Deep Thoughts on DeepSeek

AI Spotlight
The “AI Sputnik” Moment?

TL;DR: AI companies are in a…umm…appendage measuring contest after a Chinese start-up embarrassed OpenAI and upended the AI world. The market reacted very poorly. People are worried that China is stealing all of our data…again. Oh, and it’s a huge leap forward for AI.

If you haven’t lived under a rock this week, you have likely heard of DeepSeek. The company recently released its DeepSeek-R1 reasoning model, with benchmarks showing it outperforms the major US tech giants' models, including OpenAI. The US then had it's AI Sputnik moment when they woke up to the fact that they aren't the only world player making strides in AI innovation.

The first shockwave was that a non-US tech company caught up to Silicon Valley’s golden child, OpenAI. The second shockwave was that DeepSeeks’ latest model only cost $5.6 million to create, largely attributed to a few innovative technical approaches that allowed it to train much more effectively.

Compare that to OpenAI’s alleged $100M it spent to build GPT4, and well…you have the panic sell-off in the market that cost NVIDIA $600 billion in market value, the largest single-day capital loss for a US company. Even after a rough trading day, NVIDIA still said DeepSeek was “an excellent AI advancement.”

Of course, there’s drama. First, DeepSeek reported “large-scale malicious attacks” that caused them to limit new user registration. Maybe it was a DDoS attack? Or maybe it was just the flood of users trying to register? This same thing happened to OpenAI late last year when they released Sora to the world.

Second, OpenAI suspects that DeepSeek violated its terms of service and used its API for model distillation, which transfers knowledge from a large model (OpenAI) to a small model (DeepSeek). This is equivalent to OpenAI pointing at DeepSeek and saying, “Hey, you cheated!” As many have pointed out, the irony is not lost that OpenAI also used a crap ton of data without its owners’ consent.

But what about the quality of responses? People flocked to the mobile app, making it the top downloaded app in the US (sorry, ChatGPT). From the reviews I’ve seen, people agree it’s on par with ChatGPT’s responses…but with a catch.

We know that China likes to control the information its citizens see. Naturally, people started testing DeepSeek’s chatbot to see how it handled certain sensitive topics, like Tiananmen Square, Taiwan, and whether a hot dog is a sandwich…err, scratch that last one.

This video shows the output difference for the same questions across ChatGPT, Gemini, and DeepSeek. No surprise here, the answers are vastly different.

Then there’s the security and privacy outcry. They are the same as the TikTok arguments — you’re giving your data to China! And yes, you are…just like you give OpenAI your data when you use ChatGPT.

Are we still surprised that tech companies collect and store troves of information on us in the cloud?

Of course, the added media attention attracts the attention of security researchers. These researchers are good at finding things, like Wiz, who found a publicly exposed DeepSeek database with zero authentication. The database contained “a significant volume of chat history, backend data and sensitive information, including log streams, API Secrets, and operational details.” To Wiz’s credit, they responsibly disclosed the issue, allowing DeepSeek to fix it before publishing their findings. Nothing like a little ambulance chasing for marketing purposes…

What about safety and ethics? OpenAI and other AI models put a lot of effort into implementing guardrails to prevent their chatbots from returning harmful content—obvious things like not giving instructions on making explosives or not encouraging harmful acts. They also put effort into avoiding jailbreaks, where users give specific prompts to bypass those guardrails.

The threat intelligence company Kela investigated the ability to jailbreak DeepSeek’s models to retrieve harmful content. Spoiler alert: It was easy. Many old jailbreak techniques easily bypass DeepSeek’s guardrails. This is unsurprising, given that DeepSeek is only a few years old and has only a fraction of OpenAI's budget.

So, what does this all mean for us? My take is that it’s a net positive for everyone. We shouldn’t live in a world where innovation and competition exist in one country’s vacuum. Not to mention that the US was hitting diminishing returns on training newer models. Competition fuels more innovation. It seems that Sam Altman, CEO of OpenAI, agrees.

Of course, the security implications are that if you use DeepSeek’s chatbot, your data will go to China, just as your data goes to Google, OpenAI, and Meta when you use their products. We shouldn’t be shocked by this — it’s just a fact with today’s tech products.

Am I excited that this will fuel innovation in AI to make things even more efficient? Hell, yes! The more efficient we can make things, the more viable it is for startups to push the boundaries and drive faster (and hopefully safer) progress towards an AI-assisted world.

Am I going to switch to DeepSeek? No.

This is just one new model in a sea of many other models that offer similar performance. We’re still so early in the AI game that there shouldn’t be any watershed moment that permanently shifts to one player or the next. We’re in a time of rapid innovation, and for the consumer, that means rapid testing to see what you like and don’t like.

More competition → more innovation → more consumer upside → more market upside.

That sounds like winning to me.

Security & AI News
What Else is Happening?

🫢 A pro-Palestinian hacking group compromised an Israeli electronics firm that operates emergency alert systems in schools, including kindergartens. They used their access to broadcast rocket alerts before wiping the company’s systems.

🚪 It took just a few weeks from when a few vulnerabilities went public for SimpleHelp RMM to security firms seeing malicious activity. While it’s unclear if the two are related, the timing is suspect. Remember, anything that provides remote access to your environment is of interest to hackers.

🔁 Do you ever feel like you’re just watching Scooby Doo episodes on repeat? Same story, different victim? Poland accused Russia of attempting to recruit Polish citizens to help spread disinformation ahead of Poland’s presidential election in May.

🤦 A new report from the Electronic Privacy Information Center (EPIC) found that “of the 19 states that have passed comprehensive consumer privacy legislation, nearly half received failing grades, and none received an A.” Looks like we still have a lot of work to do there.

🚨 The FBI seized several domains tied to popular hacking forums. It’s impressive that law enforcement has stepped up the pressure to disrupt cybercriminals’ operations. While these efforts aren’t long-term solutions (because cybercriminals will create new infrastructure), the added pressure is a nice touch to raise operational costs for the bad guys.