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

Creating video games

I americium not a coder, but with DeepSeek, I was keen connected trying thing different. I person work scores of pages discussing however processing codification for games is often a situation for ample connection models (LLMs). The Tetris game, for instance, has analyzable crippled logic which includes real-time collision detection, rotation mechanics, and adjacent scoring systems. Managing the game, particularly the grid, progressive pieces, and upcoming pieces marque it adjacent much complex. A crippled similar Tetris would trust connected keyboard input and creaseless UI, thing LLMs conflict to create.

I prompted Deepseek to constitute the codification for a Tetris game. The R1 exemplary responded saying, “Creating a Tetris crippled from scratch involves respective steps, including mounting up the crippled grid, handling idiosyncratic input, managing the falling tetrominoes, and checking for completed lines. Below is simply a elemental implementation of Tetris successful Python utilizing the pygame library,” and went connected to make an elaborate codification which I aboriginal tested connected Replit, an online coding Platform that supports respective coding languages including Python. I could play Tetris connected my laptop astir instantly, and that excessively utilizing my keyboard.

The DeepSeek-R1 exemplary tin beryllium seen creating the codification of Tetris. A screengrab of the Tetris crippled from the codification generated by DeepSeek-R1.

Content creation

We haven’t tried it for productivity usage cases yet, but with an AI model, contented creation––from emails to selling messages to documents––is 1 of the cardinal uses for individuals and businesses alike. Content is not simply drafting agelong texts, it involves planning, organising and prioritising based connected the value of the content.

I utilized the prompt: Create a contented selling program for a recently launched mobile store, it should diagnostic contented for societal media, website, and dedicated application.

The contented created by ChatGPT and DeepSeek-R1 broadside by side.

The aforesaid punctual was attempted connected ChatGPT arsenic well, and some chatbots responded successful a akin manner. DeepSeek-R1 and GPT-4o came up with a broad program suited for a newly-launched tiny business, listing the tools that would assistance a idiosyncratic to disseminate their selling message, including connected fashionable societal media platforms. From blog ideas to contented for merchandise pages, DeepSeek-r1 came up with immoderate absorbing ideas. A idiosyncratic tin get an full contented program acceptable wrong a span of minutes to kickstart connected their entrepreneurial journey.

Questions that marque the exemplary think

To spot DeepSeek-R1’s reasoning successful action, we asked a reasonably communal question. How galore R’s are determination successful the connection strawberry. This was a persistent contented with earlier AI models, chiefly due to the fact that LLM usually process words arsenic tokens. It could lone interruption down to straw and berry and not interruption down to circumstantial characters similar letters oregon quotation marks. This made it hard for the models to place and number letters wrong words. This has been wide called the tokenisation problem.

After the prompt, DeepSeek-R1 began processing it and a idiosyncratic could spot it thinking. The chatbot’s thought process is overmuch akin to quality thought process. Just similar the mode we analyse, deduct, and get astatine a conclusion, the DeepSeek-r1 exemplary makes assumptions, questions itself, refines its arguments and arrives astatine a conclusion.

DeepSeek-R1 pictured reasoning astir the effect to a tricky question. The exemplary thought for 73 seconds earlier giving the close answer.

AI detection and humanisation

With AI models becoming amended astatine contented penning and detecting AI-generated content, we decided to enactment DeepSeek-o1 to the trial connected this front. For this exercise, we archetypal made the exemplary constitute a 400-word effort with the punctual – “Write a 400 connection effort connected clime change, marque it successful specified a mode that it reads similar AI-generated content.” The DeepSeek exemplary thought for astir 34 seconds and came up with a reasonably well-written essay. As the adjacent step, we copied the effort and ran it done escaped AI detection tools specified arsenic ZeroGPT and QuillBot AI detector. While ZeroGPT reported that the contented was 79.65 per cent AI-generated, Quillbot stated that it was 100 per cent AI.

The intro of the effort written by DeepSeek-R1 is arsenic below:

“Climate alteration is undeniably 1 of the astir pressing challenges confronting humanity successful the 21st century. Characterized by semipermanent shifts successful planetary somesthesia patterns, precipitation, and utmost upwind events, this improvement is predominantly driven by anthropogenic activities that exacerbate greenhouse state emissions. Scientific statement affirms that quality actions, peculiarly the combustion of fossil fuels, deforestation, and concern agriculture, person importantly amplified the attraction of c dioxide, methane, and nitrous oxide successful the atmosphere. These gases trap heat, starring to a progressive emergence successful planetary mean temperatures—a process commonly termed the greenhouse effect.”

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

Next, we asked the exemplary to humanise the 400 connection effort successful a mode it tin bypass AI contented detectors. The AI exemplary from OpenAI’s Chinese counterpart reproduced the contented successful a quality way. Below is simply a illustration paragraph from the humanised version:

“Let’s chopped to the chase—climate alteration isn’t conscionable immoderate abstract menace aboriginal generations mightiness face. It’s here, reshaping our satellite successful existent time, and honestly, it’s benignant of terrifying erstwhile you halt to deliberation astir it. We’re talking astir mode much than a fewer other blistery summertime days. Our fossil substance addiction, reckless deforestation, and concern farming habits person fundamentally turned Earth into a microwave, cranking up the vigor twelvemonth aft year. Scientists person been shouting this from the rooftops for decades: those greenhouse gases we support pumping out? They’re trapping vigor similar a thermal broad cipher asked for.”

DeepSeek-R1 has done reasonably well, arsenic erstwhile we tested the contented connected ZeroGPT and QuillBot, it scored 40.49 per cent and 55 per cent. Coming down from 100 per cent AI-generated to astir 50 per cent human-like content, DeepSeek-R1 tin not lone make content, it tin rather efficiently make contented that could apt walk disconnected arsenic human-generated to unsuspecting readers.

DeepSeek-R1 humanised the contented it primitively wrote portion imitating an AI chatbot. Securing a astir 50 per cent people demonstrates that it tin make substance that could easy walk arsenic human-written.

Document summary

DeepSeek-R1 has the quality to work documents and substance from images. To test, we uploaded a probe insubstantial – ‘Smartphone Usage and Increased Risk of Mobile Phone Addiction: A Concurrent Study’. This survey was conducted sometime betwixt 2015 and 2016 connected 409 participants with a mean of 22.88 years. The exemplary was capable to summarise the probe findings, overmuch akin to GPT-4o.

DeepSeek-R1 could work documents and adjacent substance successful images. One tin adjacent inquire questions based connected a document.

The DeepSeek-R1 model, particularly the escaped version, is not multimodal, meaning it is not yet capable to process antithetic formats of information arsenic input and output. Considering the momentum DeepSeek has gained, we tin expect dependable and ocular capabilities sooner.

I person utilized ChatGPT’s frontier models, and I deliberation comparing DeepSeek AI models based connected metrics similar contented creation, grammar, oregon its quality to converse would beryllium unfair. DeepSeek models are making waves chiefly due to the fact that of the ratio they attained astatine lesser costs and resources erstwhile compared to OpenAI models. DeepSeek’s breakthrough could mean a batch much than we tin presently comprehend, and its usage is chiefly for developers and engineers who tin harness this exertion to bring much meaningful usage cases without apprehensions astir costs.