DeepSeek AI – Understanding what the hype actually means

In 1957, a small metal sphere sent shockwaves through America. The Soviet Union's Sputnik satellite wasn't just a scientific achievement – it was a wake-up call that showed the world that innovation could come from unexpected places. Today, we're witnessing AI's own Sputnik moment with DeepSeek-R1 (following what Marc Andreessen wrote on Twitter:)

To be honest, it does feel like it. DeepSeek-R1 has emerged from relative obscurity to knock the entire AI industry off its limbs, even causing the US markets to crash, with Nvidia losing 20% of its stocks amounting to an insane $600 billion! Overall, US stocks lost more than $1 trillion in market cap on Monday because of DeepSeek.

David vs. Goliath: The Numbers That Matter

Think of it this way: until now, building cutting-edge AI has been like constructing the world's tallest skyscraper – you needed enormous resources, countless workers, and mind-boggling amounts of money. Silicon Valley giants like OpenAI and Google have been pouring billions into building these AI "skyscrapers."

DeepSeek-R1, did something remarkable: it matched the performance of leading American AI models while spending way less money.

Imagine spending $6 million to build something that performs better than products that cost hundreds of millions. That's exactly what DeepSeek did. Their latest AI model, V3, is going toe-to-toe with (and often beating) the biggest names in AI:

• It's better at math than almost every other AI, scoring 90.2% on advanced problems while Claude-3.5 Sonnet scores 78.3%

• It's crushing coding challenges, scoring more than twice as high as other leading AIs

• In general knowledge tests, it's neck-and-neck with the best, scoring 88.5% compared to industry leader Llama 3.1's 88.6%

• It’s open source.

And here's the kicker: they did all this using just 2,048 computer chips for two months. To put that in perspective, it's like building a skyscraper with a small construction crew while your competitors are using entire armies of workers.

Enter Liang Wenfeng

The story gets even more interesting when you look at DeepSeek's founder, Liang Wenfeng. At 40 years old, he's not your typical tech startup founder. Before DeepSeek, he built an $8 billion hedge fund called High-Flyer. Then, almost as a side project, he decided to venture into AI. His philosophy? Keep prices just slightly above costs - he's not in it to make "excessive profit."

This approach is already visible in their pricing: they're charging $1.10 per million tokens for output - 25x cheaper than OpenAI.

But how do you spend so less yet come out on top?

DeepSeek's secret sauce lies in their clever engineering. Instead of just throwing more computing power at the problem (like most big tech companies do), they're being smarter about how they use their resources:

1. They use something called a "Mixture-of-Experts" approach - imagine having a team of specialists where only the relevant experts weigh in on each question, rather than asking everyone every time

2. They've found ways to make their AI more efficient - it has 671 billion parameters (think of these as knowledge points), but only uses 37 billion at a time

3. They've trained their AI on an enormous amount of information - 14.8 trillion pieces of data - but did it more efficiently than their competitors

But Is It Really That Cheap?

Here's where things get controversial. Some experts are skeptical about DeepSeek's claims of doing more with less. Scale AI's CEO suggests they might be secretly using about 50,000 high-end Nvidia chips - the kind that Chinese companies aren't supposed to have access to because of U.S. export controls. Even Elon Musk jumped into the debate, calling this theory "obviously" true.

But renowned AI expert Andrej Karpathy points out that DeepSeek's V3 used about 11 times less computing power than similar models while achieving better results. This suggests they might really have found a more efficient way to build AI.

But what’s the catch?

Personally, their shady Terms of Service.

Using DeepSeek's AI comes with some important considerations:

1. Data Privacy: Their terms of service let them keep and analyze everything users input. Given that they operate under Chinese regulations, this raises questions about data privacy and government access.

   

   Data to be stored in PRC. Sus.

2. Content Restrictions: The AI won't discuss certain sensitive topics due to Chinese regulations. This built-in censorship could spread if other developers build on DeepSeek's technology.

3. User Responsibility: If something goes wrong with the AI's output, it's mostly on you. They're clear about this: users need to verify everything the AI produces.

   

   Like, why would you be comfortable with DeepSeek collecting my keystroke patterns?

The Open-Source Game Changer

Here's what makes DeepSeek's move especially dramatic: they're giving away their code for free. It's like Tesla suddenly publishing the complete blueprints for their best car and saying "Here, build your own if you want to!" (Elon would never do that, LOL)

This "open-source" approach means that:

• Other developers can look under the hood to see how everything works

• Anyone can modify and improve the technology

• Companies can build their own applications using DeepSeek's foundation

• The community can collaborate to make it even better

It's particularly ironic because OpenAI - despite its name suggesting openness - has become "by far, the most closed in every way possible," according to AI developer Reuven Cohen. Meanwhile, DeepSeek is doing what OpenAI originally promised: sharing AI advances with everyone.

What This Means for the Future

This feels like a pivotal moment in AI development. DeepSeek has shown that you might not need billions of dollars to compete in advanced AI - clever engineering could be enough.

The impact is already visible:

• Their app hit #1 in app stores across multiple countries, with 1.6 million downloads

• Major tech stocks took a hit, with the Nasdaq having its third-worst day in two years

• The U.S.-China AI race is intensifying, with President Trump backing OpenAI's massive $500 billion infrastructure plan

The Bottom Line

Whether or not all of DeepSeek's claims are entirely accurate, they've proven something important: breakthrough innovation doesn't always need massive resources or come from expected places. They've shown there might be a smarter way to build AI, which could open the door for more companies to enter the field.

Just as Sputnik launched a space race that transformed our world, DeepSeek might have just sparked a new kind of AI race - one where clever engineering matters as much as deep pockets. The future of AI might look very different from what Silicon Valley expected.

DeepSeek made its move. And we are watching.