The world of AI is a fast-moving place, shown very clearly by the arrival a few weeks ago of a new AI-model from a Chinese company called Deepseek. This caused a stir with the release of their V3 model, the company’s logo of a whale being very fitting as it certainly made a splash!  

Although perhaps the logo of a shark may have been more appropriate given the effects, it immediately wiped almost $600bn (£482bn) off the market value from Nvidia, the biggest one-day loss in US history, and set panic amongst American AI companies and AI policy. But why the fuss and how does this relate to the future of AI in relation to education? 

Why is it such a big deal?

So there were a few reasons this was such a big deal, Thinking model of free rivals Open AI the Thinking model which was its low cost, power-usage and the fact it’s been made open-source. Let’s explore what each of those points means. 

Cost

DeepSeek was reportedly trained on a budget of 6 million dollars, a fraction of the cost compared to the American models, and the results have been that DeepSeek V3 is comparable to the most powerful models that competitors like ChatGPT have developed, more importantly DeepSeek is able to give access to their model for free. Whereas you would have to pay almost $200 a month to get access to ChatGPT’s comparable model. 

From testing the results these are essentially the same, OpenAI has a few more features but the models processing and speed are roughly on par. 

Power usage

AI models need a lot of processing power to train them, the best way to do this is to use what are called GPU’s (Graphic processing units) which are able to do advanced processing of data. Now America has restricted Chinese access to Nvidia’s H100 GPU’s because of their usefulness to training AI models. By restricting them with the American policy aimed to prevent China developing strong AI models or so it was thought! 

But necessity is the mother of invention and through Chinese companies like DeepSeek being forced to use much weaker GPU’s, they used Nvidia’s much slower H800 GPU’s, they were able to work out ways of working smarter, programming models that got the same result using much less processing power. DeepSeek researchers then published a paper showing how they did this and how others can do the same. 

Open source

DeepSeek’s model has also been released as open-source, this means that the model can be used for free by anyone locally meaning it would run entirely from an individuals own computer (there is a cut down model that can be hosted on a reasonably powerful computer and a full model which requires a very powerful computer to run so it’s still out of reach for most people but closer than it was). This is in opposition to the other AI models such as OpenAI Google and Microsoft’s which are closed meaning you cannot run your own version of it but can only run it from those companies’ servers.  

Running your own model means you have complete access to how your data is used and aren’t beholden to a company which has now clear advantages we’ll cover later. 

Should we be using it?

There’s no real reason to be less trustful of DeepSeek then any other model, DeepSeek was created in China and so has some censorship built into it due to the Chinese government, however American companies’ models have strong connections to the current American government and their own biases and potential data issues.  

You could argue that DeepSeek is currently the most ‘trustworthy’ (used loosely) of the big models or at least the best of a bad bunch given their release of an open-source model and free sharing of their research compared to the American companies secrecy. Also, at the time of writing, the only one to not have a contract with defence or defence adjacent weapons companies. (Google having recently withdrawn their policy against developing weapons systems). It also uses the lowest amount of power to run queries making it the most “sustainable” (again used loosely) of the current crop of models. 

We should be using the same critical thinking when using any AI model regardless of which company it comes from, this means don’t put any private or personal information into any AI model that’s not been approved for use by the institution. Don’t download any AI tools directly to a computer unless it’s been approved, assume all models will have biases and potentially be aligned to certain views and values. Don’t assume anything it tells you is automatically true, it’s healthy to apply this to any AI model you use regardless of where it’s come from. 

The ideal path forward for AI models at Universities 

DeepSeek does however point us in the direction of a more ethical and safer future for AI, which is to have smaller more efficient open-source models that are locally hosted and run by individuals or institutions, entirely under their own control without the data being taken or used by an outside company. This would also allow for more ethical models with less bias and censorships. Currently there are cost barriers to these but a model like DeepSeek can run on a computer that costs a fraction of what ChatGPT and similar take to run. 

The advantages of these are multifaceted, first is a decoupling from what are largely unethical and unsustainable companies, meaning that students and staff could use AI models without worrying that their data is being harvested. Secondly, models could be tuned that are more free of biases and censorship,  

Of course, for this to come to fruition we’d need a few more technological leaps to make even smaller models that use much less power, but DeepSeek’s progress is an encouraging step in the right direction and more competition will hopefully breed more innovation.  

For better or worse we cannot avoid the proliferation of AI into our society and institutions, and so taking an optimists view, I’d like to hope we can move to a less corporate ownership model and put such tools in the hands of Universities that can use them for the public good. We can but hope!