I gave GPT4 a simple problem.
My Question: If my 3 towels dried outside in 3 hours, how long would it take if I had 10 towels out to dry?
GPT4 Answer: That’s a good question. Assuming the drying rate is constant and proportional to the number of towels, you can use a simple formula to find the answer:
time=number of towels /drying rate
Since you know that 3 towels dried in 3 hours, you can find the drying rate by dividing 3 by 3, which is 1 towel per hour. Then, you can plug in 10 for the number of towels and solve for the time:
time=10/1 =10 hours
So, it would take 10 hours for 10 towels to dry. However, this is only an estimate, and it may vary depending on other factors such as humidity, temperature, airflow, etc.
With all the massive training and millions of dollars invested in foundational models – how could the AI get something so simple wrong? Is this an issue of poor prompt engineering or the lack of common sense? We have often said that while generative AI can far exceed human ability on many tasks – there are many “common sense” things that a child could do that the AI gets egregiously wrong. The assumption often made is that AI is advancing on an exponential curve and will get better and better with more data. But can data teach common sense? Is most common sense even coded, or is it largely unsaid? What if we run out of data? Can synthetic data (generated by generative AI itself) be used for training new AI? Or will this just reinforce the pervasive biases and errors in the data?
It seems that the more likely solution is to place much more attention on training data rather than rely only on automation or scale effects for AI training. Humans “learn” common sense by interacting with the world through multiple modalities. Therefore, curating a “layer” of training data and reinforcement learning (with human feedback) might be the more promising path. It can create self-corrective mechanisms that reflect not only common sense but also the values needed to keep the AI on track.
Leave a Reply