Progress is not linear and AI is no exception.
Scholars, independent developers, and the world's largest tech companies have come closer to artificial general information, a still hypothetical form of intelligence consistent with human capabilities, and have collided with several obstacles. Many emerging models are prone to hallucinations, misinformation, and simple errors.
Google CEO Sundar Pichai called this AI stage Aji or “artificial judge intelligence” in a recent episode of Lex Fridman's podcast.
“I don't know who used it first, maybe Karpathy did it,” Pichai said. He mentioned Deep Learning and Computer Vision Specialist Andrej Karpathy, who co-founded Openai before leaving last year.
Horse horse mackerel is a small factor for the trajectory of AI development. Jagged is marked at once by sparks of genius and basic mistakes.
In a 2024 X post entitled “Jagged Intelligence,” Karpathy described the term as “the words I came up with were words that I came up with to describe the (odd and unintuitive) fact that cutting-edge LLMS can perform very impressive tasks (such as solving complex math problems) while still being able to fight very stupid problems at the same time. He then posted an example of a state-of-the-art, large-scale language model that he can't understand that 9.9 is bigger than 9.11, making “non-sensical decisions” in the game of TIC-TAC-Toe, and struggles to count.
The problem is that unlike humans, “many knowledge and problem-solving skills are all highly correlated, and everything from birth to adults is improved linearly,” AI's jagged edges aren't always clear or predictable.
Pichai repeated the idea.
“You can see what they can do and you can just slightly find them making numerical errors, counting R with strawberries, etc. “We feel like we're in a dramatic progression, some things don't work out, but overall there's a lot of progress.”
In 2010, when Google Deepmind was launched, the team spoke about AGI's 20-year timeline, Pichai said. Google then obtained DeepMind in 2014. Pichai thinks it will take a little longer, but by 2030 he emphasizes that “it doesn't matter what the definition is, as there are amazing progress on many dimensions.” ”
By then, the world needs a clear system to label AI-generated content to “distinguish reality,” he said.
“Progress” is an ambiguous term, but Pichai speaks at length about the benefits seen from AI development. At the UN's future summit in September 2024, he outlined four specific ways in which AI can advance humanity. This means improving access to native language knowledge, accelerating scientific discovery, reducing climate disasters, and contributing to economic advancement.
But first, you need to learn how to spell “strawberry.”
