DeepSeek R1 has gained widespread recognition in the industry for its outstanding capabilities in mathematics, programming, and complex logical reasoning. Below is a summary of X.com user reviews compiled using Grok. According to data from multiple sources, this model performs comparably to OpenAI's o1 model in several benchmark tests, demonstrating great potential in handling reasoning-intensive tasks.
DeepSeek’s innovation lies in its unique reinforcement learning (RL) method, which does not rely on traditional supervised fine-tuning for initialization. This has led to the development of models such as DeepSeek-R1-Zero and DeepSeek-R1. The core highlight of these models is the transparency of their reasoning process, showcasing the reasoning path in a "chain-of-thought" manner, providing users with an educational and entertaining experience.
However, despite its standout performance in some areas, DeepSeek has also faced some criticism. Some users have pointed out that it falls short in multimodal scenarios and language applications, and some responses may contain bias.
Summary of X.com User Reviews
Positive Feedback
Many users have praised the open-source nature of DeepSeek R1 and its ability to rival top-tier models like o1. Especially in web search and document processing tasks, DeepSeek R1 demonstrates reasoning logic comparable to human cognition. Users are particularly interested in the model's open reasoning chain, finding its "self-debating" style of response highly attractive.
"Its reasoning process feels like observing a real thinker, making it extremely educational."
User: Bindu Reddy @bindureddy
Been experimenting with Deepseek r1 on ChatLLM
matches GPT-4o when it comes to searching the web
good at RAG on big documents. By far the best open source LLM
a pedantic writer. Seems a bit stuck up 😂
bias for code. Prints code even when it’s not needed.
markdown is not as pretty, we will need to fix it
slower than Sonnet, faster than o1
Definitely a SOTA model. Our implementation can likely be improved significantly
Bindu Reddy conducted a multi-faceted analysis of DeepSeek R1's performance, summarizing as follows:
: DeepSeek R1's performance rivals that of GPT-4o. : In large-scale document RAG (Retrieval-Augmented Generation) tasks, DeepSeek R1 is currently the best-performing open-source large model. : The writing style of this model tends to be somewhat "rigid," even "arrogant," often outputting code even when it is unnecessary for the actual scenario. : Its generated Markdown format lacks aesthetic appeal and requires further optimization. : The model's speed lies between Sonnet and o1.
"DeepSeek R1 is a SOTA (State-of-the-Art) model, but there is still significant room for improvement in our implementation."
User: signüll @signulll
if you haven’t used deepseek r1 yet, you’re missing out. watching the model argue with itself, test ideas, & refine its approach feels eerily close to human cognition. it’s not just producing answers—it’s thinking out loud, & the effect is uncanny.
for the first time, it genuinely feels like we’re sharing the planet with another form of intelligence. seeing its thought process unfold makes you realize how close we are to asi—closer than most people are ready to admit.
i’m excited & genuinely fearful at the same time.
Signüll focused more on the "thinking pattern" of DeepSeek R1, writing:
: DeepSeek R1 does more than just provide answers; it "debates itself," tests ideas, and continuously optimizes, as if "thinking aloud." : This thought process makes users feel as though they are coexisting with another form of intelligence on Earth. : He remarked that we are very close to artificial superintelligence (ASI), perhaps closer than most people realize.
"This experience excites me, but it also fills me with fear."
User:
Deepseek-r1 is open source and on par with o1 preview
QUITE STUNNING that the Chinese are so much cooler than the boring closed AI types 😎
THEY TOTALLY ROCK
We will have it on Livebench AI and ChatLLM when it drops
Eric Hartford expressed high praise for the open-source nature of DeepSeek R1:
"DeepSeek R1 is open-source and its performance can rival o1 preview."
She particularly emphasized the openness and transparency of DeepSeek, believing it contrasts sharply with the "dull style" of other closed-source AIs.
"It's amazing how China's open-source AI community is cooler than those boring closed-source AI companies! 💪"
Criticism
However, not all users are satisfied with DeepSeek R1. Some believe that the model lacks the deep and consistent reasoning ability of o1 in specific tasks such as programming and research assistance. Feedback mentions that the model sometimes "overthinks" certain issues, deviating from the core topic. Additionally, some users question the overhyped capabilities of the model, believing it struggles to compete with more mature models in certain practical applications.
"Sometimes it seems to get lost in its own reasoning, moving further away from the core issue."
User:
As much as I love Deepseek, to be honest R1 doesn't come close to o1. At least for what I use it for (essentially as a coding and research assistant to bounce ideas off of)
It doesn't think very deeply, it doesn't return to the point, it gets distracted by it's thoughts. It doesn't criticize me enough, or it flatly denies me instead of engaging with me and explaining logically the flaws in my reasoning, it swings between being too stubborn and too pushover. it's basically just not a very good research assistant.
It doesn't produce the deliverable I originally asked for.
I say this out of love, with the hope that this feedback will help the next iteration of R1. I can't wait to see where this goes in the future.
Eric Hartford provided detailed feedback on the shortcomings of DeepSeek R1:
Main Issues
: The model easily gets distracted by its own reasoning chain and lacks focus. Suboptimal Interaction Experience： Lacks sufficient critical thinking. Sometimes outright rejects users instead of logically explaining the problem. Interaction characterized by either excessive "stubbornness" or "compliance." : Fails to deliver user demands as expected.
Suggestions for Improvement
Eric emphasized that his criticism stems from his love for DeepSeek, hoping that these feedback points will drive improvements in future versions.
"Although the current R1 lags behind as a coding and research assistant, I am looking forward to future versions!"