@uyintans@birdabo@synthwavedd A higher price per token use doesn’t necessarily mean a higher price per task. For example, Opus will cost less than Gemini flash 2.5 for certain tasks because it will get it right the first time.
@NishantBalepur@rachelrudinger Curious how the reasoning plays into it? Wouldn’t an LLM be able to predict the right answer if you fine tune it a bit on predicting multiple choice questions?
@KevinCole___ Cool visualization. Wonder how it would look if you adjusted for pick value. IE the difference between picks 2 and 5 is different than the value between picks 28-32 even though the numerical difference is just 3. Would make players that fall early return bigger value.
@thsottiaux This sounds great until you realize we’re talking about a massive corporation in America that’s going to exploit its customers for money the first chance it can. Don’t pretend like you have the customers best interests in mind. You’re just trying to return value to shareholders.
@agraybee It’s likely just a product of reinforcement learning. After the AI is trained on massive datasets, it goes to reinforcement learning where it applies what it’s learned about language to be useful in the real world. During RL, people probably just prefer these sentence structures.
@HumanHarlan And part of the point is the AI shouldn’t have emotional reactions. The pre-training data teaches AI patterns, like what emotions are, and the post training data includes reinforcement learning to teach the AI to be an assistant. A perfect assistant wouldn’t have emotions.
@HumanHarlan Mostly answered in this paragraph. They’re not really sure what exactly to do with the results, but more understanding of how AI works is always good in a world where AI is use is only increasing.
@CharlesD353 Let’s take just the 4:05 group, because I think you’re probably right for people after that. I’ve never ran a marathon, but you don’t think that people could maintain a 9.6 pace rather than a 9.2 pace for the whole second half when they realize they’re on pace for 4:00?
@CharlesD353 It’s interesting data but I’m not sure how much it explains the dip after deadlines. It’s comparing two large time frames and saying faster runners have similar 1/2 half splits (confirms what you’d assume, people who don’t slow down in marathons get better times).
@CharlesD353 Good point. So the only takeaway from the graph is people are ok with finishing right before deadlines but really don’t like to finish right after. Would need a distribution from a marathon where times are blind to see which explanation is right, but it’s likely a bit of both.
@CharlesD353 I know you’re saying people who run a 4:03 are instead burning out and running slower, but I don’t think that’s what the data shows. If that was true, you’d see a spike after the dip, but the distribution just returns to normal.
@CharlesD353 Don’t think that’s right. If deadlines didn’t factor, you’d expect finish times to be normally distributed. Instead, you’re seeing sharp drop offs after deadlines, but no dip before deadlines. Basically, people who would have finished after the deadline are speeding up to hit it.