DeepSeek R1: The AI model you should be using

China has just blown the AI community away setting vibrations of panic across the West. Their newly released DeepThink-R1 model is just as intelligent as OpenAI’s o1, yet is 27x cheaper, open-sourced, smaller and was made in less time. 

So what exactly is DeepSeek-R1?

It is an LLM out of China, competing with the likes of ChatGPT, Claude and Gemini. Think of DeepSeek R1 as an AI with a curious, thoughtful mind that can access the web. Educated on vast, diverse data, it has a strong reasoning core that solves challenges through prioritisation, trial-and-error and reflection.

R1 employs a prioritisation technique called ‘Mixture-of-Experts’ so that it uses different parts of its "brain" for different problems. This method makes it cost-efficient, and allows it to tackle various tasks efficiently. It has developed its own way of taking notes internally, or what we could call "mental tags," which help it make sense of new information or problems it encounters. When faced with a challenge, R1 doesn't just jump to conclusions; it reflects, much like a thoughtful person might ponder over a decision, drawing on lived experiences.

One of the most human-like traits of R1 is its understanding that more time spent on thinking often leads to better outcomes. This wasn't explicitly programmed into it; rather, like someone learning from life, R1 figured out that patience and deep thought yield superior results. 

In essence, DeepSeek R1 has grown into an entity that learns, adapts, and optimises its approach to problems, much like a person would, by continuously expanding its understanding of the world. 

It feels like the next generation of AI

DeepSeek-R1 marks an evolution -  shifting from implementation models like GPT-4o, to new models that reason and are simply more capable. 

When we first tried DeepSeek-R1, it felt like uncanny valley. The intelligence is intelligent, but it’s thoughts felt eerily familiar. It thinks, controls attention, self-corrects, and is curious. These are all properties absent in older models yet hardwired into humans.

Our recent essay probed this initial thought gap, celebrating human intelligence because of our “unique” metacognition. But what happens when you face an AI that can imitate metacognition? An emergence of thought that feels like ours, yet is quicker, more precise, and larger in scale. This is how DeepSeek-R1 feels.

A key noticeable, was that R1 wasn’t just interested in our prompt, it cared about the meaning behind it. It tried to figure out why I asked a certain question, and the relevance this had to what I told it before. This intrigue, tells us two things:

1. R1 truly wants to know the user, so it can personalise it’s output with meaningful context. This made us want to engage with the model so much than with other LLM’s.

2. It can build a level of predictability of our behaviour. So, if in time these models can predict our prompts/actions, can it preempt our needs, offering solutions or information before we even prompt?

This is a must for project professionals

Planning is the essence of a project, and this an expert planner. At its core, it’s all about enhancing your approach to planning—helping you understand the “why” behind your project’s structure by reasoning through its tasks, context, and methods. It’s not just a tool; it’s a partner that helps you learn more about how you manage projects while improving your strategies along the way. Since it’s open-source, you can customise it to fit your unique needs, making it as adaptable as your projects demand. What’s more, its compact size means you can run it locally, giving you tighter control over your data and keeping your project information secure. With these new reasoning models, you’re not just managing projects—you’re redefining how they’re done.

Why was such a powerful model released as open-source?

Many believe DeepSeek was open-sourced as a step toward democratising AI, highlighting an innovative path where high-performance AI can be achieved with fewer resources, directly challenging the West’s dominance in the AI race. Despite U.S. export controls designed to restrict China’s access to advanced chips, DeepSeek’s decision to release open-source models has reshaped the landscape. Necessity is the mother of innovation.

By embracing global collaboration and open innovation, DeepSeek has demonstrated that hardware limitations are no barrier to creating and sharing competitive AI technology. This approach not only diminishes the impact of chip restrictions but also accelerates AI development on a global scale. Whilst posing a threat to the West’s AI bubble - R1 is a huge win for AI accessibility and we’re here for it.

3 things to look out for

Here is how you can prepare for what’s coming:

• Follow the US response. DeepSeek have redefined the approach to building powerful AI and given it to the world for free. Expect US players to follow and build on the blueprint.

• Expect AI startups to be powered by DeepSeek models. The open-sourced nature of DeepSeek’s models means that developers can now build and deploy advanced intelligence without heavy costs and with model visibility. No longer do you have to rely on expensive AI software.

• Anticipate DeepSeek’s next big model: They’re rightfully serving what OpenAI originally promised, and have purportedly done so with less time and resource. As US companies play catchup, DeepSeek will likely have something finer cooking.

Other details you should know about DeepSeek-R1

• DeepSeek R1 is a side project?!

• Make the model even more powerful with including XML tags in your prompts

• Use R1 within agent workflows

• Build DeepSeek’s Reasoning into other LLM’s

• PsyOp claims

• Read the official research paper

• R1 has an existential crisis?