#203: DeepSeek's Disruption: Turning AI into a Commodity

After dissecting DeepSeek’s “Sputnik Shock” in Newsletter #202, it’s time to explore how they’re fundamentally reshaping generative AI economics. While our earlier analysis focused on their technological leaps, the spotlight now shifts to the broader implications of their accomplishments. By reducing inference costs to an astonishing 1/30th of competitors, DeepSeek isn’t merely innovating—they’re redefining market expectations.

This sudden cost parity signals a paradigm shift: AI is no longer a premium resource for the elite few. It’s inching closer to a ubiquitous commodity, courtesy of DeepSeek’s architectural efficiencies and data curation strategies. But beneath the headlines lie questions about the trade-offs, unspoken alliances, and potential unknowns fueling this disruption.

The Fine-Tuning Microdose: When Less Became More

DeepSeek’s approach to Supervised Fine-Tuning (SFT) began as an exercise in restraint. Initially, they steered away from large-scale SFT to avoid the burden of creating meticulously annotated datasets. Yet real-world usage quickly showed that some form of SFT was indispensable to produce coherent, context-aware responses.

Their turning point came with DeepSeek-R1, which introduced a “cold-start” technique specifically targeting reasoning. By beginning with carefully selected Chain-of-Thought (CoT) examples, they achieved the right “microdose” of SFT—enough to instill both coherence and deeper thinking. This struck a balance between the raw efficiency of earlier versions (like DeepSeek-V3) and the nuanced reasoning users expect. The result? A model that not only responds faster but also thinks more carefully.

RLHF: Same Mess, Better Tools

Reinforcement Learning from Human Feedback (RLHF) often promises to tame AI with human values. DeepSeek refined this process through multi-stage reinforcement, rejection sampling, and other optimizations—ensuring the model had a solid logical framework before humans stepped in to fine-tune it further.

Yet the crux remains: humans are subjective beings. DeepSeek’s RLHF may be more polished, but it still mirrors the biases and beliefs of the people providing feedback. Sensitive topics like Tiananmen Square remain filtered, underscoring how even the most advanced RLHF pipelines ultimately reflect human constraints. For all its technical elegance, this system is still shaped by the same messy inputs as its forerunners.

The Curious Case of DeepSeek’s Cheap Inference

DeepSeek’s most headline-grabbing feat is their ability to slash inference costs by a factor of 30 compared to industry titans. Officially, they credit Mixture-of-Experts (MoE) for activating only 37 billion out of 671 billion parameters per token prediction, alongside rigorous data filtering that cuts out redundancy. These techniques undeniably boost efficiency.

Still, some suspect there’s more at play. Sparse architectures require specialized hardware for optimal performance—could DeepSeek have access to custom chips or infrastructure that most rivals lack? Others theorize that their open-source strategy spreads development costs across a broader community. Rumors even hint at undisclosed partnerships or subsidies. Whatever the truth, DeepSeek’s achievement stands as a compelling mix of engineering prowess and possible behind-the-scenes advantages.

Conclusion: A Disruption with Lingering Questions

DeepSeek’s sweeping cost reductions are more than a mere technical win—they’ve reset the economic calculus of generative AI. For now, their innovations in MoE, data curation, and possibly specialized hardware offer a clear path to lower prices and faster outputs. Yet questions remain: How sustainable is this model? Are there hidden alignments or compromises behind their public claims?

As the industry chases DeepSeek’s lead, the challenge isn’t just catching up on speed or affordability—it’s grappling with the nuanced trade-offs that come with commoditizing AI. DeepSeek may have shown us what’s possible, but the final chapter of this disruption is still being written.