The DeepSeek Disruption: Implications of Low-Cost LLMs
Last week, DeepSeek disrupted the AI discourse by open-sourcing advances that slash the cost of training and serving large language models (LLMs) like GPT-4 and Claude. Their R1 model rivals top proprietary systems in benchmarks, but at an order of magnitude lower cost. By optimizing compute usage through innovations like GRPO (Grouped Relative Policy Optimization), they've unlocked three seismic shifts:
cheaper tuning and inference enables a flourishing of AI agent capabilities
smaller deployments become possible, from on premise for enterprise down to smartphones, drones, etc.
the future of data centres is debated, efficiency vs. demand. Likely some of the newly committed capex spending is held back.
Quality vs. Cost for leading LLMs
Markets today are opining that this is not good news for chip makers, as on the surface it appears these innovations will reduce demand for GPUs. There’s uncertainty about exactly how much it cost to train R1, it may be more than reported, and some commentators are holding to the thesis that the organization with the most compute will still be able to produce the most capable models. Another line of thinking holds that lower costs will unlock enough latent demand to offset the compute savings, and thus the spending on hardware is justified. It’s an interesting time to be paying attention, and as always this will take time to play out. However what is already clear is that more capability for less cost will ultimately enable more complex use cases that require multiple agents working together, through an agent framework. It’s possible the winners here are those who can build the most useful and trustworthy AI agents.
The Rise of Agent Frameworks
Agent frameworks are the crucial orchestration layer between agents and the services they use, like memory and tools. Adding workflows such as retrieval, self-critique, and context management enables already powerful LLMs to tackle ever more complex, real world tasks. Similar to different brain regions with specific cognitive abilities, or teams with specialized roles, we are seeing LLMs being composed with existing software tools to execute chains of reasoning and action with increasing precision.
The dramatic reduction in compute costs fundamentally transforms how AI agents can be deployed and orchestrated. When each API call costs a fraction of a penny rather than several cents, developers can architect solutions that leverage multiple specialized agents working in concert. This enables AI systems to tackle increasingly complex challenges through extended reasoning chains and collaborative problem-solving approaches that were previously cost-prohibitive.
Consider a complex research task that previously required careful prompt engineering to fit within token limits. With significantly reduced costs, the same task can now be broken down across multiple specialized agents: one conducting initial research, another validating sources, a third synthesizing findings, and a fourth generating the final output. This distributed approach not only improves accuracy but also enables more nuanced handling of complex tasks.
An example of a multi-agent workflow. https://engineering.peerislands.io/future-of-genai-applications-from-rag-to-multi-agent-collaboration-3d43e3871ffb
The Edge Computing Opportunity
The ability to run powerful LLMs on edge devices represents another transformative shift in the AI deployment landscape. By enabling local processing, organizations can now offer sophisticated AI capabilities while addressing privacy concerns and reducing latency. This isn't just about running stripped-down models – DeepSeek's efficiency improvements mean that edge devices can now run models approaching the capabilities of today's cloud-based services.
This shift has profound implications for application architecture and user experience. Healthcare providers can process sensitive patient data locally, financial institutions can run complex analysis without exposing data to external services, and mobile applications can maintain sophisticated AI features even without reliable internet connectivity.
The Data Center Question
The debate between compute efficiency and user demand presents a crucial tension for hyperscalers and their investors. While individual model operations require less compute, the lower barrier to AI access may actually increase overall data center demand. As costs decrease, more organizations can afford to deploy AI at scale, leading to higher aggregate compute usage even as per-operation efficiency improves.
This dynamic is further complicated by the trend toward larger and more sophisticated models. The efficiency gains enabled by innovations like GRPO don't just reduce costs – they may also make it feasible to train and deploy even more powerful models than we see today. It’s possible that as these new ideas are replicated, they are used to extend current scaling trends and the appetite for compute resources may continue to grow.
AI Agents Present the Clearest Opportunity
The impact of DeepSeek's innovations will take time to fully manifest across the AI landscape. While the closed source LLM ecosystem gradually adapts to these efficiency breakthroughs, and edge deployments find their footing in specific use cases, one area stands poised for immediate transformation: agent workflows. The convergence of reduced costs and maintained quality has created the perfect conditions for agent development to flourish.
The frameworks supporting this development are still in their infancy, but they're evolving rapidly. Venture-backed platforms like LangChain and CrewAI stand alongside promising open-source initiatives like Eliza, each offering distinct perspectives on agent architecture and team orchestration. These varying approaches to agent design echo the early days of neural networks, when researchers and developers worked to discover the most effective architectural patterns – patterns that would eventually become the foundation of modern AI.
Dynemetis is Building Agents
Experimentation has begun on a handful of agents, focused on interpreting data in equity options markets. More writing to come on this. If you want to work on building an option trading agent, or you are curious what such an agent could do for your portfolio, please don’t hesitate to reach out.