DeepSeek R1 is an open sourced model. DeepSeek is a Chinese AI research company backed by High-Flyer Capital Management, a quant hedge fund focused on AI applications for trading decisions. They have released models under open-source licenses like MIT.
How did they match or even surpassing OpenAI’s O1:
Reinforcement Learning Focus: DeepSeek-R1 and its variant, DeepSeek-R1-Zero, were developed using a pure reinforcement learning (RL) approach, which is a departure from traditional methods that often rely on supervised learning. This method has allowed the model to develop reasoning capabilities autonomously, without initial reliance on human-annotated datasets. This approach has proven effective in allowing the model to achieve high performance in reasoning tasks.
Performance on Benchmarks: DeepSeek-R1-Lite-Preview has demonstrated comparable or superior performance to OpenAI’s O1 on several benchmarks, such as AIME and MATH, which are focused on mathematical reasoning and problem-solving. This performance is attributed to DeepSeek’s use of chain-of-thought reasoning, where the model explicitly shows its reasoning process, which not only aids in transparency but also in refining the model’s approach to complex problems.
DeepSeek’s first-generation reasoning models are achieving performance comparable to OpenAI’s o1 across math, code, and reasoning tasks!
Give it a try! 👇
7B distilled:
ollama run deepseek-r1:7bMore distilled sizes are available. 🧵 pic.twitter.com/FdF1U3qvev
— ollama (@ollama) January 20, 2025
DeepSeek has largely replicated o1-mini and has open sourced it. pic.twitter.com/2TbQ5p5l2c
— Aravind Srinivas (@AravSrinivas) January 20, 2025
Reinforcement Learning is well-suited for tasks that involve sequential decision-making, where the AI must learn to take a series of actions to achieve a goal. The goal of DeepSeek-R1 is to generate coherent, contextually appropriate responses in conversational AI or other interactive applications. Reinforcement Learning lets DeepSeek-R1 learn to optimize long-term outcomes rather than just immediate rewards, which is critical for maintaining context and coherence over extended interactions.
The DeepSeek R1 model has a 671 billion parameters architecture and has been trained on the DeepSeek V3 Base model. It is focused on Chain of Thought (CoT) reasoning to compete on advanced comprehension and reasoning. Just 37 billion parameters are activated during most operations which is similar to DeepSeek V3.
We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive – truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely.
DeepSeek-R1 not only open-sources a barrage of models but… pic.twitter.com/M7eZnEmCOY
— Jim Fan (@DrJimFan) January 20, 2025
DeepSeek R1 ecosystem has six distilled models fine-tuned on synthetic data derived from DeepSeek R1 itself. These smaller models vary in size and target specific use cases. Developers can use lighter, faster models while maintaining great performance.
Deepseek R1 can be downloaded from Github.
DeepSeek 50 Times Lower Cost
DeepSeek hsa gotten great results with a lot less computational resources compared to what is typically required for training models of similar capabilities. DeepSeek offers competitive performance at about 2% of the cost, both in terms of training and inference.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Open AI is in so many ways in both humans, and nations “think of them, and try to dominate them” in the same “mindset” of game changing technology before. But AI is unlike so much other “stuff” because it’s so, weirdly predictable. The transistor, integrated circuit, the “DARPA net, becoming the Internet”, hell Kevlar, became part of our lives not because people set out to invent these things. But intelligent men and women, got answers to questions they never thought to ask, and recognized “interesting thing’s”, when they saw not what they were looking, for, but recognized something, “very interesting”. A good scientist (or entrepreneur) see’s “possibilities” in what they observe, but may have not been looking for. But they see, know, and FEEL that. When that “ah ha!” moment happens, which for me, has been a few times in my life, that joy, is absolutely, priceless.
AI is “weird” because it seems so pre-planned, like a recipe, or “5year plan” in the late Soviet Union. (Those 5year plans, never worked out). AI is not “new” technology, it’s the application of technology we’ve had a while, in unanticipated and dynamic ways. Point: Ways we CAN”T anticipate. The potential is great. But I’d feel better if a person, was “in the loop” and said, “I’m not comfortable going there”. That person might be wrong, But I would listen to that person. Hope the computer would, or could, as well.
So called “deep data”, IMO has an inherent handicap. Data is not information, information is not wisdom, and wisdom is not insight. For a long time, we’ve been able to intercept almost all the electronic communication from anywhere on Earth. We’re really very good at that. I like that. Problem? It’s not knowing what people say, (rather easy), it’s knowing what they mean (aka: intend to do) Much harder… AI process “data” based on anticipatory metrics. It builds it’s “data base” based on accumulating “data” . and associating one reaction, will inevitably lead to another. Uhm, sorry, human behavior is both very adaptable, and deeply, asymmetric. (Translation? Means it really follows a straight light from 1to2to3, etc.)
Hell I’ve had computers who got a 1or0 “out of joint”, in several millions line of code, and the damn primitive thing almost blew up in my face. Depressing, but true. Biology handles the “unexpected” better. It’s not “as fast” in “doing math”, but it’s much better at dealing with problems, it never imagined let alone dealt with, before.
This is cool because anyone worried that the CCP has somehow baked something nefarious into it can go through the source code with a fine tooth comb.
How much compute is needed to run the models? The market will purchase open source AI models pre installed on a home or business system. In order to be useful the model must be able to learn new data provided by the customer.
I would not trust communist Chinese, where “free” speech is so loved. By why doesnt Musk criticize Chinese and their anti free speech communist state?
Why? It would cost Musk money. It really depends on what values a person has. Or not.
This is big. A reasoner with full replication steps to make another. All the OSS SOTA candidates becoming worthless overnight.
Seems we are in a flaunting stage, a game of dare between rival nations.
While OpenAI and others restrict their top models and keep their secret sauce secret, China releases them with replication steps, probably to force their hand and erode the moats.
Which is ironic, given OpenAI was supposed to do that.