There are more and more high-profile AI models that didn’t come out when they were supposed to.
Elon Musk, the founder and CEO of the AI company xAI and a billionaire, said last summer that Grok 3, xAI’s next big AI model, would be ready by the “end of the year” 2024. Grok, xAI’s answer to models like OpenAI’s GPT-4o and Google’s Gemini, can look at pictures and answer questions. It also runs some parts of Musk’s social network X.
Musk wrote in a July post on X, “Grok 3 should be really something special by the end of the year after training on 100,000 H100s.” He was referring to xAI’s huge cluster of GPUs in Memphis. In a post in the middle of December, he said, “Grok 3 will be a major leap forward.”
But it’s January 2, and Grok 3 still hasn’t come out, and there aren’t any signs that it will soon.
In fact, AI spy Tibor Blaho found code on xAI’s website that makes it look like an intermediate model called “Grok 2.5” might land first.
It’s true that Musk has set big goals before and failed to reach them. It is well known that Musk’s predictions about when products will come out are, at best, not accurate.
Also, to be fair, Musk did say that Grok 3 would “hopefully” be out in 2024 “if we’re lucky” in a chat with podcaster Lex Fridman in August.
Now Grok 3 is missing, which is interesting because it’s part of a bigger trend.
Anthropic, an AI company, failed to make a follow-up to its top-of-the-line Claude 3 Opus model last year. A few months after saying that the Claude 3.5 Opus, a next-generation model, would be out by the end of 2024, Anthropic took the model out of all of its developer documents. One report says that Anthropic finished training Claude 3.5 Opus sometime last year, but they chose not to release it because it wouldn’t make money.
It is said that Google and OpenAI have also had problems with their top models in the past few months.
This could be proof of the flaws in the current AI scaling rules, which are the ways companies are trying to make their models smarter. It wasn’t too long ago that training models with huge amounts of computer power and bigger and bigger data sets could lead to big improvements in performance. But the gains from each new generation of models have started to go down, which is pushing companies to look for other ways to do things.
In the talk with Fridman, Musk himself made a reference to this.
“You think Grok 3 should be cutting edge?” Fridman told him.
“I hope so,” Musk responded. “I mean, this is the point.” This goal might not work out. And that’s the goal.
Also Read: Elon Musk’s X is Changing Its Privacy Rules So That Other People Can Use Your Posts to Teach Ai
Grok 3 might be late for other reasons as well. Some of xAI’s competitors have much bigger teams than it does. Whatever the case, the pushed back start date is more proof that the usual ways of training AI are hitting a wall.
What do you say about this story? Visit Parhlo World For more.
