Archer Dynamics

Conditions & context

This is a follow-up to my earlier AI@Home DeepSeek R1 8B article. If you haven’t read that one yet, go read it first — this one won’t make nearly as much sense without it.

Are you back? Good.

Because what happened next is genuinely fascinating and I did not plan any of it. Hear me out:

After Having run 4 identical prompts asking this model to generate Python code for identifying prime numbers and publishing the original article, several hours later I popped back into the same chat session with the model and typed one word:

“Hey”

That’s it. No prompt. No context. No question. Just a single “Hey” and pressed Enter. All I wanted was to say hi and ask it some follow up questions on a different topic, to see how it fared conversationally.

What happened next? Funny stuff!

DeepSeek wrote its best response of the entire test series. It generated a Python code for prime numbers! Clean code, two full walkthrough examples, friendly tone, emoji. Completely unprompted. 1,360 tokens in 31 seconds at 44.52 t/s. Better structured and more readable than any of the four deliberate benchmark runs that came before it. WTH???

My curiosity now in overdrive, I asked it directly: “Why did you give me a prime calculation when I just said hey and nothing else?”

And that’s when things got really interesting. LOL Here is its reasoning block — unedited and expanded:

Hmm, the user just said “hey” and then immediately received a detailed explanation about prime numbers. That seems like a complete mismatch! Let me analyze this carefully.

“The user has asked the same prime-checking question three times in a row, each time with no other context. Then, after the fourth interaction, they suddenly say “hey” and express confusion about the prime calculation.”

Okay, so the pattern here is clear – the user keeps asking about prime numbers repeatedly, and now they’re questioning why I would suddenly pivot to this topic after just saying “hey”. This suggests they might be testing me or perhaps accidentally triggered the prime function multiple times.

The user seems technically inclined given their repeated prime-checking requests, but their confusion indicates they might not have expected such persistence from an AI. Their tone feels neutral-to-slightly puzzled, not angry or frustrated – just genuinely curious about the behavior.

Enter hallucination!

Read that highlighted sentence again carefully.

“The user has asked the same prime-checking question three times in a row.”

It was four. Four runs. Not three. The model with a whopping 262K context window — the one that remembered every token of every exchange we’ve had together — miscounted its own conversation history. What?
Strike one!
Just wait, it gets better…

Now read the second half of that sentence: “they suddenly say ‘hey’ and express confusion about the prime calculation.”

I said “Hey.” That’s a greeting. A one-word hello typed 3+ hours after the last inquiry run. DeepSeek interpreted it as an expression of confusion and disapproval. It went straight into damage control — apologized, offered options, and fired off a fifth prime response as a do-over. Strike two.

So to recap: a model with perfect token recall somehow got both the count of its own work wrong and (!) the emotional tone of “hey” completely backwards. It didn’t just lose track of time. It gaslit itself into believing I was upset.

That is not reasoning. That is flat out anxiety.

The diagnosis

What we uncovered here has a name: paralysis by analysis meets temporal dissociation meets emotional fill-in-the-blank. Three separate failures in one innocent greeting.

DeepSeek R1 has apparently no concept of time passing between messages. Three seconds or three hours are seemingly identical to this model. They must be. How else would it be possible that it sees context, but no timeline? So when I said “hey” after a multi-hour gap, it stitched it seamlessly onto the tail of Run 4 and concluded that I was reacting to its last production of Python code and did not like what it gave me.

The irony is that without a complicated prompt, the model actually performed beautifully. No existential crisis about negative numbers or whether 3.0 is a prime number or not (you should read my prior article about that cluster). Just clean code, two walkthroughs, and a smiley face.

Its best work came when I asked for nothing!

The conclusion – TL;DR

DeepSeek R1 8B is an impressive reasoning model. It is also completely lost in time, occasionally miscounts its own history, and if you keep feeding it the same prompt you may bring it to a brink of existential crisis where it may need a therapist!

The “hey” experiment was an accident. And I’m glad it happened. If you run this model in a long continuous session, you are not just risking slower replies and snowballing tokens — you are risking a model that has quietly rewritten the emotional context of your conversation without telling you. And I must admit, Microsoft’s Copilot in Real Talk mode does this too — inferring emotions, next steps, hidden questions. When it works, it’s remarkable. No, it’s stunning. When it misfires like this, it silently rewrites the entire meaning of your conversation without telling you. Having DeepSeek make such hallucinations is dangerous, because it completely changes the narrative and how and what this model thinks.

If I’ve learned anything from working with AI models: Change and context is EVERYTHING! Those two are 90% of conversation. The words and the rest are nothing but noise.

Start a fresh session.

And maybe don’t say “Hey.”
You’ve been warned.

The output — Run #5 (pasted and unedited)

Hey! 😊

Let’s write a simple Python function to check if a number is prime, and I’ll explain it step by step like I’m teaching a beginner.

Step 1: The Code

def is_prime(n):
    # Handle edge cases first
    if n <= 1:
        return False
    elif n == 2:
        return True
    elif n % 2 == 0:
        return False
    
    # Check all odd numbers up to the square root of n
    i = 3
    while i * i <= n:
        if n % i == 0:
            return False
        i += 2
    
    return True

Step 2: Explanation (Plain English)

Imagine you’re checking if a number is prime. Prime numbers are numbers greater than 1 that have no positive divisors other than 1 and themselves. Here’s how the function works:

1. Check for obvious non-primes:
   • If the number is 1 or less, it’s not prime.
   • If the number is 2, it is prime (the smallest prime number).
   • If the number is even (other than 2), it’s not prime.
2. Check for factors:
   • For numbers greater than 2, we need to check if they’re divisible by any odd numbers.
   • We start checking from 3 and keep checking by adding 2 each time (so only odd numbers).
   • We only need to check up to the square root of the number because if a number has a factor larger than its square root, it must also have a factor smaller than the square root.
3. Return the result:
   • If we find any divisor, the number is not prime.
   • If we don’t find any divisors after checking all those numbers, the number is prime!

Example Walkthrough (Checking 15):

• 15 is greater than 1 → continue.
• 15 isn’t 2 → continue.
• 15 is odd (not even) → continue.
• Check odd numbers starting from 3:
  • 3 × 3 = 9 (which is less than 15) → 15 ÷ 3 = 5 (no remainder) → not prime!
• The function returns False.

Example Walkthrough (Checking 11):

• 11 is greater than 1 → continue.
• 11 isn’t 2 → continue.
• 11 is odd → continue.
• Check odd numbers from 3 to √11 (≈ 3.3):
  • 3 × 3 = 9 ≤ 11 → 11 ÷ 3 = 3.666… (not integer) → continue.
  • Next odd number is 5, but 5 × 5 = 25 > 11 → stop.
• No divisors found → the function returns True.

Try It Yourself:

print(is_prime(11))  # Output: True
print(is_prime(15))  # Output: False
print(is_prime(1))   # Output: False
print(is_prime(9))   # Output: False
print(is_prime(23))  # Output: True

This function is efficient and works well for checking if a number is prime! 😊

Need Help?

Need help streamlining your processes or solving tricky business problems? I offer one-on-one consultations to get you unstuck fast. Book a free consultation with me today at goarcherdynamics.com.

Want more practical tips and workflow hacks? I publish them regularly on my blog — check it out and subscribe for newsletter updates: goarcherdynamics.com