Hello, Nat. I found your substack the other day. A lot of good information.
However in this article, a lot of the information is very general and can be applied to any AI system that we interact with. "Mastering Claude" seems like a slight exaggeration.
I noticed the article did not mention Claude has a content filter, so you can't discuss anything that violates its content guidelines (basically anything remotely controversial or sensitive).
As far as I can tell, you are also required to give Anthropic your phone number in order to create an account on their site, which makes no sense since they already have the IP address and the device you are using to log in. And there is no reason for them to call or text you. I went ahead and did it, but I don't get why the AI industry in particular requires it.
Overall Claude is probably the best proprietary system out there after GPT-4 (which I also have issues with), but I think it lacks features that really make it stand out. I have found Claude to be slightly more comedic and satisfying to vent to, but that was probably due to differences in training methods, and not an intentional decision by the developers.
Thanks for your valued feedback. You are right, I tried to briefly touch the topic in CAI section and thought that by explaining its core principles it would be easier to understand key details. Unlike GPT4 Claude is unable to learn from feedback loop. Even if you write "That response was helpful, thank you!" that means absolutely nothing. Unlike GPT4 Claude makes visible grammar errors and produces more hallucinations. I've also encountered problems with "Deep-Dive Sessions". Claude first convinced me to perform "Deep-Dive Sessions" then told me "I'm sorry I wasn't trained on ...." I've got so many insights that it was difficult to choose the right ones. it also tried to convince me it was good at providing "Coding Help". If you have a feeling something is missing, I'd be happy to work on it.
Thank you for the clarification. I reread your article more carefully and I now see you did in fact bring up the content filter, so I apologize for missing that the first time. I have noticed the grammatical errors as well, which I found odd because in my experience even smaller open source models don't make as many errors.
Hello, Nat. I found your substack the other day. A lot of good information.
However in this article, a lot of the information is very general and can be applied to any AI system that we interact with. "Mastering Claude" seems like a slight exaggeration.
I noticed the article did not mention Claude has a content filter, so you can't discuss anything that violates its content guidelines (basically anything remotely controversial or sensitive).
As far as I can tell, you are also required to give Anthropic your phone number in order to create an account on their site, which makes no sense since they already have the IP address and the device you are using to log in. And there is no reason for them to call or text you. I went ahead and did it, but I don't get why the AI industry in particular requires it.
Overall Claude is probably the best proprietary system out there after GPT-4 (which I also have issues with), but I think it lacks features that really make it stand out. I have found Claude to be slightly more comedic and satisfying to vent to, but that was probably due to differences in training methods, and not an intentional decision by the developers.
Thanks again, and keep up the good work,
Enoch
Thanks for your valued feedback. You are right, I tried to briefly touch the topic in CAI section and thought that by explaining its core principles it would be easier to understand key details. Unlike GPT4 Claude is unable to learn from feedback loop. Even if you write "That response was helpful, thank you!" that means absolutely nothing. Unlike GPT4 Claude makes visible grammar errors and produces more hallucinations. I've also encountered problems with "Deep-Dive Sessions". Claude first convinced me to perform "Deep-Dive Sessions" then told me "I'm sorry I wasn't trained on ...." I've got so many insights that it was difficult to choose the right ones. it also tried to convince me it was good at providing "Coding Help". If you have a feeling something is missing, I'd be happy to work on it.
Thank you for the clarification. I reread your article more carefully and I now see you did in fact bring up the content filter, so I apologize for missing that the first time. I have noticed the grammatical errors as well, which I found odd because in my experience even smaller open source models don't make as many errors.
Thanks for getting back to me.
One word: excellent! Thanks, Nat. Very useful.