OpenAI 19 Facts You Didn't Know?

Open AI is one of the fastest growing companies in the world, and here are 19 surprising facts.

First, despite the huge success of Open AI, it only has 375 employees, which is a bit crazy relative to the company's achievements. CEO Sam Altman said, "I shouldn't be bragging about Open AI, but this density of talent is unprecedented in the tech industry." 375 people is certainly not a small company, but the number is really surprising compared to Google or Metastatic.

Second, Microsoft holds a 49% stake in Open AI and invested a total of $11 billion. Not only that, but before the $11 billion is paid off, Microsoft will also receive 75% of Open AI's profits. This is a win not only for Microsoft, but also for Open AI, because they have access to Microsoft's supercomputers and have $11 billion in funding.

Third, Greg Brockman is no longer the CTO, but now Mira Murati, an Albanian AI engineer. Murati began his career at Goldman Sachs and then spent three years at Tesla. In 2018, she joined Open AI and most recently became Chief Technology Officer, specializing in robotics and product development.

Fourth, a few years ago, Open AI created a team of five AI bots to master the game Dota 2. These bots are so good that they can beat even the best esports teams in Dosta. These robots learn by playing games with themselves every day. If you don't know Dota 2, this achievement may not sound as impressive, but in reality the game is much more complex than chess or Go, which is an important milestone in the development of artificial intelligence.

Fifth, Open AI is not actually open, and the company has essentially shut down the source code. Elon Musk came up with the name Open AI, which means open source. However, now the situation is different. One of the key reasons for OpenAI's transformation is security considerations, as casually releasing these powerful AI models, such as GPT4, can be seen as irresponsible and unsafe. However, Microsoft's investment is also a reason. When Microsoft invested in the company, they wanted exclusive access to GPT3, which was the most powerful model at the time, and OpenAI granted it to them, signaling that OpenAI was no longer open source.

Sixth, the free version behind the GPT3 model, ChatGPT, actually existed two years before ChatGPT was released. This highlights the complex and time-consuming process of fine-tuning the model. The reason why GPT3 was not a huge success when it was first released was that it was not easy to use. ChatGPT is very easy to use and can be used as long as you can use a computer, but GPT3 can only be used through APIs, and most people don't even know what an API is, let alone how to use it.

Seventh, Open AI currently has seven board members and is not democratically elected. Members include Shivan Zillis, Matthew Mohr, Alejandro Cortujo, Jade Long, Kelly Sims, Helen Toner and Atlas Benjalo. Interestingly, Shivon Zillis is actually the mother of Elon Musk's twin brother, so Elon may still have some influence on Open AI.

Eighth, Elon Musk terminated Open AI's access to Twitter data because he believed the company was underpaid. OpenAI pays Twitter $2 million a year for access to data, but Elon doesn't think that's enough. After the release of GPT, Elon strongly criticized Open AI, saying that chatbots are biased and even lying. Since then, Elon has accused OpenAI of being a profit-maximizing company and signed a letter calling for a six-month moratorium.

Ninth, Open AI's top AI researchers are paid more than some professional athletes. Even during the nonprofit period, many top researchers earned more than a million dollars. Back in 2016, OpenAI's chief scientist, Ilya Sutskever, earned $1.6 million, even though the company was still nonprofit.

Tenth, the training cost of GPT4 is more than $1 billion, which is much higher than the training cost of GPT3 (only $120 million). Yes, those supercomputers are not cheap, but exactly how the $1 billion will be distributed is not clear. There is speculation that only 40% of the computational cost, or $400 million, is used, with the remaining 60% going to the salaries of researchers and engineers.

Eleventh, we still don't actually know how many parameters GPT4 has. GPT4 is reportedly about six times the size of GPT3, which has 17.5 billion parameters, meaning GPT4 may have more than 100 billion parameters. Experts estimate it at around 100 billion, but no one at OpenAI has confirmed or denied that number.

Twelfth, one of Open AI's first successful projects was Dactyl, a robotic hand that eventually learned to solve the Rubik's Cube, a major achievement in robotics. Dactyl actually learns from scratch, using the same reinforcement learning algorithm as the OpenAI five-person Dota 2 robot. Dactyl shows that it is feasible to train agents in simulations and transfer them to real-world tasks.

Thirteenth, Reid Hoffman, one of OpenAI's original investors, recently resigned from the board. The billionaire LinkedIn founder said this was to avoid a potential conflict of interest, which is why Elon left (at least officially). Reid himself said, "By stepping off the board, I can proactively eliminate any potential downstream issues between OpenAI and all of the Greylock portfolio companies I support."

Fourteenth, in November 2022, the same month that ChatGPT was released, Microsoft, OpenAI, and GitHub suffered class-action lawsuits alleging that the code created by GPT violated the rights of programmers. The person who filed the lawsuit was Matthew Butterick, a programmer and lawyer. ChatGPT relies on an unprecedented scale of software privacy, he said. Butterick then assembled a legal team to file a second class action lawsuit against the three companies. The three companies sought dismissal of the lawsuit. Microsoft actually says the complaint fails on two inherent flaws: lack of damage and lack of visible claims, meaning that the programmer was clearly not harmed by ChatGPT and didn't have any real claims, at least according to Microsoft.

Fifteenth, OpenAI is not actually a for-profit company, but a so-called limited profit company. This is a very unique and unusual corporate structure that is not seen anywhere else. The limited profit structure is more like bonds than equity. Owning equity means you can sell at any time and make a profit. Owning a bond is more like buying it in anticipation of a return for the same amount you put in plus interest, but disappears once paid out. This unique corporate structure is designed to ensure that all obligations are paid to investors and employees, after which OpenAI will become non-profit again, at least by design they want us to believe.

Sixteenth, according to Ilya Sutskever, OpenAI's main KPI is technological advances in AI models. This means not just increasing the number of parameters, but also really understanding the model and how it works. That means better data and more control, better predictions of the model's outcome, and as Sam Altman says, you can know at the beginning of training how smart the model will be, which is kind of crazy.

Seventeenth, OpenAI has said that the era of easy scaling in AI models is over, which means that more attention will now be paid to data than just adding more internet parameters. This begs the question: How does OpenAI get all this high-quality data? Make no mistake, larger neural networks will still continue to improve, but so will the cost and effort required to improve them. One potential solution is the concept of synthetic data, where AI creates its own data. Because GPT4 is already very smart, you can use GPT4 to create new books, new articles, and even new images. Maybe with some reinforcement learning human feedback or human supervision, you can filter it to see if the image is junk or if the book is really bad, and the book could be a bestseller. Now, if AI starts generating data (so-called synthetic data), we can basically have unlimited data at our disposal, which will solve the data problem once and for all.

Graylock partner Greg Brockman publicly stated that Elon Musk's departure was a huge failure for OpenAI. Unfortunately, he did not elaborate on the reasons, but it is clear that the loss of Elon was a mistake, since Elon left voluntarily, mainly because of money. Elon is one of the richest people in the world. He pledged to donate $10 billion to OpenAI, but only donated $100 million in total, because he again believes that the company is not moving in the right direction and believes that OpenAI is lagging behind Google in AI development. There are many speculations as to the actual reason Elon left. Some say it's just because he's overworked to manage Tesla, SpaceX, SolarCity, the board, and so on. Others said he wanted to take control of OpenAI and become CEO, but the likes of Sam Altman, Ilya Sutskever, and Greg Brockman wouldn't let him do that, so he walked away and took the money he promised.

Nineteenth, you might be wondering how much money OpenAI actually makes, whether it's profitable, and how long it will take to repay Microsoft's $11 billion investment. According to anonymous sources cited by Reutez, in December 2022, OpenAI expected 2023 revenue of $200 million. By 2024, revenue is expected to be $1 billion. This prediction has two important premises: it comes from December, meaning that ChatGPT has only been released for a month, and today it is more popular than December. As a result, projected revenue could be well above $200 million.

All in all, OpenAI is a very promising and exciting company with some of the brightest AI experts and modeling capabilities. Still, the way the company operates is puzzling, claiming to be non-profit, but apparently committed to making a profit and paying back the investment. It's also no longer open source, which runs counter to the original vision. However, as long as OpenAI continues to push the boundaries of artificial intelligence and make a positive impact on the world, these contradictions are justified. The field of artificial intelligence needs leaders, and OpenAI is one of them.

#2023高考季 #