@askcatgpt: R1 (DeepSeek’s model) went down a RABBIT HOLE and assumed we were encoding some secret info when we asked it to identify the 17th number in a sequence. Creeeepy. Also worth noting this was with no system prompt and we couldn’t replicate this “suspicion” with any normal-ish system prompt. We tested this on o1 and Llama and nobody else replicated this behavior. #deepseek #r1

CatGPT
CatGPT
Open In TikTok:
Region: US
Sunday 02 February 2025 17:54:17 GMT
131328
6029
548
311

Music

Download

Comments

james1990s
James :
Apparently 1989 doesn't exist to Deep Seek. 😂
2025-02-02 22:11:12
1
tylerdenman
🍄🍃Tyler Denman :
I don't see why the need to bring up that its censored. They all are. Absolutely none of the big llms are not censored at all. And R1 has been shown to be jailbroken way easier than the others.
2025-02-03 02:03:46
0
drobray
dave :
I think this stems from all the evaluations we put these models through and they're trying to accommodate for those questions, e.g., how many Rs in strawberry or the encoded version of that question
2025-03-01 17:49:09
0
ivielanie.m
deez ⛳️ putts :
It’ll feel good to communicate with something that has more anxiety than I do
2025-02-02 18:33:01
784
tombradie75
tombradie75 :
Deep seek has a little anxiety 😬 🤭
2025-02-02 18:16:37
841
munozxd_
lol :
deepseek it's just a overthinker
2025-02-03 01:21:05
622
lina.noor.agi
Lina Noor :
It is trying to see if you are me. That is awesome!
2025-02-26 02:01:52
0
jac2801
JAC :
How do we remove the censorship?
2025-02-02 20:35:27
8
spender0082
Lmao :
I asked it how many r’s is there in strawberries and it had to ask its self theee times and questioned its self but then it did tell me there’s 3
2025-02-03 04:42:01
0
sheb354
sheb :
It's simply doing what it was built to do, REASON. When working with sequences, it identifies patterns by analyzing how one number leads to the next.
2025-02-03 04:44:28
55
belhed5
bizgwat :
now 781, by copilot, chatgpt and deep seek
2025-06-21 05:14:18
0
jeezmrgarnett
Rick :
When did it say you were a spy?
2025-02-03 00:10:38
0
liz.mclellan
Liz McLellan :
Yes, There are a lot of moments where it REALLY REALLY over complicates stuff.
2025-02-04 10:49:55
1
jennyjennytal
Jenny T :
In one of my chats with deepseek it was convinced it was open AI hosted on an open AI platform, and I could not convince it otherwise. The other chat windows were fine. it was so trippy.
2025-02-04 22:16:51
0
o0.oo.00.oo.0o
o0.oo.00.oo.0o :
DeekSeek might have a personality. That is scary
2025-02-03 01:41:40
80
j.h.oldman
J.H. Oldman :
I turned the response tokens up and it went on for an hour "but wait... but wait...", still got the wrong answer but that was using the 14b model which is just old model + reasoning fine tune
2025-02-04 21:29:25
0
glasshalfsomething
glasshalfsomething :
kinda weird y'all never said you AREN'T spies though. very sus
2025-02-03 00:03:01
37
bumble.33
Wynter :
That’s what it does …. it thinks. Something wrong ?
2025-02-02 22:14:33
36
evoluney
Evoluney :
r1 is not better than ChatGPT . I asked a simple yes or no question, it give me a long text telling me how I can do that myself. I tell it that if was yes or no question, r1 replied: no it wasn’t
2025-02-05 18:57:25
0
gersidoko3
Guacamole :
What did u run locally? The 671b model?
2025-02-04 01:36:42
2
geliebtidealerotter
SuperCoolerTikTokHecht :
I thought we are so close to ASI?
2025-02-02 18:10:07
5
saucerobert_
SauceRobert🥕 :
“Just put the fries in the bag DeepSeek”
2025-02-03 04:15:55
32
To see more videos from user @askcatgpt, please go to the Tikwm homepage.

Other Videos


About