Sam Altman Says the GPT-5 Haters Got It All Wrong

August launch The large language model of its GPT -5 was somewhat disastrous. During livestream was glits, the model obviously formed the chart with the wrong number. OpenAI employee, with user in a Reddit AMA Complaint The new model that was not friendly, and called on the company to recover the previous version. After all, the critics grabbed it GPT -5 OpenAI is less than the stratospheric expectations that have been juice for years. Promised as a Game Changer, GPT -5 may probably play the game better. However it was still the same game.

The skeptics occupied at the moment to promote the end of the AI ​​Boom. Some even predicted the beginning of the other AI winter. “GPT -5 was the most hypeod AI system of all time,” Fulltime Bubble -Popper Gary Marcus told me during the packed Shidiul of Vijay Laps. “It was supposed to provide two things, AGI and PhD-level knowledge, and it did not provide any of them.” What is more, he says, the seemingly unnecessary new model proves that the ticket in Openai’s AGI is scaling to make its systems smartly smart – and cannot be bribed. For once, Marcus’s views were echoed by a large part of the AI ​​community. In the next days, GPT -5 looked like II’s new coke version.

Sam Ultman is not doing that. A month after the inauguration, he is traveling to a conference room at the company’s new headquarters at San Francisco Mission Bay, interested in me and my colleague Kyley Robissis that GPT -5 is forcing everything he was saying, and his epic exploration for the AGI. He acknowledged, “The siblings were bad at the time of launch.” “But now they are great.” Yes, GreatThe It is true that criticism is dead. Indeed Recent revelation A Mind -Banded Tool to produce impressive AI video OP is removed from the disappointing GPT -5 debut. Although Ultman’s message is that the heroes are on the wrong side of history. He emphasized that the AGI journey is still on the track.

Number of games

Critics will see GPT -5K AI as summer ends, but Ultman and Team have argued that it is an essential home teacher, search engine -killing source, and especially for scientists and coders. Ultman claims that users have begun to see his way. “GPT -5 is the first time where people are, ‘Holy Choda It’s doing this important part of physics.’ Or a biologist says, ‘Wow, it really helped me find out this thing,’ ‘he said. “There are some important events that have not happened to a pre -GPT -5 model, which is the beginning of AI that helps to accelerate the rate of new science discovery.” (Opeena did not quote these physicists or biologists.)

So why is TPED initial reception? Ultman and his team have revealed several reasons. One, they say, since GPT -1 hit the street, the company provided versions that they themselves were converters, especially sophisticated rational modes they added. “There was a jump from 4 to 5 Great Than to jump from 3 to 4, “Ultman says.” We only had a lot of things on the way. “OpenAI’s president Greg Brockman agrees:” I am not shocked that it was for many people [underwhelmed] The response, because we showed our hands. “

Open is also said that since GPT -5 is favorable for special use such as science or coding, everyday users are taking some time to appreciate its qualities. “Most people are not physics researchers,” Ultman observed. Mark Chen, the head of the Openai research, explained that if you do not give yourself mathematics, you will not care too much of the GPT -5 rank in the top five of the Mathematics Olympians, where the system is ranked in the top 200 last year.

GPT -5 shows how scaling does not work shows that as a charge, Openai says that it comes from a misunderstanding. In contrast to previous models, GPT -5 did not get its main progress from a larger large datasate and more calculations. The new model has received its profits from reinforcement learning, a technique that reacts to the expert. Brookeman says that Openai made its models at a point where they could produce their own data to strengthen the reinforcement learning cycle. “When the model becomes dumb, all you want to do is train a big version of it.” “When the model is smart, you want to sample from it you want to train your own data.”