Skip to main content

Nvidia Dismisses China AI Threat, Says DeepSeek Still Needs Its Chips

3 months 2 weeks ago
Nvidia has responded to the market panic over Chinese AI group DeepSeek, arguing that the startup's breakthrough still requires "significant numbers of NVIDIA GPUs" for its operation. The US chipmaker, which saw more than $600 billion wiped from its market value on Monday, characterized DeepSeek's advancement as "excellent" but asserted that the technology remains dependent on its hardware. "DeepSeek's work illustrates how new models can be created using [test time scaling], leveraging widely-available models and compute that is fully export control compliant," Nvidia said in a statement Monday. However, it stressed that "inference requires significant numbers of NVIDIA GPUs and high-performance networking." The statement came after DeepSeek's release of an AI model that reportedly achieves performance comparable to those from US tech giants while using fewer chips, sparking the biggest one-day drop in Nvidia's history and sending shockwaves through global tech stocks. Nvidia sought to frame DeepSeek's breakthrough within existing technical frameworks, citing it as "a perfect example of Test Time Scaling" and noting that traditional scaling approaches in AI development - pre-training and post-training - "continue" alongside this new method. The company's attempt to calm market fears follows warnings from analysts about potential threats to US dominance in AI technology. Goldman Sachs earlier warned of possible "spillover effects" from any setbacks in the tech sector to the broader market. The shares stabilized somewhat in afternoon trading but remained on track for their worst session since March 2020, when pandemic fears roiled markets.

Read more of this story at Slashdot.

msmash

DeepSeek Piles Pressure on AI Rivals With New Image Model Release

3 months 2 weeks ago
Chinese AI startup DeepSeek has launched Janus Pro, a new family of open-source multimodal models that it claims outperforms OpenAI's DALL-E 3 and Stable Diffusion's offering on key benchmarks. The models, ranging from 1 billion to 7 billion parameters, are available on Hugging Face under an MIT license for commercial use. The largest model, Janus Pro 7B, surpasses DALL-E 3 and other image generators on GenEval and DPG-Bench tests, despite being limited to 384 x 384 pixel images.

Read more of this story at Slashdot.

msmash