4 min read
“How To Leverage DeepSeek R1, And Surpass The 99%”
Introduction
Recently, there’s been a new AI technology blowing up, going by the name of ‘DeepSeek’. It’s like ChatGPT, but completely free-to-use, with access to its most advanced model, and safe to use.
In this post, we’ll go over 1) What DeepSeek is, 2) How To Use DeepSeek, and especially: 3) How to use DeepSeek and maximize it as a student. Answering all of these questions will be the central focus of this post, and hopefully, they’re your central goals too.
We haven’t made a post solely on AI yet, and we’re glad that this is the first. If you’re new here, we go by ‘GraduationGurus’, or ‘GG’ for short; and here, we look to answer student’s most common questions and resolve their biggest concerns.
Here, we’ll be doing something a little different, so, we hope you like it.
Background (not important)
Feel free to skip this section and move onto Part 1. However, if you’re interested in the context of the AI, and what it is, then you might want to read ahead.
DeepSeek is a Chinese Artificial Intelligence Company that has become a world (or just US/China) -renound name, gaining major attention in the AI landscape. The company was founded in July of 2023 (not even 2 years old yet) by founder Liang Wenfeng, where they built advanced AI models that rival those of major US companies.
The company offers a free AI-powered chatbot, just like ChatGPT, which can understand and generate responses in Chinese and English. Their latest model, going by R1, has even out-performed several other models like those from Google, Anthropic, Meta, and OpenAI. Keep in mind: this was all done on a SIGNIFICANTLY smaller budget.
When we say, “smaller budget”, we’re talking 20x-50x smaller. Yeah, that much. Unlike these major US companies, it only (relative to, pretty much all other AI models in the space) costed DeepSeek around $5.8 million to build the model, unlike these other bigger companies, which usually have between $150 million to $1 billion to spend JUST ON THESE MODELS. That’s the reason for all the craze.
What Makes It So Good
If you’re still interested on the inner gears behind the model, then you might want to check this out.
Firstly, we must say this: Although OpenAI already has their own AI model, going by ChatGPT o1, which is just like DeepSeek in all aspects (meaning that it already existed), DeepSeek R1 is special because of how it was able to replicate the model for so much cheaper; while simultaneously revealing the curtains behind OpenAI’s “Strawberry thinking model”. This “strawberry” model allows users to ask complex queries to AI, and recieve well-thought answers where ChatGPT takes its time to thinking and create an answer. DeepSeek R1 replicated exactly this, revealing the AI’s “thinking” process in the conversation directly, and making it available to users for free (which OpenAI charges for).
Ok, back to how it works. Firstly, again, DeepSeek utilizes a technique named “inference-time computing”, which focuses on improving model performance during the inference phase. Put it like this:
- R1 is firstly trained.
- Techniques are applied to the model after training (to enhance performance) without re-training.
- methods include dynamic model adjustment, ensembling, quantization, and batch processing
- The goal is to balance accuracy and efficiency according to system constraints
Put simpler, this is when the model adjusts it computational effort based on the complexity of the query.
What Makes It So Good (cont.)
The next part of the model’s functionality is through an architecture by the name of ‘Mixture-of-Experts’, in which DeepSeek segments its larger model into numerous smaller submodels, or ‘experts’, with each one specializing in a parcticular task. Only the relevant submodel is used when its expertise is necessary for the present situation.
This means that R1 can esentially niche-down its knowledge to specific sectors of itself to provide better-quality responses, and enhance performance at the same time.
Some other cool, but not as cool features behind its functionality are:
- Only utilizes 37 billion parameters (not as many as it seems) despite the most advanced and complete modeling requiring 671 billion parameters –> saving on computational effort
- It uses FP8 low precision processing for most operations, with the selective use of higher-precision choices whenever necessary,
- DeepSeek uses dynamic adjustments instead of a typical penalty-based system to prevent ‘expert overload’ (no idea what this is)
- This is critical considering that most reinforcement learning models leverage (or maybe not in this case) penalty systems to punish the model for incorrect answers, and reward it for good answers. It’s really interesting.
Why Any Of This Matters
DeepSeek is a, what you could call ‘breakthrough’, that has meant for much for the field of AI. And because it’s open-source, anyone can take advantage of it to learn more about how it works and use it too in their own applications – something OpenAI does not offer.
It means so much for the growing space of AI, and makes a bigger and bigger questions as to how close we are to AGI (Artificial General Intelligence), just 1 step behind Super Intelligence. It means so much for growth of human potential, as well as the potential for Artificial Intelligence, across all fields.
How To Use DeepSeek R1 (2 Methods)
First off, in case you’re still wondering why you can’t just stick with ChatGPT, or Gemini, you could! The only reason we really want to use DeepSeek in the first place is because of its thinking model, which offers capabilities that in other AI applications, you have to pay for.
Method #1: Search Engine
The first, most straight-forward way, is to go to DeepSeek Official Website, and access the model directly there. It’s easy, and just like that, you can use a GPT-like model right from your device.
Additionally, there’s also the mobile application, which also might likely require to sign up for an account (I’m not sure as I haven’t made an account).
However, if you don’t feel all that secure with giving some of your data to their company, you can do the second method instead.
Method #2: Downloading The Model
This option takes a few more steps, but nonetheless, is still really easy to do.
- First, you need to download Ollama. This is used to run LLMs right from your device, and not on their servers.
- Next, you need to install Ollama and give it permissions to your device.
- Third, you need to open up a terminal. On Windows, this is called “terminal” on your search bar. You can search for it directly, and it should open up an application called “Windows Powershell”.
- Once you’ve gotten to this point, on the Ollama homepage, you need to click on the link directing you to the DeepSeek-R1 Model. Here, you’ll see a bunch of different model options to choose from. The most lightweight model is the 1.5b model. This takes up the least amount of RAM on your device, and should run the smoothest.
- However, other options like the 7b, 8b, and 14b could run just as smoothly depending on the amount of available RAM in your computer/laptop. Make sure to check (unless you are already aware) how much you actually have, and download according to your device. If you don’t, the model you select likely will not work as expected, and will run poorly.
- As you might see, the most advanced option (the model that was tested in the metrics against OpenAI’s o1 model) is the 671b option. This requires like several NVIDIA GPU processors to run, so don’t even bother even looking over that one.
- However, other options like the 7b, 8b, and 14b could run just as smoothly depending on the amount of available RAM in your computer/laptop. Make sure to check (unless you are already aware) how much you actually have, and download according to your device. If you don’t, the model you select likely will not work as expected, and will run poorly.
- Once you’ve selected a model, copy-and-paste its download code statement and paste it into your terminal. This shouldn’t take very long to download.
- Done! You can now use DeepSeek’s R1 Model right from your computer.
How To Download R1 (Video Demonstration)
If a video-format is easier to understand, I recommend checking out the video below. I personally found it to be the most easy-to-understand and to-the-point tutorial there was.
How To Leverage DeepSeek As A Student
First, let’s get something straight. DeepSeek is a good option for those looking to to simplify concepts and receive human-like feedback on the concepts they’re looking to master – something that ChatGPT kind of lacks.
Point is: if there’s a specific study that you’re in that is just hard to grasp, and that involves mostly text-format explanations, R1 is an excellent option. For example, if I were in the shoes of a regular high schools student, DeepSeek might be the most applicable to:
- Math
- Sciences (Chemistry, Physics, Biology)
- History
- Languages (English, Spanish, French, etc.)
- any others that can be explained by words
And…that pretty much describes it. If it can be explained with words, R1 can definitely explain it too.
How To Really Use R1
- DeepSeek can break down complex concepts into easy-to-understand explanations tailored to the student’s specific level of understanding.
- It can also generate customized study guides, summaries, and as I said previously: anything with words.
- MOST NOTABLY: DeepSeek has a unique feature that sets it apart from other AI models. It thinks, and it shows what it thinks and how it gets to where it ends up. If it makes a mistake, or gets something back-tracked, you’ll know where. From there, you can point out a mistake if you see one, and it will correct itself; leading itself back to where it got wrong, and get it right.
- This is unlike ChatGPT or Gemini, where they tend to suffer a problem known as hallucinations. Under this, a model can tend to falsify information, and claim it as right – when it isn’t.
And, that’s pretty much it. That’s how you can use R1, and use it right.
How To Surpass 99% Of Other Students
Chances are: not everyone is going to take advantage of R1 as you might. Sure, some here and there might, but a majority probably won’t.
The goal here is to do better at academics, and to use it in order to strengthen your current academic capabilities – not to weaken them. A common mistake I see with the application of AI in school environments is that they tend to be abused, and “leveraged” to the point that it becomes “cheating”, and unethical in an academic sense. The goal here is to solely reinforce your knowledge, and make your human capability even higher – not to lower it.
If you can use R1 in that exact fashion, to absorb information at faster speeds, you will surpass anyone that dares to challenge your position.
Incorporate it into your daily routine of tasks, using it to streamline what would be the regular process of researching and gathering information, and you scores and grades will skyrocket.
Conclusion
Artificial Intelligence is a very potent tool capable of making or breaking almost anything. It can add layers to your learning structure and routine, and simplify the stress and maximize the efficiency of what you do on the regular, or turn you into a lazy and unmotivated person with no desire to dedicate any effort to learning and progressing.
Leverage it appropriately, and you will reach levels of success beyond the and above the clouds.
