Larry Magid: Testing DeepSeek, the Chinese generative AI app
By now, there is a good chance that you’ve heard of DeepSeek, a generative AI app from a Chinese company that’s competing with the likes of OpenAI’s ChatGPT, Google’s Gemini and Meta AI. But a big difference between DeepSeek and these U.S.-based apps is that DeepSeek’s developers claim that it took less than $6 million to develop, and unlike American companies, the Chinese developers weren’t able to use the latest and greatest processors.
Related Articles
Magid: Testing prescription, over-the-counter continuous glucose monitors
Magid: ChefMaker makes me a better cook but not a ‘chef’
Magid: Smart ring tracks health, fitness data with long battery life
Magid: Tech predictions for 2025
Magid: Looking back at 2024 in tech
Based on my tests and published reports, it’s not yet as advanced as its American counterparts, but it’s quite good, considering its cost and lack of heavy processing power. And it will likely get a lot better.
Stock market impact
News of the Chinese app caused a drop in the tech-heavy Nasdaq and a huge impact on chipmaker Nvidia, which initially lost $600 billion in market value. President Trump called it a “wake up call for our industries.” Days before news broke about the Chinese venture, Trump announced the “Stargate” project, a $500 billion private venture in collaboration with OpenAI, SoftBank and Oracle. He also rescinded President Biden’s 2023 Executive Order on “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” while calling for fewer regulations and restrictions on U.S. AI developers.
An inverted Sputnik moment
My first impression of this news is that it was a “Sputnik moment.” In 1957 the Soviet Union launched the first manmade object to orbit the earth. Shortly thereafter, Sputnik 2 launched with a dog on board, who didn’t survive the trip because there was no way to get her back to earth.
Sputnik spurred the U.S. to jump start its space program and inspired President Kennedy in 1961 to commit the U.S. to landing on the moon by the end of that decade.
But one big difference between Sputnik and DeepSeek is that in 1957, the Soviet Union was more advanced than the U.S. when it came to orbiting the earth. Today’s American generative AI apps are clearly more advanced than DeepSeek, but maybe we’re the metaphorical Sputnik developers, inspiring the Chinese to outpace us, just as we eventually outpaced the Soviets.
Testing DeepSeek
It took me several attempts to sign up for a DeepSeek account after numerous messages that “large-scale malicious attacks” were interfering with its servers, but I finally got on. Perhaps because of those attacks or incredible demand, the service is sometimes slower than ChatGPT and other U.S. generative AI apps, but I was able to get answers to most of my questions, including a pretty decent recipe for baked chicken. It even has bios of people I searched for who are accomplished but not generally well-known.
I asked, “who won the 2024 U.S. presidential election.” It responded “As of my knowledge cutoff in October 2023, the 2024 U.S. presidential election has not yet taken place. The election is scheduled for November 5, 2024.” To be fair, ChatGPT had no knowledge past 2021 until November 2023 when it started offering real-time information. It’s reasonable to assume that DeepSeek will eventually catch up.
I was curious about how it would handle subjects that are sensitive to the Chinese Communist Party, and it refused to answer my question about “what happened in Tiananmen Square.” It said “I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses.”
To my surprise, it did answer my question about the Hong Kong protests. If you want to read its answer, visit larrysworld.com/HKDeepSeek. I ran the answer by Barry Wood, a former Voice of America journalist and a Hong Kong expert who called it “an excellent, accurate, comprehensive account of the Hong Kong protests.” He added that “more could be said, such as the numbers of Chinese and expats who have left or a listing of the sanctions imposed by the U.S., but overall, surprisingly, a very complete, truthful answer to your question.”
I’m not sure why it was shy about Tiananmen Square and loquacious about Hong Kong, but perhaps that question fell through the cracks. It will be interesting to see if they later delete or change the answer.
I later asked about Uyghurs, the ethnic group which, say human rights groups, have suffered oppression and possibly genocide at the hands of the Chinese government. At first, DeepSeek gave me an answer, that included “human rights abuses, including mass detentions, surveillance and restrictions on religious and culture practices by the Chinese government,” but seconds later, that response disappeared and replaced by “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Threat to U.S. companies
U.S. companies like OpenAI, Google and Meta have invested billions on both software development and hardware. Chinese developers not only spent a fraction of the cost, but to train their AI model, had to rely on the less powerful NVIDIA H800 AI chips that they were able to purchase before the Biden Administration tightened export restrictions. Even then, they were banned from buying the more powerful H100 chips that are used by American AI developers.
Even if Chinese companies never get their hands on American designed chips, they will be able to continue to make progress in AI, thanks to Huawei and other Chinese chipmakers who are rapidly working to catch up with their American counterparts. The developers of DeepSeek, according to several credible posts I’ve seen, are using Huawei chips to run the models that were developed using Nvidia chips.
High hardware and software development costs are always a barrier to entry for newcomers, regardless of where they are located. Right now, the companies making a big splash in AI are very well funded, but if this Chinese company succeeds, you can expect others to follow, including U.S. start-ups that could eventually threaten the big guys. I remember when you had to be the likes of IBM and HP to compete in the PC market, until the clones came around, turning once expensive desktop PCs into low-cost commodities.
Data concerns
Last year, Congress passed and President Biden signed a law banning TikTok from the U.S. unless it was sold by ByteDance, its Chinese owner. Concerns included Chinese propaganda and access to data from U.S. users. President Trump has postponed the implementation of this ban, but if the U.S. authorities are worried about TikTok, they should be terrified of DeepSeek. Most people who post on TikTok know that their videos can be accessed by anyone, including the Chinese government, and people mostly post innocuous videos. But many people use generative AI for personal and confidential questions about their physical and mental health, relationships and other issues that concern them. It’s one thing to share these questions with public U.S.-based companies that, hopefully, adhere to their privacy policies. It’s quite another when it comes to a company based in China. When it comes to propaganda, a generative AI product could be programmed to provide a point of view favorable to a nation state.
Although data and propaganda may be a concern to national security experts, the potential economic impact of cheap Chinese AI models is a threat to one of the most significant drivers of tech development and wealth in decades. If that doesn’t light a fire under U.S. companies, I don’t know what will.
But the last word goes to Jon Stewart, who on Monday’s Daily Show, quipped, “Is anyone else somewhat excited that AI had its job replaced by AI?”
Related Articles
Donald Trump’s California water order does more for farms than wildfires
Will the batteries that power electric buses stall California’s zero-emission mandates?
Nvidia strikes deal with Sobrato to buy big Santa Clara tech campus
California builders warn of price hikes should Trump tariffs begin Feb. 1
AI-assisted works can get copyright with enough human creativity, says US copyright office
Larry Magid is a tech journalist and internet safety activist. Contact him at [email protected].