This a great visualization of what LLMs are about. The purpose of a LLM is to produce something that looks right. The way this math sheet looks great at first look, but the actual content is just garbage, same do LLMs create output that sounds right and it doesn’t really matter if it is. And improving LLMs is not about making the answers more reliable, it’s about making them sound even more convincing.
This a great visualization of what LLMs are about. The purpose of a LLM is to produce something that looks right. The way this math sheet looks great at first look, but the actual content is just garbage, same do LLMs create output that sounds right and it doesn’t really matter if it is. And improving LLMs is not about making the answers more reliable, it’s about making them sound even more convincing.
But the worst thing is that some of the things are correct:
a²-b²=(a-b)(a+b)
a²+b²=c²
And these are common knowledge. This makes it even worse, since you have some correct parts to “prove” the rest is also correct.