In case you missed it, the tech and financial press has been getting VERY excited about DeepSeek, a powerful new entrant into the AI race from China. The implications of a new challenger to the AI throne has wide-ranging consequences for geopolitics, censorship and how we access large language models. Up until now, the AI throne had been passed between US-based AI heavyweights such as OpenAI, Google, Anthropic, Meta, and xAI.
There has been so much written about DeepSeek this week that it’s been hard to synthesise it all, and this is coming from someone who has been excited about pairing it (or a similarly optimised LLM) with an Nvidia Digits personal AI supercomputer as soon as I can get my hands on one! Rather than just opine on my own, I thought I would collate the thoughts of some of my favourite thinkers leading the conversation about the dawn of the AI Era.
According to Ethan Mollick, Associate Professor at The Wharton School and author of Co-Intelligence (the best book I’ve read on AI to date), the DeepSeek model is the first non-US state-of-the-art AI, trained at a fraction of the cost than competing models. While privacy and censorship concerns exist using DeepSeek’s platform, alternative hosts will emerge as it is an open source model that can be hosted on your own servers.
One of DeepSeek’s strategies was to start with Meta’s existing Llama LLM and fine-tune it by using reinforcement learning methods to boost the AI's reasoning capabilities without heavy reliance on human intervention. This approach allowed them to train the model at a fraction of the typical cost - around $5.6 million, compared to the $100 million or more that others have spent.
Arshad Hisham, founder and CEO of InGen Dynamics, recently documented the DeepSeek journey and its agile and innovative approach has not only overcome the US -ban on supplying high-performance AI chips to China but also how it has done so using unorthodox hiring philosophies. He also got DeepSeek working on his MacBook M1 Max laptop - being able to run a 70B parameter on a laptop is pretty wild!
DeepSeek’s models face restrictions preventing content critical of China’s government from bring generated: If one asks DeepSeek about the events that occurred on June 4, 1989 in Tiananmen Square they are met with a message saying "Sorry, that's beyond my current scope. Let's talk about something else.'"
Using the DeepSeek-hosted model could also leave users open to feeding personal or confidential information to a company that is not free from the watchful eye of the Chinese-state. However, it remains to be seen just how "safe" those Ai's developed in the US are – the weaponisation of AI as a geopolitical tool could well be about to start.
However, Reuven Cohen, Agentic Engineer, aiCTO and Consultant, has shown that DeepSeek R1 can be hosted using Google Cloud Run and even developed a step-by-step tutorial on how to create an uncensored/self-hosted version of the LLM. The guide enables users to set up a scalable API that handles both model inference and fine-tuning, enabling unrestricted AI interactions.
No sooner had it been announced the DeepSeek app had topped downloads on major app marketplaces, the AI startup suffered a significant cyber-attack that forced the company to temporarily limit new user registrations outside of China while allowing existing users to continue accessing their accounts.
According to Paul Ziegler, CEO of cybersecurity training company Reflare, such an incident was to be expected and has raised concerns about data privacy and security, particularly given DeepSeek’s operations in China. Critics have highlighted potential risks of state surveillance and inadequate regulatory oversight over user data. He stresses the incident highlights the importance of integrating robust cybersecurity measures into AI systems and their usage, but is also excited that it provides optionality for those wanting to host LLMs locally and avoid overreliance on US-developed and hosted AIs.
What is so exciting about this is that it brings the possibility of a locally hosted LLM even closer – one that could avoid the challenges of having to send data into the cloud at all!
My marketing hero, Mark Schaefer, wrote an excellent article discussing five positives to come from the DeepSeek apocalypse – it’s an article I highly recommend everyone reads.
Most intriguingly, based on some of the conversations I’ve had around the sustainability (and lack thereof) of AI usage, Mark highlights how this may impact Meta’s announcement of a plan to build a $60 billion data centre in Louisiana that would be nearly as big as Manhattan – with the energy demands to match.
The era of assuming AI is just a technical tool is over - its strategic implications are now front and centre. Adam Cox, Managing Director of the strategy advisory firm adamcox.com commented:
“DeepSeek’s arrival isn’t a surprise - and seasoned strategists have been expecting credible Chinese entrant for some time. Now that the real AI race has begun, expect more competitively credible disruptors to follow. What is interesting is that the flawed assumption that AI dominance would remain exclusively American was clearly held by many, evidenced by how the DeepSeek announcement spooked the markets. However, what deserves greater focus is that business leaders beyond just the CIO and COO must now uncover and assess the potential technical debt arising from the geopolitical risks, and how those impact their AI dependencies and their businesses as large.”
DeepSeek’s emergence underscores how quickly competitive advantage can vanish in the AI era. Security, regulatory constraints, and national interests will play a larger role in AI adoption than ever before, and the companies that prepare for this reality will be the ones that ride the wave of a dynamic and fast-evolving marketplace.
If you’re looking for help in plotting your approach to implementing AI in your marketing function, then we’d always be happy to take a call.