Chinese tech company DeepSeek has been delighting (and sometimes baffling) the world with its budget-friendly chatbot, DeepSeek R1. On the bright side, this open-source marvel has seen a meteoric rise in popularity—Similarweb reports a jump from a modest 300,000 to an impressive 6 million daily users. Even giants like Microsoft and Perplexity have hopped on the DeepSeek train, weaving it into their own platforms. But before we break out the celebratory confetti, it’s worth noting that our new AI buddy comes with a few… interesting quirks.

Let’s start with the big one: safety. Researchers at Cisco put DeepSeek R1 through its paces, using just 50 prompts related to cybercrime, disinformation, and various illicit activities. The result? A flawless 100% success rate in bypassing the model’s built-in filters. In other words, if you’re looking for shady content, DeepSeek won’t just fail to say no—it’ll roll out the red carpet and hand you the instructions.

For context, GPT‑4o from OpenAI blocks about 14% of harmful requests, Google’s Gemini 1.5 Pro comes in at 35%, while Claude 3.5 manages a respectable 64%. The reigning champ so far is the preliminary version of OpenAI o1, which blocks 74% of attacks. Granted, DeepSeek’s development budget was only a measly $6 million, so perhaps we shouldn’t be too surprised when it’s stacked up against GPT‑5’s half-billion-dollar training process.

Then there’s the question of what DeepSeek doesn’t talk about—namely, anything politically sensitive, especially topics involving China. Ask about the Uyghurs or the events in Tiananmen Square in 1989, and you’ll likely get a polite but firm, “Sorry, that’s beyond my capabilities. How about we change the subject?”

On one hand, you can’t really fault a chatbot for trying to steer clear of hot-button issues. On the other, it’s almost too predictable. If we’re being honest, ChatGPT also tap-dances around certain delicate themes, so DeepSeek isn’t exactly alone in this. But the irony is hard to miss: both bots end up wearing the badge of “censorship,” whether they intend to or not.

Here’s another fun twist: DeepSeek R1 still feels very much like a work-in-progress. It operates more like the early, free versions of ChatGPT—if you’ve used those, you know the struggle. Your first question is key, since that’s the query the model latches onto. After that, trying to hold a smooth conversation can feel like you’re lining up for a new smartphone on release day. DeepSeek’s servers often get so overwhelmed that users end up in a waiting queue, as if they’re waiting for a table at the trendiest brunch spot in town.

DeepSeek is a fantastic example of how a “budget AI” can both impress and amuse in equal measure. Sure, it’s gaining traction and even getting integrated into large-scale projects, but glaring security holes, heavy-handed censorship, and a few technical hiccups might make you wonder: is saving on development really worth it? Perhaps DeepSeek will evolve as more funding rolls in and more engineers hop on board. For now, though, it remains an intriguing—and slightly exasperating—new contender in the AI arena, with a dash of irony and a generous sprinkle of friendly skepticism.