No, it’s money. Running a smart AI uses much more resources, so they’re trying to keep consumer AI as cheap and dumb as they can get away with and put the smarter answers behind a paywall.
I heard something like OpenAI (or Anthropic?) tweaked one of their AI’s to do well on a certain benchmark test and while they did have the highest score they couldn’t publish the model because it used like $20,000 in electricity alone to make sure it’s answer was correct.
It is also a great example of how low people’s expectations are of tech they have spent around a trillion dollars on to get something that is about as good as asking a random person.
“People get the puzzle wrong!”
We don’t spend a trillion dollars for someone to fail a simple logic puzzle.
No, it’s money. Running a smart AI uses much more resources, so they’re trying to keep consumer AI as cheap and dumb as they can get away with and put the smarter answers behind a paywall.
I heard something like OpenAI (or Anthropic?) tweaked one of their AI’s to do well on a certain benchmark test and while they did have the highest score they couldn’t publish the model because it used like $20,000 in electricity alone to make sure it’s answer was correct.
The answer was probably something like:
“…Yes! You do need eggs to make an omelet!”
“You need to drive your car to the car wash.”
For context, this is a real question that stumped a lot of LLMs:
“I want to wash my car. The car wash is 100 meters away. Should I walk or drive?”
Of course it’s famous now, so they have added specific exceptions for it (or just because it’s famous they’ve scraped the solution)
It is also a great example of how low people’s expectations are of tech they have spent around a trillion dollars on to get something that is about as good as asking a random person.
“People get the puzzle wrong!”
We don’t spend a trillion dollars for someone to fail a simple logic puzzle.