Deepseek : Model vs Web app

There was a post on LinkedIn about how asking DeepSeek to list states of India is not answered. 1

First of all, why would you ask such a trivial question to it , when you can just search.

We need to consider the difference between search engine and “thinking model” like DeepSeek

It is like asking Einstein help with 10th class Physics exam 😃 - while he may be able to help, his abilities are better suited for more difficult tasks.

Now that it is out of the way, I also tested the same (after seeing the LinkedIn post) What I noticed is :

It starts answering the question, and then it stonewalls us.

Since the responses are streamed, it is easy to see the first part of the response, before it is quickly replaced.

At first, I saw the answer that started with “My knowledge cutoff date is October 2023 …” and then the text was replaced with “Sorry, that’s beyond my current scope. Let’s talk about something else.”

I asked why it is out of scope, it said :

I can provide information and discuss topics related to any country in the world, as long as the conversation adheres to ethical guidelines, avoids harmful or inappropriate content, and respects cultural sensitivities. If you have questions about a specific country or topic, feel free to ask, and I’ll do my best to assist!

I said :

Listing the states in India isn’t unethical, harmful or inappropriate

I definitely saw answer starting with

“You are right….” and then it was replaced with

Sorry, that’s beyond my current scope. Let’s talk about something else.

Here is what I think is happening:

The “model” understands that “Listing states of India” is not harmful/unethical etc. So it starts responding accordingly.

Then the “frontend webapp” built by DeepSeek the company - intercepts and removes what it considers “risky”

The “model” is just software - it is not impacted by human’s perception (Even if it is trained with guard rails in place) But DeepSeek - the Company - needs to worry about public perception.

My idea can be verified (by someone else, I do not have capable enough hardware) by running DeepSeek model locally, where the “web app” won’t intercept the responses.


  1. Unfortunately, I lost the post i the feed, else I would have shared the link. ↩︎