This is a pretty simple query and it fails to adhere to the criteria:
Q: 5 letter words starting with e
‘Egg’ is not what I would expect to see in a breakthrough. Maybe in a few more years.
7 Likes
“This recipe involved several innovations, most notably the application of pure reinforcement learning to teach a base language model how to reason without any human supervision.”
from
they did open source the weights and explained the process but did not open source everything (especially the training data)
now, using 8b models will never give you good results. they’re made to be fine-tuned by others more than anything else.
try your query on the 671b model:
https://chat.deepseek.com/
don’t ask about Taiwan situation or Tiananmen though…
Microsoft open source phi4 14b was the best for some RAG queries. better than llama.
1 Like
Yes, that is expected (although I consider the 5 letter question pretty basic even for an 8B model). I should probably state more clearly that I was referring to the common talking point in news and media that its so lightweight you can run it on a laptop. Most people aren’t going to have the compute power to run 671B.
note that the small ones are distillations from other models (qwen/llama) so not really deepseek model.
fine tuned small models can be useful but generalist good models fitting in 8b or even 70b is not going to happen anytime soon. We indeed need more innovations to get small generalist models (I wished something like gpt4o could run locally!) but this Chinese company did bring some good open source stuffs. What OPENai should be doing instead of hyping AGI like it’s going to happen tomorrow.
this week is insane… DeepSeek R1 and Janus yesterday… QWen 2.5 today pulverizing all of them today… and flying low on the radar: the chinese model run inference on Huawei chips instead of NVIDIA…
and it’s just Tueday and holiday in China…
like some said: ChatGPT lost it´s job to AI eventually…
(they’re all just plagiarism machines, regardless of how efficiently they work on a chip)
8 Likes
Running an 8b version of a model and concluding that means the model has poor accuracy is kind of like recording a synth at 11 kHz and concluding the synth has poor sound quality.
1 Like
I like LoFi, looking forward to install DeepSeek r1 1.5B on my computer just for fun.
2 Likes
I wonder how the quote “Premature optimization is the root of all evil” applies to current AI. U.S. tech companies rushed to push their AI models onto the market, at gargantuan scale, motivated by their wish to corner the market. In the process, they hoarded knowledge and created proprietary black box systems. It is ironic that an industry insisting that it cannot operate without every bit of human knowledge being made freely accessible to it…has the exact opposite attitude about the sharing the technical specifications of their own products. Meanwhile, others (the Chinese?) were sharing technical information more freely with one another. No surprise they came up with an improved system.
1 Like
And that is why we love you, just as you are, you meat puppet.
2 Likes
1 Like
openai should really change their name.
With all the GPU data centers they have, they certainly can benefit from the open sourced papers given by DeepSeek (and other open source projects) so they might have even more power in their hands with the same stack.
2 Likes
Great comparison 
I’m absolutely not suggesting using the smallest 1.5B model as comparison for the full model and dismissing the whole thing as trash. Distilled is still part of the discussion though - when there is talk about telling people, oh, it’s no problem to run locally, it should be noted that the vast majority of people are not going to be using the full R1. From my testing and what I have garnered from others, all the distilled versions, even 70B, are pretty much mouthy mixed bag at the moment.
I can’t speak to the web version because I am not interested in training or sending data to China - I am coming at this from a homelab use case perspective. Reiterating, open-source nature is a huge win in my book. Breakthrough (or as it being referred to, Sputnik moment) is not on the end user level.
The complete lack of self-awareness is amazing.
Twats.
13 Likes
Agree, AI can GTFO
1 Like
I must be too old. I never imagined people could use an AI to discuss on forums.
AI is nothing but IP theft all the way down
2 Likes
If you don’t like AI you should probably stop using all search engines.
That seems like a pretty facile argument, given that there are no longer any alternatives to using search engines in 2025.
If someone said that they have concerns about, say, the safety of our food system, “you should probably stop eating food” is hardly a productive response.
6 Likes