DeepSeek has returned to the AI scene with a significant new research paper introducing MHC (Manifold Constrained Hyperconnections), which addresses a longstanding challenge in scaling AI models. Traditionally, making AI models smarter by adding complex connections often led to instability during training, causing models to “blow up” or collapse. Standard architectures like ResNets and Transformers use residual connections to maintain stability by combining new and old information at each layer. However, DeepSeek’s innovation, hyperconnections, aimed to allow multiple streams of memory to interact, potentially increasing intelligence and reasoning within the same computational budget. The problem was that these hyperconnections caused instability at scale, especially in large models with over 10 billion parameters.
The breakthrough with MHC is that it solves the instability issue that plagued hyperconnections. Hyperconnections allowed for free mixing of information between streams, but without constraints, this led to uncontrolled amplification or dampening of signals, breaking the training process. MHC introduces strict rules: all values in the mixing matrix must be positive, each row and column must sum to one, ensuring no signal amplification or loss. This enforces conservation of “energy” in the network, stabilizing both the forward pass and gradients during deep training. As a result, MHC enables the benefits of richer internal memory and cross-layer reasoning without sacrificing the stability that residual connections provide.
DeepSeek’s focus is now on three main research directions: mathematics and code (as testbeds for AGI), multimodality (enabling AI to interact with the real world), and natural language (fundamental to human intelligence). The company’s mission is to unravel the mystery of AGI with curiosity, and their CEO has stated that achieving AGI could take anywhere from two to ten years, but is expected within our lifetime. The next major model, DeepSeek R2, has faced delays due to dissatisfaction with performance and challenges in training on Huawei Ascend chips, a result of US export restrictions on Nvidia hardware. The new release is now expected in early 2026.
However, DeepSeek’s rapid rise has triggered significant security and censorship concerns. The US House Select Committee on China and OpenAI have accused DeepSeek of unauthorized model distillation, and security researchers have found code capable of transmitting user data to Chinese state-controlled telecoms. Cisco’s testing revealed DeepSeek’s model failed to block harmful prompts effectively, and all user data is stored in China under laws requiring cooperation with intelligence agencies. The model also collects extensive user data, and censorship is built-in, with the AI refusing to discuss sensitive topics like Tiananmen Square and aligning with Chinese government narratives. These issues have led to bans in multiple countries and regulatory probes in Europe.
Despite these controversies, DeepSeek has established itself as a disruptive force in the AI landscape. Its models are significantly cheaper than OpenAI’s and have matched or exceeded some benchmarks, though recent updates from competitors have widened the gap. DeepSeek’s approach challenges the prevailing assumptions about AI development, emphasizing efficiency over brute-force scaling. The company’s open research and cost-effective models have pressured competitors and demonstrated that the AI race is more open than previously thought. However, questions remain about the true costs of their achievements, their ability to sustain innovation under export controls, and whether users can trust a model that enforces state censorship and fails basic security standards.
