We are definitely at the "try to break the model" phase. This is a super important phase with any emerging tech I want to actually use -- I have to make sure I fully understand what it CAN'T do, then figure out what I can do with it.
Of course, the things we can't do tend to get lower over time, too!
Yeah. Understanding the data which you're asking Code Interpreter to run analyses on is critical. You can't just assume that it understands the data in the way that an informed human would understand the data.
Great example, and important. Both a reasonable wrong request wording, and a far better request wording. Total # of CA drivers ( over many years) at 576 mln , is less good than total # of CA drivers in 2018. 27 mln is the (more?) right answer. More folks should guess first. I would have guessed 3/5s - 3/4s or up to 4/5s of CA pop of 36 mln.
Smart folk willing to estimate first, then check more carefully if the AI answer is outside the range, seem more likely to avoid wildly wrong answers.
We are definitely at the "try to break the model" phase. This is a super important phase with any emerging tech I want to actually use -- I have to make sure I fully understand what it CAN'T do, then figure out what I can do with it.
Of course, the things we can't do tend to get lower over time, too!
Yeah. Understanding the data which you're asking Code Interpreter to run analyses on is critical. You can't just assume that it understands the data in the way that an informed human would understand the data.
Great example, and important. Both a reasonable wrong request wording, and a far better request wording. Total # of CA drivers ( over many years) at 576 mln , is less good than total # of CA drivers in 2018. 27 mln is the (more?) right answer. More folks should guess first. I would have guessed 3/5s - 3/4s or up to 4/5s of CA pop of 36 mln.
Smart folk willing to estimate first, then check more carefully if the AI answer is outside the range, seem more likely to avoid wildly wrong answers.