DeepSeek and the coming AI Cambrian explosion

The excitement about DeepSeek is understandable, but a lot of the reactions I’m seeing feel quite a bit off-base. DeepSeek represents a significant efficiency gain in the large language model (LLM) space, which will have a major impact on the nature and economics of LLM applications. However, it does not signal a fundamental breakthrough in artificial general intelligence (AGI), nor a fundamental shift in the center of gravity of AI innovation. It’s a sudden leap along an expected trajectory rather than a disruptive paradigm shift.

DeepSeek’s impressive achievement mirrors the broader historical pattern of technological progression. In the early 1990s, high-end computer graphics rendering required supercomputers; now, it’s done on smartphones. Face recognition, once an expensive niche application, is now a commodity feature. The same principle applies to large language models (LLMs). The surprise isn’t the nature of the advance, it’s the speed.

For those paying attention to exponential technological growth, this isn’t shocking. The concept of Technological Singularity predicts accelerating change, particularly in areas of automated discovery and invention, like AI. As we approach the Singularity, breakthroughs will seem increasingly rapid. DeepSeek is just one of many moments in this unfolding megatrend.

DeepSeek’s architectural innovations: impressive, but not new

DeepSeek’s main achievement lies in optimizing efficiency rather than redefining AI architecture. Its Mixture of Experts (MoE) model is a novel tweak of a well-established ensemble learning technique that has been used in AI research for years. What DeepSeek did particularly well was refine MoE alongside other efficiency tricks to minimize computational costs:

Parameter efficiency: DeepSeek’s MoE design activates only 37 billion of its 671 billion parameters at a time. This means it requires just 1/18th of the compute power of traditional LLMs.

Reinforcement learning for reasoning: Instead of manual engineering, DeepSeek’s R1 model improves chain-of-thought reasoning via reinforcement learning.

Multi-token training: DeepSeek-V3 can predict multiple pieces of text at once, increasing training efficiency.

These optimizations allow DeepSeek models to be an order of magnitude cheaper than competitors like OpenAI or Anthropic, both for training and inference. This isn’t a trivial feat—it’s a major step toward making high-quality LLMs more accessible. But again, it’s a stellar engineering refinement, not a conceptual leap toward AGI.

The well-known power of open-source

One of DeepSeek’s biggest moves is making its model open-source. This is a stark contrast to the walled-garden strategies of OpenAI, Anthropic and Google – and a nod in the direction of Meta’s Yann LeCun. Open-source AI fosters rapid innovation, broader adoption, and collective improvement. While proprietary models allow firms to capture more direct revenue, DeepSeek’s approach aligns with a more decentralized AI future—one where tools are available to more researchers, companies, and independent developers.

The hedge fund HighFlyer behind DeepSeek knows open-source AI isn’t just about philosophy and doing good for the world; it’s also good business. OpenAI and Anthropic are struggling with balancing research and monetization. DeepSeek’s decision to open-source R1 signals confidence in a different economic model—one based on services, enterprise integration, and scalable hosting. It also gives the global AI community a competitive toolset, reducing the grip of American Big Tech hegemony.

China’s role in the AI race

Some in the West have been taken aback that DeepSeek’s breakthrough came from China. I’m not so surprised. Having spent a decade in China, I’ve witnessed firsthand the scale of investment in AI research, the growing number of PhDs, and the intense focus on making AI both powerful and cost-efficient. This isn’t the first time China has taken a Western innovation and rapidly optimized it for efficiency and scale.

However, rather than viewing this solely as a geopolitical contest, I see it as a step toward a more globally integrated AI landscape. Beneficial AGI is far more likely to emerge from open collaboration than from nationalistic silos. A decentralized, globally distributed AGI development effort—rather than a monopoly by a single country or corporation—gives us a better shot at ensuring AI serves humanity as a whole.

DeepSeek’s broader implications: The future beyond LLMs

The hype around DeepSeek largely centers on its cost efficiency and impact on the LLM market. But now more than ever, we really need to take a step back and consider the bigger picture.

LLMs are not the future of AGI

While transformer-based models can automate economic tasks and integrate into various industries, they lack core AGI capabilities like grounded compositional abstraction and self-directed reasoning.

If AGI emerges within the next decade, it’s unlikely to be purely transformer-based. Alternative architectures—like OpenCog Hyperon and neuromorphic computing—may prove more fundamental to achieving true general intelligence.

The commoditization of LLMs will shift AI investment

DeepSeek’s efficiency gains accelerate the trend of LLMs becoming a commodity. As costs drop, investors may begin looking toward the next frontier of AI innovation.

This could drive funding into AGI architectures beyond transformers, alternative AI hardware (e.g., associative processing units, neuromorphic chips), and decentralized AI networks.

Decentralization will shape AI’s future

The AI landscape is shifting toward decentralized architectures that prioritize privacy, interoperability, and user control. DeepSeek’s efficiency gains make it easier to deploy AI models in decentralized networks, reducing reliance on centralized tech giants.

DeepSeek’s role in the AI Cambrian explosion

DeepSeek represents a major milestone in AI efficiency, but it doesn’t rewrite the fundamental trajectory of AGI development. It’s a sudden acceleration along a predictable curve, not a paradigm shift. Still, its impact on the AI ecosystem is significant:

It pressures incumbents like OpenAI and Anthropic to rethink their business models.

It makes high-quality AI more accessible and affordable.

It signals China’s growing presence in cutting-edge AI development.

It reinforces the inevitability of exponential progress in AI.

Most importantly, DeepSeek’s success should serve as a reminder that AGI development isn’t just about scaling up transformers. If we truly aim to build human-level AGI, we need to go beyond optimizing today’s models and invest in fundamentally new approaches.

The Singularity is coming fast—but if we want it to be beneficial, we must ensure it remains decentralized, global, and open. DeepSeek is not AGI, but it’s an exciting step in the broader dance toward a transformative AI future.

We've featured the best AI phone.

This article was produced as part of TechRadarPro's Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro