How Did Deepseek Change AI?
The AI world freaked out about DeepSeek this past week. And that made the markets freak out. And basically everyone else freaked out, without totally understanding what they were freaking out about.
Then very quickly everyone tried to un-freakout. Including AI leaders and markets.
So where’s the reality here? Clearly what DeepSeek did is a big deal - but why, exactly? What did they actually *do* that roiled the AI universe?
That’s what we’re breaking down today.
THE STORY SO FAR
This story starts back in 2023, when DeepSeek emerged from a Chinese hedge fund called High-Flyer. Nobody really noticed. But then...
December 2023: Released their first chatbot
May 2024: Dropped DeepSeek-V2, focused on being super cost-efficient
December 2024: Launched V3 with 671B parameters (that's a whole ton of AI brain cells - aka, really smart)
January 2025: Released DeepSeek-R1, and suddenly everyone's paying attention
Why?
Because they just did something that the Big Players hadn’t been able to figure out. And in doing so, they combined powerful and efficient in a way that got heads turning.
THE IMPOSSIBLE MATH
Here's what DeepSeek did: They trained an AI model comparable to GPT-4 for just $5.6 million, instead of, like, a hundred million, as it cost OpenAI. For context, that's like building a Ferrari for the price of a Toyota Corolla.
Now - did they really do it? There are skeptics out there around those numbers. But there’s no denying that they are building and training their models for WAY cheaper.
The part that really turned heads here is that their model performs nearly as well as the big players:
79.8% on advanced math tests
Matches GPT-4 on many benchmarks
Ranked third globally after Google and OpenAI
Uses 95% less computing power
API costs up to 90% less per million tokens
HOW DID THEY PULL THIS OFF?
Remember playing Super Mario as a kid? You didn’t read a manual. You learned by trial and error - jump here, don't touch that, grab these coins.
DeepSeek used the same approach (technically called reinforcement learning) instead of force-feeding their AI massive amounts of labeled data. They basically let their AI figure things out through practice rather than memorization.
Plus, they found creative ways around China's chip restrictions. Sort of like those movies where they’re on a desert island and they turn coconuts into radios.
BUT WAS IT ABOVE BOARD?
There’s a bunch of evidence, of course, that they probably violated a bunch of rules in doing this. In other words, they used ChatGPT to train their own model. Why is there not more stink about this? Because let’s face it, a lot of models train in exactly this way - they steal data, train, and apologize later.
But let’s take a look at what’s been happening in January 2025:
Microsoft's security teams detected some suspicious data transfers
OpenAI launched an investigation into potential data misuse
The U.S. Navy banned personnel from using DeepSeek
Sam Altman announced OpenAI would speed up their releases in response
WHAT THIS MEANS FOR ALL OF US
1. Innovation isn't just for tech giants anymore
Small teams with clever ideas can compete with the big players. DeepSeek proves you don't need unlimited resources to build competitive AI.
2. AI is about to get way cheaper
Think Netflix vs. cable TV prices. When someone figures out how to deliver the same product for 90% less, the whole market changes.
3. Open source is having its moment
DeepSeek made their work open source - meaning anyone can use, modify, and build upon it. It's like they invented a super-efficient engine and shared the blueprints with the world.
The Trade-offs (Keeping It Real)
Nothing's perfect. Here's what you should know:
Content restrictions (it follows Chinese regulations)
Occasional routing errors (sometimes sends you to the wrong "specialist")
Not as smooth with complex conversations
Struggles with certain coding tasks
SHOULD YOU SWITCH?
Perplexity (my favorite AI search engine, by the way) just integrated DeepSeek R1 into their platform. But they're doing it their way:
Running it on US and European servers
No data going to China
No censorship (unlike the Chinese version)
Limited to Pro users ($20/month)
This is huge. It means you can test drive DeepSeek's capabilities through a trusted platform that's already handling your data properly.
For most users though, stick with ChatGPT or Claude for your day-to-day AI needs. The real story isn't about switching - it's about what this means for the future of AI.
Think about it: a major US company just validated DeepSeek's technology while solving the trust issues. That's going to make other companies sit up and take notice.
THE BOTTOM LINE
DeepSeek is proof that the rules of the game are changing. The White House is investigating. Nvidia lost $589 billion in value (though it’s making a big comeback). OpenAI is scrambling to respond.
You don't need unlimited resources or the latest chips to build competitive AI anymore. What comes next? We’ll find out together.
---
*Want to level up your entire organization's AI game? We have a digital course designed to create behavior change at scale. DM me to learn more.*