dolly.
But let's go to the next lecture.
We're going to have a crash course on how does GPT actually work.
Autoscroll
COURSE CONTENT
OVERVIEW
Q&AQUESTIONS AND ANSWERS
NOTES
ANNOUNCEMENTS
REVIEWS
LEARNING TOOLS
0:18
Styles
1000
CancelSave note
All lectures
Sort by most recent

###### tags: `Udemy`,`OpenAI Overview`