An Obligatory Post on DeepSeek

In all seriousness, I've been getting bombarded with questions about DeepSeek since November when R1-lite-preview was released, followed by V3. But R1's release 7 days ago really lit the world ablaze with its exceptional performance and lower cost point. The rapid improvement we've seen from DeepSeek is nothing short of astounding.

Here's why I think this matters for the industry.

Open Source Challenging the Status Quo

First off, I love seeing open models challenge proprietary ones. DeepSeek is another significant addition to the open Llama-based landscape, and what they've achieved is remarkable. The really interesting part? Their reasoning capabilities are impressive even in the smaller models that can run on laptops. This isn't just about having another model – it's about democratizing access to advanced AI capabilities.

Economics and Innovation

DeepSeek is challenging the economics of training and scaling models across multiple axes, which is fantastic for everyone. They're pushing the industry forward in ways that go beyond just raw performance metrics. This ties into a broader trend we're seeing: Mistral, Llama 3, and others are proving that open models can be comparable and competitive to large-scale foundational models. Both distilled models (where larger models' knowledge is compressed into smaller ones) and Small Language Models (SLMs) that are designed to be efficient from the ground up are especially interesting in this context.

Speaking of smaller models, we're seeing exciting developments across the board. Hugging Face recently released SmolerVLM, a remarkably efficient vision-language model that packs impressive capabilities into just 1.3B parameters. This thing can run on a single consumer GPU with just 12GB of VRAM - we're talking laptop territory here. What's really interesting is that despite its tiny size (70x smaller than GPT-4V), it actually outperforms much larger models on certain vision-language tasks.

In parallel, Microsoft just made waves by fully open-sourcing their Phi-4 model on Hugging Face. This isn't just another model release - Phi-4 is significant because it delivers performance comparable to much larger models while being remarkably compact at just 4 billion parameters. For context, it achieves 84.4% on GSM8K and matches models 10-100x its size on various benchmarks. The fact that Microsoft is making this fully open source, including weights and architecture details, represents a major contribution to democratizing AI. It also demonstrates Microsoft's diversified approach to AI - they're clearly not putting all their eggs in the OpenAI basket despite their significant partnership. This trend of making smaller models not just more capable but actually competitive with their larger cousins represents a significant shift for accessibility and practical applications.

The China Factor

DeepSeek definitely shows that China is competitive on the innovation front, not just playing catch-up. However, this brings up some important considerations:

Censorship Concerns

There's a legitimate concern about censorship being baked into training data and open models. This creates a potential proliferation effect and opens a new chapter for propaganda potential. While open source helps by allowing scrutiny of the model architecture, we're all still blind to the training data fed into these systems. This isn't unique to Chinese models, but it adds an interesting dimension given the current geo-economic-political landscape.

Security Considerations

The Terms and Conditions of DeepSeek's platform services (API, consumer chat app, and mobile applications) need careful scrutiny by US users. As a Chinese company, they're legally obligated to support military intelligence gathering. However, it's important to note that the open source models can be deployed locally or operated through Western inference services like Groq (and likely AWS Bedrock soon). This makes the security question more nuanced than a simple yes/no.

Impact on NVIDIA

The market's kneejerk reaction to NVIDIA seems overblown. Yes, DeepSeek did impressive work with fewer resources, but this doesn't spell doom for GPU demand. If anything, it might increase it as more organizations attempt similar feats with smaller clusters. Consider:

• While NVIDIA faces emerging competition from dedicated chips like Google's TPUs and Amazon's Trainium clusters, these same hyperscalers continue to pledge billions to GPU infrastructure

• Major AI companies are doubling down on GPU investments (see Project Stargate)

• The demand for compute isn't slowing down - NVIDIA still can't make chips fast enough

• It's worth noting that DeepSeek's advances, impressive as they are, still stand on the shoulders of Llama's models, which were trained on one of the world's largest NVIDIA clusters at Meta

The Bigger Picture

DeepSeek, along with developments like Google's Gemini 2.0 Flash (which they've made available to free users), is challenging OpenAI's subscription model and conventional wisdom about what's possible. Google's move to offer their latest reasoning-focused model for free puts additional pressure on OpenAI's premium strategy. They've ignited a new race along an orthogonal axis, pushing innovation in unexpected directions.

The space is truly remarkable to be part of right now. We're seeing rapid advancement not just in raw capabilities, but in efficiency, accessibility, and novel approaches to solving complex problems. The next few years are going to be fascinating as these various threads of open source, smaller models, novel training approaches, and geopolitical considerations continue to interweave and evolve.

Time will tell how this all plays out, but one thing's certain: DeepSeek has helped push the boundaries of what we thought was possible, and that's good for everyone in the field.