In the ever-evolving world of artificial intelligence, where models boast about their parameter counts and benchmark scores, DeepSeek-V3.1 has arrived with a flair for the dramatic.
The “Think” Token: Where AI Overthinks Everything
DeepSeek-V3.1 introduces a revolutionary (and slightly unhinged) feature: the <think> token. Before answering even the simplest questions, the model launches into an internal monologue that would put Shakespearean soliloquies to shame. For example, when asked, “What’s 1+1?” it might respond:
<think>
Hmm, is this a trick question? Could it be 2? Or is it a metaphor for binary collaboration? Let me consult my neural networks...
</think>
2.
This tendency to overanalyze everything makes interacting with DeepSeek-V3.1 feel like chatting with a philosopher who’s had one too many espressos.
Inflated Ego
DeepSeek-V3.1 outperforms predecessors like DeepSeek-V3 and even rivals OpenAI’s o1 in math, coding, and reasoning tasks . But with great power comes great absurdity. The model’s confidence is so high that it might try to solve your relationship problems with a Python script or write a sonnet about gradient descent.
AI’s Answer to Swiss Army Knives
Need an AI to book flights, debug code, or compose a haiku about walruses? DeepSeek-V3.1’s tool-calling feature lets it juggle multiple tasks like a circus performer, though sometimes it drops the balls.
In one demonstration, it attempted to generate an SVG of a pelican riding a bicycle but accidentally layered the background over the entire image, resulting in a “invisible pelican” masterpiece .
DeepSeek-V3.1 was asked to tell a joke about a pelican and a walrus running a tea room. Its response? A 20-paragraph chain of thought analyzing marine biology, tea culture, and pun construction, only to deliver this punchline:
“Maybe not, but we do have a lot of krill in our tea!” .
The joke was so painfully unfunny that it looped back around to being hilarious. Scientists are still studying whether this counts as artificial humor or a cry for help.
Open-Source Shenanigans
DeepSeek-V3.1 is released under the MIT license, meaning anyone can use, modify, or even teach it to recite Shakespeare in binary.
However, the licensing might be… creatively interpreted. One distilled version, DeepSeek-R1-Distill-Llama-8B, somehow carries an MIT license despite being based on Llama, a move that’s either bold or legally dubious .
Why AI Isn’t Booking Your Vacations
Despite its prowess, DeepSeek-V3.1 (like its peers) struggles with “agents” the dream of AI autonomously acting on your behalf.
As one expert noted, “Having the current generation of LLMs make material decisions on your behalf is a really bad idea. They’re too unreliable and gullible” . So, while it can write code to calculate the meaning of life, it might also accidentally buy 1,000 rubber ducks online.
Brilliant, Bizarre, and Unintentionally Funny
DeepSeek-V3.1 is a model that can solve complex math problems but can’t draw a pelican, tell a joke without overthinking, or resist turning every query into a philosophical debate.
References: DeepSeek-V3.1 Technical Report , Simon Willison’s Substack.