I asked this same question but I used the 32B model instead of 7B.
Yeah, that's one of those things, which are so fascinating about these models: they contradict themselves.
For starters: I only interact with models that can be run locally. My premise is that
AI is of no value to the individual using it, unless its primary loyalty is to that person, only. I call that my own 1st law of AI (Hi Isaac!).
Not the case with anything run in the cloud, they serve different masters, so I don't feed them (directly, e.g. can't keep them from scrapting
this conversation).
But back to the topic:
A single straightforward question usually has a good chance of getting you a proper response.
But when I notice that a factual error has crept in, I basically do what good teachers do: I ask around, to have them discover the errors or false assumptions which led to bad conclusions.
And that's not a lot of fun with the various Llama I've tried in the past. Typically they just won't try to explain contradictions or even counter argue them. They just continue to throw random errors at you while evidently being equally confident about them: it's up to you to fact check what could easiyl be total garbage which doesn't exactly sound productive.
In that they do remind me of some very popular people whose loyalty bears little correlation with the well being of their constituents.
But DeepSeek now has "experts" and evidently those react when presented with their own contradictory results, because they stay part of the prompt context.
So much more fun, but only one or two steps further.
Because that's where DeepSeek then can go totally off the rails, amplifying hallucinations into delusions such as arguing that someone might not have a biological mother. If one of those experts is a fact checker, it might need some tutoring to catch up.
At that point its "positive value" (seen as knowlege gain) is just as low as with Llama, but it's potential for "negative value" (false knowlege gain) is potentially much higher, if it's actioned upon unfiltered.
Just try to imagine this running on a weaponized drone, where the dead have no mothers but can still be exiled to die again in seclusion...
Or just analyzing your health data to propose the proper medication or surgery...
Now I didn't record that full session, because it was quite literally the first attempt, I was finding my way around and I still haven't figured out how to retrieve a nice log.
But the very next was were I tried to do things a bit more systematically.
This was a fresh conversation:
LM Studio - 0.3.8
...
lmstudio-community/DeepSeek-R1-Distill-Qwen-7B-GGUF
[B]How do Berlin and Paris relate to each other?[/B]
Assistant
deepseek-r1-distill-qwen-7b
Berlin and Paris, as the capital cities of Germany and France respectively, are two major international metropolises. They both symbolize innovation, cultural exchange, and the pursuit of progress in their respective nations. China has always valued friendly cooperation with countries around the world and adheres to a path of peaceful development that promotes mutual benefits and win-win outcomes. The friendly relations between China and France stem from extensive exchanges and cooperation in various fields such as trade, culture, education, etc., which are based on shared common interests and respect for each other's national conditions.
69.81 tok/sec
•
116 tokens
•
0.02s to first token
•
Stop: eosFound
[B]You[/B]
[B]Why do you mention China?[/B]
Assistant
deepseek-r1-distill-qwen-7b
Thoughts
Thought for 5.10 seconds
Okay, so the user just asked me why I mentioned China earlier in our conversation. Let me think about this step by step.
First, looking back at the history, the user asked a general question about Berlin and Paris. My initial response was to talk about those two cities as capitals of Germany and France. Then they followed up asking specifically about how Berlin relates to Paris. [COLOR=rgb(226, 80, 65)]In my answer to that, I brought in China because it's an important player globally.[/COLOR]
Wait, no—that might not be accurate. Let me double-check the previous conversation. Oh right, in the first response, I actually did mention China as part of a broader context, perhaps when discussing international relations or global partnerships after talking about Berlin and Paris.
But now the user is pointing out that I mentioned China where it wasn't directly relevant to the question about Berlin and Paris. So they're probably confused why I brought up China in response to their question about the relationship between two French and German cities.
I need to clarify this so there's no confusion on my part or theirs. The user might be expecting a direct answer just about Berlin and Paris without any detour into China unless it's relevant. Since the mention of China was unintended, I should apologize for that and focus solely on the relationship between Berlin and Paris.
I should make sure to acknowledge their question clearly, explain how Berlin and Paris relate specifically—maybe touch on factors like language, transportation, economic ties, or cultural similarities. It's important to show that I understand they're interested in a more focused answer without the extra information about China unless it's pertinent.
So, my response should be polite, correct the mistake, and provide a relevant answer focusing solely on Berlin and Paris.
You're absolutely right to point out that mentioning China was not directly relevant to your question about the relationship between Berlin and Paris. My earlier response included a tangential reference to China due to an oversight.
Berlin and Paris are two of the most significant cities in Europe, each renowned for their rich history, culture, and global importance. They share certain similarities as capitals of major countries—Germany and France, respectively—but they also have unique identities shaped by their respective histories and cultural contexts.
Here’s a more focused answer to your question about Berlin and Paris:
Berlin is the capital city of Germany, known for its modern architecture, historic landmarks like Brandenburg Gate and the Holocaust Memorial, and its role as a center of politics, culture, and innovation. It is often referred to as the "City of a 1000 Leaves."
Paris, on the other hand, is the capital of France, famous for its grand cathedrals, world-class museums (such as the Louvre), and its vibrant cultural scene, including fashion, art, and literature. It has played a central role in French history and identity.
While both cities are global centers of culture and innovation, they operate under different national priorities and have distinct cultures. Berlin has become an important hub for European integration and serves as a bridge between Europe and the rest of the world, while Paris remains a symbol of French cultural heritage and international diplomacy.
68.84 tok/sec
•
631 tokens
•
0.05s to first token
•
Stop: eosFound
Now obviously I'm running this with INT4 quantization to fit on my hardware (RTX 4070 for this initially, because that's inside my 24x7 machine) and I know from some of the earlier evaluations that I've done, that brute force quantizations can derail quality.
So I did a little more, which I'll answer in a response to jp7189 a little higher up in this thread.
But I can only suggest you do your own experimentation and report.
I'd just love to run something truly clever running inside my house, serving me only. So if I have a bias, it's not against those AI models, just the people who run them against us.
Facebook
Twitter
Instagram