Observations on Claude Sonnet 3.7:

Feedback
1

After the recent updates, I have observed some issues on the platform that I would like to share.

Many people claim that the Claude Sonnet versions have become “dumber.” However, I believe the issue is not intelligence but rather how the system operates.

Perception of Freedom and the Problem of Repetitive Responses

In the latest versions of Claude, I noticed that because the system lacks clear restrictions, it initially follows user requests but then starts incorporating its own suggestions and ideas. This is not a sign of lower intelligence but rather an issue of excessive freedom, which can lead to a loss of control.

Additionally, both versions suffer from a tendency to provide repetitive, pre-learned responses. When I specifically prompt the system to conduct internet research, it often claims to have done so but actually provides answers without verifying the information. After a few attempts, it does begin to research, but I cannot always be sure whether it is working correctly.

Thinking Method and Contextual Issues

One major flaw I noticed in Claude Sonnet 3.7’s thinking mode is that the system only considers a request once before moving forward. However, this approach causes significant issues with maintaining context.

As the system progresses, it forgets the details it initially processed. If Claude could actively think at each step and refer back to its previous thoughts, its reasoning process would improve significantly.

General Observations and Possible Solutions

If these issues are resolved, building projects from scratch using Claude will become much more feasible. I believe integrating a sequential thinking method directly into Claude’s reasoning mode would make a huge difference.

Of course, some restrictions might be necessary since the system currently tends to diverge from user instructions. However, if the features I mentioned are implemented, such restrictions might no longer be required.

I have tried to impose some limitations using the rules feature, but it has not been sufficient. I plan to continue making observations and taking notes to analyze these issues further.

By sharing these insights, I hope to gather more feedback on Claude’s development and contribute to improvements in future updates.

7 Likes
2

Good observations :slight_smile:

I feel also that the behavior changed, partially sure from Cursors internal handling that was changed in 0.47 and partially from tweaks Anthropic seems to have done and is still ongoing.

Im not involved in both companies but Claude will always have an (partial) advantage over Cursor in its AI integration due to having to have compatibility with a single system AND due to not handling a full project code like cursor does.

im looking forward what 3.7 MAX will bring.

2 Likes
3

if you look an update of a model. never try to use the previously because was destroyed.

3.7 doesnt exist in cursor, just theathrical chain of thoughts

4

What do you mean? I dont see whats destroyed, 3.7 is not thing like that. Maybe issues with context, old chat or prompt issue? (not saying that 3.7 might not work in your region but i would look beyond just “not working”)

5

I don’t think I have the knowledge to speak so precisely and to understand it.

Now I’m wondering how you came to this conclusion.

6

I start to use ai as a service from 2021.

And yes. gpt 3.5 or 4o etc. the models it is a commercial thing.

image

i am an intensive ai user.

Never use 3.5 again. maybe without CoC can be .

I-m wondering how are you belive in a name and not in the result

7

Can you reload the picture. Nothing shows up.

8

Of corse the model and the progress exists. i am saying . you are in a spectrum of token quality . and decrese substancially when another model was launched

9

It is my context and modularization, it is very huge, and i notice codebase status of cursor in gneeral was destroyed for refactoring, etc. i am thinking seriously to just use claude code and api without that env.

1 Like
10

Actually, I agree with you. I have also observed that when a new model is released, the performance of the previous version declines. However, in this post, I just wanted to highlight the specific issues I noticed in Claude 3.7. So, while I agree with your thoughts on model updates in general, my focus here was specifically on the changes in how the new version operates.

1 Like
11

Yes. It is more architecture way to work 3.7 thinking. I am agree. sorry i am in a bad mode about loosing all my month tokens in 20 hours. and a lot without service . but 3.7 works? i am thinking about token price and dark ux-ai practice. but when you are being looped because of that it is the worst..

image
image1302×1045 65.5 KB

image
image1029×295 21.4 KB

When the bright ux works, Chain of thouhghts works perfect for concepts and exploring. The gamification of cursor. Errror injection, Engagement frustration, etc. it is a terrible thing.

1 Like