DeepSeek's Disruption: Redefining AI Development Economics 🖥️
By Polly Barnfield, OBE, CEO of Maybe*
The artificial intelligence landscape has just experienced a seismic shift, and it's not coming from Silicon Valley. A team of 150 engineers in Hangzhou, China, has accomplished what many thought impossible: developing a competitive AI model at a fraction of the cost and time typically associated with such endeavours. This achievement by DeepSeek isn't just another AI advancement—it's a fundamental challenge to the established paradigms of AI development.
The Economics of Innovation
At the heart of DeepSeek's breakthrough lies a compelling economic narrative. While Silicon Valley engineers command salaries upwards of $700,000-$1,000,000 annually, DeepSeek's team operates at roughly one-tenth of that cost. This dramatic difference in operational expenses challenges the conventional wisdom that cutting-edge AI development requires massive capital investment.
The implications are profound. DeepSeek's success demonstrates that with the right team and approach, developing sophisticated AI models doesn't necessarily require billions in funding. This democratisation of AI development could spark a new wave of innovation from unexpected quarters.
Open Source as a Disruptive Force
The most significant aspect of DeepSeek's approach is their commitment to open source. While industry giants like OpenAI and Anthropic maintain strict proprietary control over their models, DeepSeek has made their research and methodologies publicly available. This transparency not only advances the field but also poses serious questions about the sustainability of closed-source AI business models.
For venture capitalists who have poured billions into proprietary AI companies, DeepSeek's success is particularly troubling. The "moats" that justified these massive investments appear increasingly vulnerable to efficient, open-source alternatives.
Global Innovation and Political Implications
The success of DeepSeek carries broader geopolitical implications. It demonstrates that attempts to contain AI innovation within specific geographical boundaries may be futile. Even China's own government was caught off guard by DeepSeek's achievements, with the Prime Minister making an unexpected visit to Hangzhou to understand the development.
This situation highlights a crucial tension in global tech policy. While some advocate for restrictive measures to maintain technological advantages, DeepSeek's story suggests that innovation will find its way regardless of barriers. The narrative of "don't trust it because it's Chinese" may hold less sway in a world where data privacy concerns exist for both Western and Eastern technology companies.
Infrastructure and Market Evolution
The rise of efficient, open-source AI development could reshape the cloud computing and infrastructure landscape. Companies like AWS, CoreWeave, and Nebius stand to benefit as AI development democratises, potentially serving a broader base of developers rather than being primarily dependent on a few large AI companies.
Looking Forward
DeepSeek's achievement signals several potential shifts in the AI landscape:
1. The democratisation of AI development, with more teams able to attempt ambitious projects at reasonable costs
2. A potential shift away from the venture capital-driven model of AI development
3. The increasing importance of efficiency and transparency in AI development
4. A more globally distributed AI innovation ecosystem
For the broader tech industry, this serves as a reminder that innovation often comes from unexpected places and that artificial barriers to competition—whether financial, political, or geographical—rarely stand the test of time.
The success of DeepSeek doesn't necessarily spell doom for established players, but it does suggest that the future of AI development may look very different from its past. Companies and investors would do well to take note: in the rapidly evolving world of AI, efficiency and openness might prove more valuable than deep pockets and closed systems.
Secure Access and Data Protection
Maybe* provides all Your AI users with access to DeepSeek's powerful capabilities via your access though Perplexiy. This ensures robust data protection and that your data is not being shared with China
Aravind Srinivas, CEO of Perplexity AI, has made it explicitly clear:
"All DeepSeek usage in Perplexity is through models hosted in data centres in the USA and Europe. DeepSeek is open-source. None of your data goes to China."
Perplexity uses the open-source DeepSeek model but hosts it on their own servers located in the United States and Europe. This means that while DeepSeek's own services may store data in China, Perplexity's implementation of DeepSeek does not send user data to Chinese servers. This is the route that Maybe* uses so that all Your AI users have secure access to the DeepSeek LLM too
It's important to note the distinction between using DeepSeek directly and using it through Your AI and Perplexity. While DeepSeek's own privacy policy states that they store information on servers in China, this does not apply to the approach that Maybe* have take with Your AI via Perplexity. i
Using DeepSeek's direct applications where data may be transmitted to servers in China, our implementation maintains complete data sovereignty. All user interactions and data remain secure within our infrastructure, never leaving our protected environment. This approach allows organisations to leverage DeepSeek's advanced AI capabilities while maintaining compliance with data protection regulations and security requirements.
This hybrid approach represents the best of both worlds: access to cutting-edge AI technology while ensuring data privacy and security remain paramount. Users can confidently utiliSe DeepSeek's capabilities knowing their sensitive information remains protected and controlled within our secure infrastructure.
Technical Notes From the Maybe* Team
The use of reinforcement learning rather than extensive supervised learning is actually more along the lines of traditional machine learning. The fact that you can see these same emergent properties organically is both encouraging and impressive.
The reinforcement learning uses rules-based rewards and learns by comparing against it's current policy (approach) vs some iteration it comes up with. It compares the range of previous answers given by it's last approach to the new one and then adopts the new policy if it gets a reward (again, traditional gradient descent stuff - but the group bit is important here)
Ability to improve existing models, but doing a form of distillation to generate smaller models that perform equally well (I've previously referred to this as compression)
The model appears particularly well-tuned for maths problems but gets a big win using majority voting. This means it'll come up with answers to the same question several times, and go with its most common answer (tradeoff is compute power/time needed at runtime)
A small amount of 'cold start' fine tuning can help, and the beginning - this seems to be more getting the output in the format you'd expect rather than anything else
Limitations
function calling and JSON support is still limited (wasn't a focus)
struggles with language mixing and was primarily trained on Chinese/English (will give answers half in Chinese/English or answer in the wrong language)
Works better with zero-shot prompting (i.e. it gets worse when given examples -->) - the tradeoff is it should give more correct answers, which is why it is better at maths than the LLMs that go before it.
What do DeepSeeks breakthroughs mean
DeepSeek's breakthrough indeed represents a significant technical achievement and a challenge to existing assumptions about AI development. The company's ability to create a model comparable to leading AI systems at a fraction of the cost demonstrates that innovation and efficiency can be as crucial as substantial financial investment.
The open-source nature of DeepSeek-R1 is democratising AI development, making advanced technology more accessible to researchers and developers worldwide, particularly those with limited resources. This approach fosters innovation and could lead to a more inclusive AI ecosystem
DeepSeek's success also highlights the potential for efficient resource utilisation and optimised training techniques to rival sheer computing power. This efficiency-driven approach could inspire a new wave of cost-effective AI development, potentially improving return on investment in the field.
While DeepSeek-R1 is more efficient, running large language models locally still requires significant computational resources. The model can be deployed on local or cloud GPU infrastructure, but this doesn't necessarily mean it can run on personal devices like smartphones or laptops
DeepSeek's breakthrough is indeed shaking up the AI industry, challenging established players, and potentially redefining the competitive landscape in AI development, its full impact on local AI deployment remains to be seen.
The next generation of AI-powered business transformation is here, and it's more accessible than ever. Get started today.
