DeepSeek R1 hands-on: 5 things we tried, including developing a game

Creating video games

I am not a coder, but with DeepSeek, I was keen on trying something different. I have read scores of pages discussing how developing code for games is often a challenge for large language models (LLMs). The Tetris game, for instance, has complex game logic which includes real-time collision detection, rotation mechanics, and even scoring systems. Managing the game, especially the grid, active pieces, and upcoming pieces make it even more complex. A game like Tetris would rely on keyboard input and smooth UI, something LLMs struggle to create.

I prompted Deepseek to write the code for a Tetris game. The R1 model responded saying, “Creating a Tetris game from scratch involves several steps, including setting up the game grid, handling user input, managing the falling tetrominoes, and checking for completed lines. Below is a simple implementation of Tetris in Python using the pygame library,” and went on to create an elaborate code which I later tested on Replit, an online coding Platform that supports several coding languages including Python. I could play Tetris on my laptop almost instantly, and that too using my keyboard.

The DeepSeek-R1 model can be seen creating the code of Tetris.

A screengrab of the Tetris game from the code generated by DeepSeek-R1.

Content creation

We haven’t tried it for productivity use cases yet, but with an AI model, content creation––from emails to marketing messages to documents––is one of the key uses for individuals and businesses alike. Content is not merely drafting long texts, it involves planning, organising and prioritising based on the significance of the content.

I used the prompt: Create a content marketing plan for a newly launched mobile store, it should feature content for social media, website, and dedicated application.

The content created by ChatGPT and DeepSeek-R1 side by side.

The same prompt was attempted on ChatGPT as well, and both chatbots responded in a similar manner. DeepSeek-R1 and GPT-4o came up with a comprehensive plan suited for a newly-launched small business, listing the tools that would help a user to disseminate their marketing message, including on popular social media platforms. From blog ideas to content for product pages, DeepSeek-r1 came up with some interesting ideas. A user can get an entire content plan ready within a span of minutes to kickstart on their entrepreneurial journey.

Questions that make the model think

To see DeepSeek-R1’s thinking in action, we asked a fairly common question. How many R’s are there in the word strawberry. This was a persistent issue with earlier AI models, mainly because LLM usually process words as tokens. It could only break down to straw and berry and not break down to specific characters like letters or quotation marks. This made it difficult for the models to identify and count letters within words. This has been widely called the tokenisation problem.

After the prompt, DeepSeek-R1 began processing it and a user could see it thinking. The chatbot’s thought process is much similar to human thought process. Just like the way we analyse, deduct, and arrive at a conclusion, the DeepSeek-r1 model makes assumptions, questions itself, refines its arguments and arrives at a conclusion.

DeepSeek-R1 pictured thinking about the response to a tricky question. The model thought for 73 seconds before giving the correct answer.

AI detection and humanisation

With AI models becoming better at content writing and detecting AI-generated content, we decided to put DeepSeek-o1 to the test on this front. For this exercise, we first made the model compose a 400-word essay with the prompt – “Write a 400 word essay on climate change, make it in such a way that it reads like AI-generated content.” The DeepSeek model thought for about 34 seconds and came up with a fairly well-written essay. As the next step, we copied the essay and ran it through free AI detection tools such as ZeroGPT and QuillBot AI detector. While ZeroGPT reported that the content was 79.65 per cent AI-generated, Quillbot stated that it was 100 per cent AI.

The intro of the essay written by DeepSeek-R1 is as below:

“Climate change is undeniably one of the most pressing challenges confronting humanity in the 21st century. Characterized by long-term shifts in global temperature patterns, precipitation, and extreme weather events, this phenomenon is predominantly driven by anthropogenic activities that exacerbate greenhouse gas emissions. Scientific consensus affirms that human actions, particularly the combustion of fossil fuels, deforestation, and industrial agriculture, have significantly amplified the concentration of carbon dioxide, methane, and nitrous oxide in the atmosphere. These gases trap heat, leading to a progressive rise in global average temperatures—a process commonly termed the greenhouse effect.”

QuillBot and ZeroGPT impeccably detected the AI-generated content written by DeepSeek-R1.

Next, we asked the model to humanise the 400 word essay in a way it can bypass AI content detectors. The AI model from OpenAI’s Chinese counterpart reproduced the content in a human way. Below is a sample paragraph from the humanised version:

“Let’s cut to the chase—climate change isn’t just some abstract threat future generations might face. It’s here, reshaping our world in real time, and honestly, it’s kind of terrifying when you stop to think about it. We’re talking about way more than a few extra hot summer days. Our fossil fuel addiction, reckless deforestation, and industrial farming habits have essentially turned Earth into a microwave, cranking up the heat year after year. Scientists have been shouting this from the rooftops for decades: those greenhouse gases we keep pumping out? They’re trapping heat like a thermal blanket nobody asked for.”

DeepSeek-R1 has done fairly well, as when we tested the content on ZeroGPT and QuillBot, it scored 40.49 per cent and 55 per cent. Coming down from 100 per cent AI-generated to nearly 50 per cent human-like content, DeepSeek-R1 can not only create content, it can quite efficiently generate content that could likely pass off as human-generated to unsuspecting readers.

DeepSeek-R1 humanised the content it originally wrote while imitating an AI chatbot. Securing a nearly 50 per cent score demonstrates that it can generate text that could easily pass as human-written.

Document summary

DeepSeek-R1 has the ability to read documents and text from images. To test, we uploaded a research paper – ‘Smartphone Usage and Increased Risk of Mobile Phone Addiction: A Concurrent Study’. This study was conducted sometime between 2015 and 2016 on 409 participants with a mean of 22.88 years. The model was able to summarise the research findings, much similar to GPT-4o.

DeepSeek-R1 could read documents and even text in images. One can even ask questions based on a document.

The DeepSeek-R1 model, especially the free version, is not multimodal, meaning it is not yet able to process different formats of data as input and output. Considering the momentum DeepSeek has gained, we can expect voice and visual capabilities sooner.

I have used ChatGPT’s frontier models, and I think comparing DeepSeek AI models based on metrics like content creation, grammar, or its ability to converse would be unfair. DeepSeek models are making waves primarily because of the efficiency they attained at lesser costs and resources when compared to OpenAI models. DeepSeek’s breakthrough could mean a lot more than we can currently comprehend, and its usage is mainly for developers and engineers who can harness this technology to bring more meaningful use cases without apprehensions about costs.