In This Story
Despite rattling global tech stocks with competitive artificial intelligence models, Chinese AI startup DeepSeek didn’t introduce anything new, a Google (GOOGL+0.65%) executive said.
Google DeepMind chief executive Demis Hassabis said the Hangzhou-based startup’s AI model “is probably the best work” from China, and is “an impressive piece of work,” during a Google event in Paris, CNBC reported. Hassabis said DeepSeek has demonstrated “extremely good engineering,” and that its AI models have deeper geopolitical implications.
However, Hassabis said DeepSeek doesn’t show “actual new scientific advance” and is “using known techniques” in the AI industry. He added that the reaction to DeepSeek was “exaggerated a little bit.”
Last month, DeepSeek released results for its latest open-source reasoning models, DeepSeek-R1, which performed comparably to OpenAI’s reasoning models, o1-mini and o1, on several industry benchmarks. In December, the startup launched its DeepSeek-V3 models which it said cost just $5.6 million to train and develop on Nvidia’s (NVDA+3.46%) H800 chips — the reduced-capability version of Nvidia’s H100 chips used by U.S. firms.
The release of DeepSeek-R1 sparked a global sell-off of tech stocks, sending Nasdaq, Dow Jones Industrial Average, and S&P500 futures falling. Nvidia’s stock plunged 17%, wiping out nearly $600 billion in value — a record loss for a U.S. company. DeepSeek’s cheaper-yet-competitive models have raised questions over Big Tech’s big spending on AI infrastructure, as well as how effective U.S. chip export controls are.
Meanwhile, Google made its Gemini 2.0 Flash Thinking Experimental AI model available to all Gemini app users last week. The AI model currently ranks as the best model in the world on the community-driven Chatbot Arena.
Gemini 2.0 Flash Thinking Experimental is trained to “strengthen its reasoning capabilities” by breaking down prompts step-by-step and showing users its “thought process” to understand how it came to its response. Through this process, users can see “what its assumptions were, and trace the model’s line of reasoning,” Google said.