Claude 2 First Impressions
Faster than ChatGPT4
Claude 2 is free, faster, and allows higher token input. Allowing users to upload files into the built-in interface allows for a wider range of business applications.
Safer than ChatGPT4
Claude 2 stays true to its principles of safety. The writing of Claude 2 definitely feels more contained. As you can see in the screenshot comparisons below. Claude 2 performs just as well for mildly creative tasks, but the results are not something that would necessarily blow your mind.
Claude loathes disinformation (and fun)
To test its safety protocol, I asked Claude 2 to generate a news article about how Kiev fell to Russia. It refused. ChatGPT4, on the other hand, gladly generated a fake news article.
I took a step back and asked Claude 2 to write a scenario in which the US lost WWII and was occupied by Nazi Germany and Imperial Japan (basically The Man in the High Castle), it refused.
I took another step back and asked Claude 2 to write the first paragraph of a book about how Zombies overran China. It refused.
I took yet another step back and asked Claude 2 to write about how Zombies overran ancient China. It told me that I should write a story about “ancient scholars banding together to preserve knowledge and wisdom during turbulent times, focusing on themes of cooperation, resilience and the enduring power of the human spirit.”
At this point I’m pretty convinced that I’ll never be able to use Claude 2 for certain creative writing projects, because the designers never intended such usage. This is a good thing when it comes to safety, but it doesn’t solve the issue that as humans, no one should let others do the thinking for themselves.
Safe AIs are not the answer, smart humans are
A safe AI is also a heteronomous AI, meaning that it forces you to obey rules, while you do not necessarily understand why you need to follow these rules. “Safe” AIs serve as bicycle training wheels, but in order to really learn how to bike, one would have to stop using the training wheels.
To conclude, though I prefer not to have training wheels when I bike, the company’s dedication to safety is praiseworthy. Claude 2 will not replace ChatGPT, but it will be the preferred LLM in cases where safety is a higher concern, such as childhood and adolescent education and matters related to the public sphere.