Deep Seek - Just how censored is it?

Deep Seek R1: Controversy, Innovation, and AI Export Controls

Recent developments surrounding Deep Seek's R1 model have sparked significant discussion in the AI industry, particularly regarding potential data distillation from OpenAI models and the implications for GPU export controls. Former OpenAI employee and current Anthropic CEO Dario Amade has shared insights on these developments through a new essay.

Data Distillation Controversy

Evidence has emerged suggesting Deep Seek may have used OpenAI's model outputs in training R1:

- The R1 model occasionally self-identifies as being trained by OpenAI

- Industry leaders, including Grock CEO Jonathan Ross, suggest Deep Seek spent significant resources distilling knowledge from OpenAI's models

- The reported $5 million training cost may not reflect total R&D investment

Scaling Laws and AI Development

Amade outlines three key dynamics in AI development:

1. Scaling Laws

- Larger models consistently show linear improvements in capabilities

- Each order of magnitude increase in scale yields measurable performance gains

- Cost increases exponentially while quality improves linearly

2. Shifting Curves

- Software and hardware improvements provide efficiency gains

- Companies typically reinvest efficiency gains into training more advanced models

- Current efficiency improvements occur roughly 4x per year

3. Paradigm Shifts

- New training approaches can fundamentally change model capabilities

- Recent focus on reinforcement learning for chain-of-thought reasoning

- Still early in exploring this new paradigm's potential

Deep Seek's Innovation

According to Amade:

- Deep Seek V3, released in December, represented the true technical breakthrough

- R1 built upon V3 by adding reinforcement learning capabilities

- Deep Seek's achievements align with expected industry progress

- Their models perform similarly to US models from 7-10 months ago

Future Implications

Amade presents two possible futures:

1. Bipolar World (2026-2027)

- Both US and China achieve advanced AI capabilities

- China could potentially leverage advantages in military applications

- Requires successful circumvention of export controls

2. Unipolar World

- US and allies maintain AI leadership through export controls

- Harder to hide large-scale chip acquisitions

- Export controls remain crucial for maintaining technological advantage

Market Impact

The market reaction to Deep Seek's announcements, including NVIDIA's 16% stock drop, appears overblown. Industry experts suggest these developments represent expected progress rather than a fundamental disruption of AI economics.

Conclusions:

- Deep Seek's achievements, while impressive, align with industry trends

- Export controls remain critical for AI development globally

- The race toward artificial general intelligence continues to accelerate

- Current projections suggest human-level AI capabilities by 2026-2027

This development highlights the ongoing competition in AI advancement and the critical role of policy in shaping the future of artificial intelligence.

Note: The YouTube video based on this article goes deeper into how much censorship there is in Deep Seek.

Scan the QR code in the picture to watch the video.

Link: Deep Seek - Just how censored is it?