{"id":4586,"date":"2025-02-12T10:26:08","date_gmt":"2025-02-12T18:26:08","guid":{"rendered":"https:\/\/www.visla.us\/blog\/?p=4586"},"modified":"2025-02-12T10:27:44","modified_gmt":"2025-02-12T18:27:44","slug":"whats-going-on-with-deepseek","status":"publish","type":"post","link":"https:\/\/www.visla.us\/blog\/news\/whats-going-on-with-deepseek\/","title":{"rendered":"What’s Going On With DeepSeek?"},"content":{"rendered":"\n

If you\u2019re into AI, you\u2019ve probably heard whispers\u2014or outright debates\u2014about DeepSeek . It\u2019s the latest large language model (LLM)<\/a> making waves, and its impact has been nothing short of seismic. Whether it\u2019s outperforming expectations on a tight budget, being accused (but not proven) of copying ChatGPT, or raising some serious privacy red flags, DeepSeek\u00a0has everyone in the AI world talking.<\/p>\n\n\n\n

What is DeepSeek?<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

Chinese AI startup DeepSeek recently released two new AI models: V3<\/strong> and R1<\/strong>. Both are powerful, but they serve different purposes. Here\u2019s what you need to know.<\/p>\n\n\n\n

DeepSeek V3: the generalist<\/h4>\n\n\n\n

V3 is a well-rounded, general-purpose model built to handle a wide range of tasks. It was trained on a massive dataset of 14.8 trillion tokens<\/a>, mostly in English and Chinese, with a strong focus on coding and math. One of its biggest strengths is its long context window, which stretches up to 128,000 tokens<\/a>. That means it can process and remember way more information at once, making it great for tasks that need deep context, like summarizing long reports or analyzing complex documents.<\/p>\n\n\n\n

V3 also predicts multiple tokens<\/a> at a time, which speeds up responses. This means that it performs at the level of top models like GPT-4o and Claude 3.5 Sonnet, but at a fraction of the cost. DeepSeek reportedly trained\u00a0this model on 2,048 Nvidia H800 GPUs over just 55 days<\/a>, spending around $5.6 million<\/a>, which is incredibly low compared to what OpenAI and Google are spending. However, some have disputed this number<\/a>.<\/p>\n\n\n\n

DeepSeek R1: the deep thinker<\/h4>\n\n\n\n

R1 is different. It\u2019s built specifically for reasoning tasks\u2014the kind where step-by-step logic matters. While it shares some of V3\u2019s foundations, it was fine-tuned to explain its thinking as it solves problems. This makes it ideal for things like complex problem-solving, decision-making, and breaking down arguments.<\/p>\n\n\n\n

DeepSeek also experimented with different training approaches for R1. Some versions were trained entirely with reinforcement learning<\/a>, skipping the usual supervised fine-tuning process. Despite the unconventional method, R1 competes with top-tier models, such as OpenAI’s o1, in reasoning-heavy tasks.<\/p>\n\n\n\n

Why DeepSeek matters<\/h2>\n\n\n\n

DeepSeek is a classic case of second-mover advantage<\/a>. OpenAI, Google, and Anthropic spent billions pushing the frontiers of AI. But DeepSeek sat back, learned from their successes and mistakes, and built something leaner, cheaper, and just as good.<\/p>\n\n\n\n

This is bad news for big AI companies. They\u2019ve relied on their massive research budgets as a competitive moat, keeping smaller players from catching up. But DeepSeek\u00a0proves you don\u2019t need a multi-billion-dollar budget to make a top-tier AI model. If they can do it, who else will?<\/p>\n\n\n\n

Did DeepSeek copy ChatGPT? The (currently unproven) allegations <\/h2>\n\n\n\n

With any disruptive tech, there are always accusations. Some industry experts suspect that DeepSeek might have used OpenAI\u2019s ChatGPT to train its model. The theory? Model distillation<\/a>, allegedly.<\/p>\n\n\n\n

Model distillation<\/a> is when a smaller AI model learns by mimicking the outputs of a larger one. Instead of training from scratch, the student model \u201clearns\u201d patterns and behaviors from a teacher model. This is a very common technique in the industry and is an effective way to create smaller, more efficient models. If DeepSeek did this using OpenAI\u2019s ChatGPT, it could explain why it is so effective despite its reportedly lower training budget.<\/p>\n\n\n\n

But here\u2019s the thing: No one has actually proven this. OpenAI is reportedly investigating<\/a>, but as of now, it\u2019s just speculation.<\/p>\n\n\n\n

Whether or not distillation played a role, DeepSeek\u2019s success suggests that OpenAI\u2019s moat might not be as wide as they thought. Even if DeepSeek played fair, other companies will find ways to train AI cheaper and faster. That changes everything.<\/p>\n\n\n\n

DeepSeek, ChatGPT, and data privacy<\/h2>\n\n\n\n

Privacy concerns are a huge deal in AI, and DeepSeek is already under scrutiny. South Korea\u2019s National Intelligence Service (NIS) claims that DeepSeek  \u201cexcessively\u201d collects personal data<\/a>. They allege that:<\/p>\n\n\n\n