I spent a week testing DeepSeek AI — here's what surprised me

Last Tuesday, I stayed up until 2 AM testing a Chinese AI model called DeepSeek. My girlfriend asked why I was yelling "it actually works" at my laptop. I didn't have a good answer.

But here's the thing — I genuinely wasn't expecting much. I've tested dozens of "GPT-killers" that were terrible. Most couldn't handle basic Python debugging. DeepSeek? It refactored a messy function from our codebase in under 30 seconds. I still don't fully understand how.

The $5.6 million question

DeepSeek claims their model cost $5.6 million to train. For context, OpenAI spent an estimated $100 million+ on GPT-4. That's not a typo.

I ran the numbers during a slow afternoon at ParallelTech. If true, this changes everything. My co-founder Oluwa thinks I'm overreacting. "They're probably subsidizing hardware," he said. Maybe. But even if it's double that — $11 million — it's still 90% cheaper.

My take: The era of "bigger models always win" might be ending. Good riddance. Smaller, efficient models mean more startups can compete. That's good for people like us.

What it's actually like to use

The interface is minimal — almost boring. No fancy UI, no "personalities." Just a text box. But the speed surprised me. Responses come back in 1-2 seconds, compared to GPT-4's 3-5 seconds.

I tested it on three real tasks from our week at ParallelTech:

• Debugging: Gave it a broken API endpoint. It found the missing await in 10 seconds. I'd been staring at it for 20 minutes.
• Documentation: Asked it to explain our authentication flow. It was 80% accurate — better than our actual docs (which I need to update).
• Creative writing: Had it write a blog intro. Too formal. But after two tries, it loosened up.

One funny failure: I asked it to roast our logo. It was too polite. AI still struggles with humor. Good for humanity, I guess.

Final verdict

Is DeepSeek better than GPT-4? For coding and reasoning, it's competitive — sometimes better. For creative writing and personality, GPT-4 still wins.

But that's not the point. The point is that AI is getting cheaper — dramatically cheaper. That means more people can build with it. Including small teams like ParallelTech.

P.S. I stayed up until 2 AM again last night testing their coding abilities. Send coffee.