So let’s take a look. Much has been made of Deepseek’s economical cost of development (reported $6m vs $100m+ for US based Large Language Models) with 2nd tier chips. Whilst there is debate on the extent of the budget, the technical chips (were they banned?), and distillation there is undoubtedly a step change and difference in the approach which has been brought forward by doing more with less, specifically computers. To give some context, the ability to run more data using fewer computers is huge - the financial cost and environmental cost of running these servers is enormous. ChatGPT-3 required reportedly 10,000 GPUs and used 700,000 litres for training and it is estimated to use 500 ml of fresh water for every 20–50 requests. Whilst we don't yet have like-for-like comparisons between these models, this efficiency is on the whole a good thing as it will bring faster, more accessible steps forward with less impact both business and planet.

 

The Strengths of Deepseek 

 

What’s really interesting in the comparison are the strengths and the “learning” approach of Deepseek. Deepseek uses Reinforced Learning; this approach has advantages. It does not require any labelling of data which can be costly and time consuming. This learning approach is also adaptive and works for complex problems. This capability is invaluable in real-world applications, such as autonomous driving or robotics, where conditions can change unexpectedly.

 

A further difference is its use of a “Mixture of Experts” architecture - the system has been designed to work on a variety tasks; these are smaller sections that can work together but critically focus on specific tasks. These lead to greater reasoning and abilities in dealing with technical tasks such as coding or logical problem solving. 

 

Deepseek is designed as an open source model - this fundamentally changes the way a user interacts with it. This breaks with OpenAI’s subscription model approach and democratises access. In the interests of brevity, I have included below a short table calling out more points of differences including bias, not to be overlooked.  

 

Having said all of that, Deepseek vs ChatGPT…does it matter? 

 

These differences are real and they matter when you are thinking about what you need an AGI system to do. 

 

But more important than this is the speed of innovation and efficiency we are starting to see. This breaks open the stranglehold and disrupts what could have been AGI’s ownership to large US based corporations with significant capital. Expect to see more Large Language Models open sourced. 

 

This battle between low cost open source vs closed & higher cost closed systems will bring even more benefits from AI to fashion: inventory analysis, listing speed & capacity, SEO optimisation, productivity, returns reduction, conversion increase & optimisation, scaled personalisation of offers, loyalty growth and trend analysis amongst many others!

 

AI powers these solutions right now.  

 

The question: what are you doing today to start to realise these opportunities? How are you starting - PoC? Trial? Test? Partnership? ….? Or are you waiting... what are you waiting for? 

 

I’d love to hear your thoughts.