--- title: 'I Tested DeepSeek vs ChatGPT: The Results Shocked Me' source: 'https://youtube.com/watch?v=P6kew9TrZYM' video_id: 'P6kew9TrZYM' date: 2026-06-15 duration_sec: 394 --- # I Tested DeepSeek vs ChatGPT: The Results Shocked Me > Source: [I Tested DeepSeek vs ChatGPT: The Results Shocked Me](https://youtube.com/watch?v=P6kew9TrZYM) ## Summary A tiny Chinese startup, DeepSeek, has built a model matching GPT-4's performance at one-tenth the training cost and offers it for free. This video tests DeepSeek V3 against GPT-4o to see if free AI can be better. ### Key Points - **DeepSeek's Mixture of Experts Architecture** [00:22] — DeepSeek uses a Mixture of Experts (MoE) architecture with 64 specialized experts, activating only 2 per query, saving computational power. - **Cost Efficiency Threatens Big Tech** [01:42] — DeepSeek's efficiency undercuts OpenAI's $20/month subscription model, potentially making paid AI obsolete. - **Coding Test: Snake Game** [02:45] — DeepSeek generated a working Python snake game 20% faster than GPT-4o, with zero errors on first try, while GPT-4o had a syntax error. - **Reasoning Test: Drying Shirts** [04:20] — DeepSeek correctly answered that drying 10 shirts takes 4 hours (simultaneous event), showing contextual understanding. - **DeepSeek Matches or Beats GPT-4o in Reasoning** [05:22] — In 5 out of 7 logic tests, DeepSeek matched or beat GPT-4o, demonstrating sophisticated reasoning. - **Competition Benefits All Users** [05:51] — DeepSeek's free model forces OpenAI and Google to lower prices or improve, benefiting consumers. ### Conclusion DeepSeek proves that high-quality AI can be free and open-source, challenging the paid AI model and benefiting all users through competition. ## Transcript A tiny Chinese startup just humiliated OpenAI. They built a model that matches GPT4's performance. But here is the crazy part. They did it for onetenth of the training cost and they are giving it away for free. Is the era of paid AI over? Today we put Deepseek V3 to the test to see if free finally means better. Before we dive into the tests, [music] it's crucial to understand what makes Deepseek so different. In a world flooded with AI models that are essentially variations on a theme, DeepSeek [music] broke the mold. This isn't just another copycat trying to chase GPT4's [music] tail. They fundamentally rethought the architecture from the ground up, leading to a breakthrough in [music] efficiency and power. The secret source, it's a concept that's been around for a while, but has only recently been perfected. They're using a sophisticated architecture called a mixture [music] of experts ore. To understand why this is a gamecher, imagine a traditional model like GPT4 as one giant monolithic brain. It's incredibly powerful, but it's also incredibly [music] expensive. Every time you ask it a question, no matter how simple, [music] the entire massive brain has to power up and process it. Deepseek, on the other hand, is like a team of 64 highly specialized experts. When you ask a coding question, [music] a smart routting network instantly identifies the two best experts for the job, say the Python Pro and the [music] Algorithm Ace, and only wakes them up. The other 62 experts remain dormant, [music] saving an immense amount of computational power. This makes it not just faster, but radically [music] cheaper to train and run. We're talking about achieving top tier performance while using only a fraction of the computational resources. >> [music] >> This efficiency is what should be terrifying for big tech. Think about it. If [music] Deep Seek can offer this level of intelligence for mere pennies on the dollar through an API, [music] why would anyone continue to pay a premium? Open AAI's entire $20 subscription model, which subsidizes [music] the immense cost of their giant brain, is suddenly on very shaky ground. This isn't just a new competitor. It's a potential extinction level event for the old way of doing AI. Talk is [music] cheap. Let's put these models to the test with a realworld coding challenge. On one [music] side, we have the reigning champion GPT40. On the other, the Challenger [music] Deep Seek Coder V2. This isn't just any model. It's an open-source [music] mixture of experts model trained on a colossal two trillion tokens of code and natural language. It boasts [music] top scores on benchmarks like human eval claiming to rival proprietary models at [music] a fraction of the cost. But benchmarks are one thing. Practical application is another. So [music] I asked both to write a Python script for a classic snake game using the Pi game library. To make it interesting, I added a twist. The snake must speed up every [music] time it eats an apple. This tests not just basic code generation, but [music] also state management and logical implementation. Right away, Deepseek's performance was [music] impressive. It generated the complete functional code about 20% faster than [music] GPT40. For developers, that speed translates directly [music] to productivity, enabling faster iteration and problem solving. But speed is meaningless if the code is broken. So the real question is, [music] does it actually work? And the answer is a resounding yes. Deepseek's [music] code ran perfectly on the very first try. Zero errors, zero debugging. GPT40, however, stumbled. It missed a crucial [music] variable definition, throwing a syntax error that broke the program. I had to go back and prompt it a [music] second time to get a working fix. This initial test highlights a key difference reliability. While both models eventually produce the correct [music] code, Deepseek delivered a flawless solution faster and on the first attempt. [music] For any developer on a deadline, that's a gamecher in this round. [music] Deepseek isn't just a contender. It's looking like the new heavyweight champion of coding. Now for the second test, reasoning and logic. This is where many models, especially earlier open-source ones, fall flat. They can perform complex calculations, but often miss the simple realworld context [music] that humans grasp instantly. This is a crucial hurdle for AI to overcome if it's going to be genuinely useful. So, I set a classic logic trap to see if Deep [music] Seek could think, not just calculate. I asked, "If I dry five shirts in the sun and it takes 4 [music] hours, how long does it take to dry 10 shirts?" The trap is obvious. A purely mathematical brain might double the time to 8 hours. It's a simple question, but it's a fantastic test for [music] contextual understanding. Deep Seek answers 4 hours. It immediately understands that drying is [music] a simultaneous event. The shirts all dry together, so adding more shirts doesn't extend the time, assuming you have enough space. [music] This ability to handle nuance and implicit assumptions is incredibly impressive. It's a sign of sophisticated training on diverse, highquality data. This isn't just a one-off trick. [music] This reasoning power extends across the board. In my tests, it excelled at debugging code, planning multi-step projects, and even breaking down complex scientific concepts. [music] It feels far less robotic than other open-source models. It's not just regurgitating [music] data. It's connecting dots and demonstrating genuine problem-solving [music] skills. In fact, in five out of the seven logic and reasoning tests I ran, it either matched or outright beat the current [music] industry leader, GPT 40. That is a monumental achievement for a model that's completely open- source. And remember [music] the best part, you aren't paying a single cent for this level of intelligence. Why does this matter if you aren't a [music] developer? Because of competition. For the last 2 years, we accepted that smart AI costs $20 [music] a month. Deep Seek just proved that intelligence is becoming a commodity like electricity. It's getting cheaper every day. This forces Open AAI and Google to either lower their prices or release something significantly better. Either way, we win. One caveat. This is a Chinese model. If you are working on top secret [music] government data, maybe stick to local models. But for learning, coding, and general tasks, [music] it's a no-brainer. If you want to run AI completely privately on your own computer, check out this tutorial next. The revolution is open source.