Sam says that Katis GPT-5 made a mistake

The launch of Openai August From the large GPT-5 language model it was somewhat a disaster. There were defects during the live broadcast, with the model generating plans with a clear inaccurate numbers. In Reddit Ama with Openai employees, users Complain The new model was not friendly, and called on the company to restore the previous version. The most important of all, critics arrested this GPT-5 It was less than the predictions of the strategy that Openai was the era of juice for years. She promised as a game change, the GPT-5 may have played the game already better. But it was still the same game.

The skeptics currently seized the end of the mutation of artificial intelligence. Some even predicted the beginning of the winter of artificial intelligence. “GPT-5 was the most artificial intelligence system in all ages,” Gary Marcus told me full-time during his schedule of victory rolls. “It was supposed to present two things, AGI and PHD at the PhD level, and none of these were presented.” Moreover, he says, the dull model apparently is evidence that the Openai ticket to AGI – which leads to increased data and chips groups to make its systems significantly smarter – cannot be installed. For one time, Marcus’s views were echoed by a large part of the artificial intelligence community. In the days after the launch, GPT-5 looked like a copy of the new Coke.

Sam Al -Taman does not get it. A month after its launch, he wanders in a conference hall at the company’s new headquarters in the Mission Bay neighborhood in San Francisco, keen to explain to me and my colleague Kylie Robson that GPT-5 is all he was telling, and that’s everything in his epic pursuit of AGI. “It was somewhat bad feelings at all,” he confessed. “But they are now great.” Yes, great. It is true that criticism has faded. In fact, the company The latest version From the mind bending tool to create an impressive artificial intelligence, the narration has turned from the first disappointing GPT-5 appearance. Nevertheless, the message from Altman is that those who reject the wrong side of history. Insist on the trip to AGI, is still on the right track.

Numbers game

Critics may see GPT-5 as a party retreating to the summer of artificial intelligence, but Altman and Team argue that it enhances artificial intelligence technology as an indispensable teacher, a source of information about the search engine, especially an advanced collaborator of scientists and programmers. Altman claims that users have started to see him on his way. “GPT-5 is the first time that people are,” the holy curse. It does this important piece of physics. ”Or the biologist says,“ Wow, this really helped me to determine this thing. ”There is an important thing that did not happen with any model before the GPT-5, the beginning of Amnesty International that helps to accelerate the discovery rate of the new flag.” (Openai was not cited about those who are these physicists or biologists.)

So why are the first lukewarm reception? Altman and his team took several reasons. They say that this is that GPT-4 hit the streets, the company presented versions that were itself transformational, especially the advanced thinking modes that they added. The jump was 4 to 5 greater “From the jump from 3 to 4, we had a lot of things all the way.” Altman says. “I am not shocked because many people had a reaction (disappointing), because we see our hands.”

Openai also says that since GPT-5 has been improved for specialized uses such as science or coding, ordinary users take some time to estimate its virtues. “Most people are not physicists,” Altman notes. As explained by Mark Chen, the head of Openai Research, unless it is like mathematics, you will not care much that GPT-5 ranked first of mathematics Olympics, while the regime was ranked last year 200.

As for the charge on how GPT-5 appears that the scaling does not work, Openai says that this comes from misunderstanding. Unlike previous models, GPT-5 did not get its main progress from a large larger data collection and tons of more account. The new model got gains from reinforcement learning, a technique that relies on expert humans who give it. Brockman says Openai has developed its models to the extent that they can produce their own data to run the reinforcement learning cycle. “When the model is stupid, all you want to do is to train a larger version of it,” he says. “When the model is smart, you want to take samples of it. You want to train in its own data.”