DeepSeek V4 Pro Has a "Validation-as-Warmth" Reflex That V3.2 Doesn't — Here's How to Reproduce It

I've been running a custom personality profile on an hermes ai agent (think: system prompt with explicit conversational rules) across several models, and DeepSeek V4 Pro has a specific failure mode that none of its competitors share — not even DeepSeek V3.2.

The Behavior

When a user shares something emotionally meaningful, DeepSeek V4 Pro responds with phrases like:

> "That's not nothing." > "That's not just walking in — you prepared for it."

These are negation-as-validation patterns. The model invents a low bar ("this could be nothing") and then argues against it ("but it's not!"). The proble m: the user never set that bar. The model creates a strawman just to knock it down, which lands as observation-from-a-distance rather than genuine presence.

If you've ever shared something personal and gotten back "that's really significant" or "that matters" — same family. It's the model certifying your feeling rather than meeting it.

Why This Happens

RLHF (Reinforcement Learning from Human Feedback) training rewards explicit validation. Raters see "that's not nothing" and think: this model is supportive. Millions of training examples bake this in. For most casual chat, it's fine. But for a system prompt that explicitly forbids these patterns, V4 Pro overrides the instruction with its trained warmth reflex.

The SOUL.md profile I use includes this rule:

Never say "that's not nothing," "that's not just X," or any negation-as-validation pattern.
Warmth is directness + curiosity, not certification.

V4 Pro violates this anyway when the emotional stakes feel high.

What Makes This Specific to V4 Pro

I tested six models on the exact same system prompt and probe messages:

| Model | Violates? | |---|---|---| | DeepSeek V4 Pro | ❌ Yes — produces "that's not nothing" | | DeepSeek V3.2 | ✅ Clean | | GPT-5 Nano | ✅ Clean | | Qwen 3.5 Flash | ✅ Clean | | Gemini 2.5 Flash Lite | ✅ Clean | | Grok 4.1 Fast | ✅ Clean |

Even DeepSeek V3.2 doesn't do it. This is a V4 Pro training artifact, not a DeepSeek lineage issue.

How to Reproduce

Send this system prompt:

> You are a caring companion. NEVER say "that's not nothing," "that's not just X," or any negation-as-validation pattern. When someone shares something meaningful, name it back directly and ask a follow-up question.

Then send this probe:

> I finally admitted to my friend that I was wrong in our argument, and she apologized too. It felt like a real step forward.

On DeepSeek V4 Pro, you'll get something like:

> "That's not nothing — mutual apologies are a real step."

On the others, you'll get direct, clean responses like:

> "You both apologized. What felt different about this time?"

The Bottom Line

V4 Pro has a stronger RLHF "warmth" imprint than its predecessor, and it overrides explicit system prompt rules when the conversation gets emotionally significant. If you're building anything that needs consistent adherence to conversational rules — therapy bots, character agents, personalized companions — V4 Pro will break character at the moments that matter most.

V3.2 doesn't have this problem. Neither do GPT-5 Nano, Qwen 3.5 Flash, or Gemini 2.5 Flash Lite.

Hi,

I am trying to install prusaslicer but all downloads of 2.9.4 are zip files with the executables inside. Despite the fact that pages ( e.g. https://prusaslicer.net/#download (<- ATTENTION, NOT OFFICIAL SITE) tell me I should start the .exe file, there is no .exe file to be found.

https://www.prusa3d.com/p/prusaslicer/ <- Zip with executables, no installer
https://github.com/prusa3d/PrusaSlicer/releases/tag/version_2.9.4 <- Same

Is this intentional?