A brief background of the important news
Last week, OpenAI released a new multimodal large model, GPT-4o, which can be said to be highly acclaimed. However, less than two days after the release of GPT-4o, Ilya Sutskever, co-founder and chief scientist of OpenAI, announced her departure.
The news has sparked speculation about disputes and turmoil within OpenAI. There have even been claims that OpenAI has a "secret" severance agreement where departing employees must sign an agreement that does not discredit OpenAI. For example, if OpenAI illegally collects video data to train AI, it will lose all its equity.
On May 19, Sam Altman posted a lengthy statement about the matter, saying that he had never taken back the equity that anyone had acquired, and that they would not do so if people did not sign a separation agreement.
However, the departure of Ilya Sutskever remains in the spotlight, with many believing that there are hidden secrets and conflicts.
So, what is the essential reason for the departure of OpenAI's chief scientist?
The following is Wang Yuquan's important news commentary:
Recently, OpenAI released GPT-4o, which has attracted a lot of attention. Today I'm not going to talk about it, there has been a bit of negative press about OpenAI recently, and its co-founder and chief scientist Ilya Sutskever has left.
This guy was actually very important, on the one hand he was the co-founder, and on the other hand he was called the chief scientist at the time.
I want to explain a little bit, we have made a lot of introductions before, the advanced technology companies in the United States, we call it the "dual-long system", must have a good CEO, and at the same time have a good chief scientist, but sometimes it is part-time.
Why? Because I'm a scientist, I'm doing research in a university, and I'll just give you the research results, and you can do the technology transformation.
Note that the person who does technology transformation in a company is not the chief scientist, but two people, one is the CEO, of course, and the other is who? It's the CTO, it's the CTO. This time, the speaker of OpenAI's GPT-4o release was called Mira Murati, and her title was CTO.
But why did OpenAI have to get a chief scientist back then? The reason is very simple, it is because OpenAI did not achieve a breakthrough back then, and its lofty goal in the future is to achieve AGI general artificial intelligence. What does it entail? We need to continue to make breakthroughs in scientific research, which means that this is a new trend now.
In fact, since the 80s of the 20th century, after the Bayh-Dole Act in the United States, a large number of advanced technology companies have taken ready-made technologies from universities for productization.
However, when we talk about this round of artificial intelligence revolution, many companies believe that they still need to continue research and development. Including a recent news, Li Feifei joined a start-up company, why?
Now many people think that artificial intelligence still needs to have a breakthrough in basic research, but a breakthrough in basic research costs money, and universities cannot support this matter, and they need to invest heavily in enterprises to support it.
It is easy for a star expert like Li Feifei to enter a company, and she will get a lot of money, and she can continue her research. And many people also believe that artificial intelligence still needs to be studied in the next step, and general artificial intelligence has not yet come.
So you can see why when OpenAI was founded, the chief scientist was even more important than the CTO.
Why are you leaving your job again? We all know that GPT has just broken through 3.5, bringing surprises to everyone, GPT can talk to people, and slowly people feel that the interlocutor can't tell whether it is a human or a machine, plus the current GPT-4o, even the voice semantics can talk directly, and you can also add vision to do all kinds of interactions, it's really the same as people.
There are already so many advanced scientific research achievements, and everyone is still expecting to further achieve general artificial intelligence, but how did the chief scientist go?
Notice that one of the interesting things about this, which most people probably don't observe, is what? These two breakthroughs of OpenAI are precisely not breakthroughs in scientific principles, but breakthroughs in engineering.
Many people are saying that GPT is so great, it can be said to be a great event in human history, and many people think that this robot is at least intellectually and human, at least special intelligence and people have a fight.
This most basic theory can be traced back to Geoffrey Hinton, which is called convolutional neural networks. And then the next key step is actually proposed by Google, which is Transformer.
GPT is based on the Transformer architecture, GPT has not made a breakthrough in the Transformer architecture, but how can it be more eye-catching than Google, and how can it be so attractive?
It's an engineering breakthrough, it's a miracle, I get bigger, I spend more money on more GPUs, I train with a larger amount of data, and as a result, all the architecture remains the same, and intelligence emerges.
At this time, do you say that it is a breakthrough in scientific research? So you'll find that this thing is also confusing a lot of scientists, and it's the engineers who make the results, and the scientists explain why you made the leap. Shouldn't this leap be the contribution of scientists? You see, it's the other way around now.
And GPT-4o repeats the same old path. We say what's so great about GPT-4o right now? It is called multimodality, and multiple large models can be fused. However, the so-called multiple large models, not each large model architecture is different, the architecture is very similar, but the tasks are different and deal with different things.
And it solves the problem with a thing that is Transformer, and even behind the Transformer is a unified language called token. I'll talk a little more in our science and technology training camp.
So let's look at Ilya Sutskever now, what is most likely the reason for his departure? Of course, as the chief scientist, he hopes to make a breakthrough in the architecture, which will lead to advanced results and significant performance improvements. As a result, I didn't expect that every performance improvement is an engineering breakthrough, which is the work of the CTO and the CEO. But either way, the structure hasn't changed much.
In other words, this time GPT-4o is still the same, that is, using Transformer, using multi-modality to solve the problem of multi-sensory fusion that everyone thought could not be solved before, and the solution is so beautiful that now emotions can be perceived and expressed when human-computer interaction. Not only does it know my emotions, but it can express them, and even those emotions can be adjusted at any time.
You see that emotions and language have been perfectly integrated, so this perfect fusion used to be difficult, but now I find that it is actually very simple, that is, multi-modal synchronous training, that is, the same mechanism is used to deal with it, or Transformer, the structure has not changed. If you're the chief scientist, can you stay?
Therefore, it is very likely that Ilya's departure is due to his discovery that the future development may be engineering breakthroughs, because engineering breakthroughs have caused significant performance improvements, and even emergent phenomena, which scientists can only explain, not contribute. As a scientist, he definitely didn't want to explain, so he had to change places and make contributions.
You see OpenAI this time, on the one hand, Ilya left, on the other hand, the CTO gave a speech and launched the amazing GPT-4o, behind which there is a profound principle of the operation of the world.
If you are interested, join our science and technology training camp to explore the principle of technology changing the world.
In addition, from June 24th to 29th, we will organize the "Outpost Japan Longevity Science and Technology Expedition", I will accompany you to Japan throughout the whole process, and we will analyze and summarize together in the process, and I will also share some of my personal meditation methods.
For more information, please scan the QR code below. The number of places in the event is limited, so if you are interested, please seize the opportunity to sign up.
In addition, this year is the 8th anniversary of the science and technology training camp, and the friends of the innovation map have prepared the 8th anniversary benefits for everyone, from May 1st to May 31st, that is, if you buy a member of the science and technology training camp this month, you can get two months of school hours.
Interested friends should seize the time to scan the QR code in the event poster below, sign up, and let us understand the future together in the science and technology training camp.
The above is today's content, more detailed industry analysis and underlying logic, I will share it in the science and technology training camp. Welcome to pay attention to the global outlet WeChat account and sign up.
Wang Yuquan has to be heard, and we'll see you tomorrow.