Interview with Altman: It is expected that there will be several different versi

2024-06-18

During my brief meeting with Sam Altman, there were many moments that gave me a deeper understanding of the worldview of this OpenAI CEO.

The first moment was when he pointed at my iPhone SE and said, "This is the best iPhone." However, the moment that revealed his clearer vision was when he described how AI tools would integrate into our daily lives better than smartphones.

"You really want," he told MIT Technology Review, "just this tool that can really help you."

At that time, Altman was in Cambridge, USA, attending a series of events hosted by Harvard University and the venture capital firm Xfund. He described the killer application of AI as "an extremely capable colleague who is well aware of my entire life, every email, and every conversation, but it feels more than just an extended application to me."

He said it could solve some simple tasks immediately, and for more complex tasks, it could try, but if necessary, it would ask you questions.This is a leap compared to OpenAI's current products. Its most advanced applications, such as DALL-E, Sora, and ChatGPT, have left us in awe with their ability to generate convincing text and hyper-realistic images and videos.

However, compared to the future killer applications, Altman calls them "incredibly stupid."

These applications are still independent tools we use for individual tasks, and they struggle to accurately understand us during our conversations with them.

Advertisement

In Altman's view, in the new paradigm, artificial intelligence will be able to assist us beyond the chat interface and will help or take over our work in the real world.The Future of AI Hardware

 

I asked Ultraman if we need a new kind of hardware to realize this future. Although smartphones are extremely powerful and their designers have integrated many AI-driven features, some entrepreneurs believe that future AI applications will require a specially customized device.

 

Some of these so-called "AI devices" have already begun to emerge.

 

Humane's wearable AI Pin has hit the market, but the response has been poor. Ultraman is an investor in the company, but he is not the driving force behind the device.

 

It is rumored that he will also collaborate with former Apple designer Jony Ive to develop a new type of hardware.But Altman said that we might not need custom devices at all. "I think this doesn't require new hardware," he told me, "the envisioned types of applications might exist in the cloud."

But he quickly added that even if this shift in the paradigm of artificial intelligence does not require consumers to buy new hardware, "I think you would be happy to have (new devices)."

Although Altman finds artificial intelligence hardware devices exciting, he also hinted that he might not be well-suited to challenge this field: "I am very interested in consumer hardware for new technologies.

But I am just an amateur who likes it, which is far from my expertise."Searching for Training Data

After hearing his vision for a powerful AI-driven agent, I wondered how this would address the current lack of training data in the industry.

To build GPT-4 and other models, OpenAI scraped content from internet archives, newspapers, and blogs as training data, as the scaling laws have long indicated that the larger the model, the better the performance.

AlphaFold3 is here! No need to input structural information, the accuracy of bio
Scientists develop high-entropy alloy nanoparticles that can simulate the reduct
Academician Yang Xueming jointly develops a new method for controlling the stere
Scientists formulate a theoretical framework for collective intelligence, helpin
The Shanghai Jiao Tong University team develops general artificial intelligence
The Southern University of Science and Technology team achieves non-diffractive
Northeastern University in the United States proposes a video data enhancement m
Xi'an Jiaotong University creates a sound-suspended bubble that exists for more
Scientists achieve controllable degradation of polystyrene, with the molecular w
Scientists synthesize a new functional group that can safely store hydrogen fluo

However, finding more training data is an increasingly daunting problem. Much of the content on the internet has already been scraped, and obtaining private or copyrighted data would plunge developers into legal disputes.

Altman is optimistic that this will not be a problem for a long time, although he did not specify the details.He said: "I believe, but I'm not sure, that we will find a way to solve this problem. You always need more and more training data. Humans are a great example that shows we have other methods that can be used to (train intelligence). I hope we can find it."

Who will create general artificial intelligence?

For a long time, OpenAI's core vision has been centered around the pursuit of general artificial intelligence, or an artificial intelligence that can reason like humans or with greater reasoning ability than humans.

OpenAI claims its mission is to ensure that this technology "benefits all of humanity." However, it is far from the only company pursuing general artificial intelligence.So, in the competition for general artificial intelligence, what is the most important tool? I asked Altman whether he thinks the player with the most chips and computing power will ultimately be the winner.

Altman believes that there will be "several different versions of general artificial intelligence, each good at different tasks." He said, "I think you have to exceed a certain computational threshold. But even so, I wouldn't say I'm sure (that having more computing power will lead to success)."

When can we see GPT-5?

Did you think he would give a clear answer? When another reporter in the room asked Altman if he knew when the next version of GPT would be released, he calmly replied, "Yes."Then he smiled, and said nothing more.

Leave a comment