The tech world is collectively losing its mind over DeepSeek, and for good reason. The Chinese startup burst into the limelight back in December after debuting the R1, a model that uses self-checking mechanisms to reason and deliver accurate answers to complex questions.
In Hardwired, AC Senior Editor Harish Jonnalagadda delves into all things hardware, including phones, audio products, storage servers, and networking gear.
Most chatbots today — like ChatGPT and Gemini — use inference models to come up with an answer, and they're trained on billions of inputs. While they excel at generalized queries, they struggle with solving complex coding and math problems, but this is slowly changing. OpenAI rolled out its o1 reasoning model at the end of last year, and it uses chain-of-thought — where it runs through multiple steps — to come up with an answer.
DeepSeek is challenging the status quo with its own solution. The DeepSeek R1 is a reasoning model that's built on the V3 large language model, and what's noteworthy about it is that it is claimed to be developed at a fraction of the cost of what U.S. entities are doing. Essentially, DeepSeek says it spent under $6 million to train its model, which is ridiculous — U.S. brands invested hundreds of millions of dollars in training each model, and billions in infrastructure.
Oh, and you don't need to pay anything to use it. It's the latter point that's causing so much consternation; to give you some context, OpenAI charges $200 a month for ChatGPT Pro, which is built on the o1 model. With DeepSeek R1, you get something that's just as good — for absolutely free. This runs counter to the fundamental ideal of Silicon Valley, where the tenet is that if there's a halfway- decent product, the main goal is to monetize it.
DeepSeek is doing things a bit differently; it open-sourced the r1, so anyone can take a look under the hood to see how it runs, or modify it as they see fit. It's no wonder, then, that DeepSeek climbed to the top of the App Store charts after just a week of availability. Its introduction had a seismic effect on the sector as a whole; the U.S. stock market lost a cumulative $1 trillion in value, with NVIDIA alone losing $589 billion in market value — the most by any organization on a single day.
Obviously, the market will correct itself, and the panic will ease as U.S. companies seek out ways to catch up — OpenAI is already touting that it will roll out "better models." But what DeepSeek exceled at doing is democratizing cutting-edge AI and making it accessible to everyone, and that alone is the biggest signal that the next big breakthrough in this category may not come from Silicon Valley after all.
While I'm not keen on generative AI, I dabble with ChatGPT, Gemini, and Claude occasionally, and they clearly have their use cases. Reasoning models, however, are actually interesting; they have the ability to use logical reasoning and understand context, so if you ask a follow-up question, you get relevant answers. The best part is that all of this runs on your phone — I had no issues using the AI assistant on the Honor Magic 7 Pro and Nubia Z70 Ultra among others.
They're particularly great at solving math problems and creative thinking, and more than anything else, this is where DeepSeek's R1 fulfils its potential. And best of all, it doesn't cost you anything to use it. The breakneck pace of this industry and the staggering amounts of money being thrown at anyone building AI tech means we don't have to wait too long to witness the next big milestone.
