The DeepSeek Revolution: Why AI Experts Are in a Frenzy

A few days ago, only the most devoted enthusiasts (I say this as one) had any awareness of DeepSeek, a Chinese A.I. branch of the equally intriguing High-Flyer Capital Management, a quantitative analysis (or quant) firm that initially established itself in 2015.

Yet in the past few days, it has arguably become the most talked-about entity in Silicon Valley. This surge in interest is significantly attributed to the launch of DeepSeek R1, a new extensive language model exhibiting reasoning capabilities akin to OpenAI’s currently best-performing model o1 — taking several seconds or minutes to respond to challenging inquiries and tackle complicated issues as it deliberates on its analysis in a systematic, or “chain of thought” manner.

Moreover, DeepSeek R1 has achieved scores equal to or surpassing those of OpenAI’s o1 across a range of third-party benchmarks (evaluations designed to assess AI performance at answering questions across various topics), and was allegedly trained at a significantly reduced cost (approximately $5 million), utilizing far fewer graphics processing units (GPUs) under a strict embargo enforced by the U.S., the home ground of OpenAI.

Conversely, while o1 is accessible solely to paying ChatGPT subscribers of the Plus tier ($20 per month) and higher tiers (such as Pro priced at $200 monthly), DeepSeek R1 has been made available as a completely open source model, which explains its rapid ascent on the charts of AI code-sharing platform Hugging Face’s most downloaded and active models.

Furthermore, due to its fully open-source nature, users have already fine-tuned and adapted various iterations of the model for specific applications, such as optimizing it to operate on mobile devices or integrating it with other open-source models. Even for development purposes, DeepSeek’s API costs are over 90% lower than the corresponding o1 model from OpenAI.

Perhaps most impressively, you don’t even need to be a software developer to utilize it: DeepSeek features a free website and mobile app for U.S. users, complete with an R1-powered chatbot interface that closely resembles OpenAI’s ChatGPT. In a further contrast, DeepSeek has outperformed OpenAI by linking this powerful reasoning model to web search — a feature OpenAI has yet to incorporate (web search is presently only accessible on the less powerful GPT family of models).

An open and shut irony

There’s a rather delectable, or perhaps unsettling irony to this considering OpenAI’s initial ambitions to democratize AI for the general populace. As NVIDIA Senior Research Manager Jim Fan stated on X: “We are experiencing a timeline in which a non-U.S. company is sustaining the original mission of OpenAI – genuinely open, innovative research that empowers everyone. It makes no sense. The most entertaining outcome is the most probable.”

In the words of X user @SuspendedRobot, referring to claims that DeepSeek seems to have trained on question-answer outputs and other data produced by ChatGPT: “OpenAI took from the entire internet to enrich itself, DeepSeek took from them and returned it to the public for free; I believe there’s a certain British folktale regarding this.”

However, Fan is not the only one taking notice of DeepSeek’s achievements. The open-source nature of DeepSeek R1, its impressive performance, and the fact that it seemingly “emerged from nowhere” to rival the former leader in generative AI, has reverberated throughout Silicon Valley and beyond, based on my discussions and observations of various engineers, thinkers, and industry leaders. While not “everyone” is in a frenzy over it as my exaggerated headline implies, it is undoubtedly the hot topic in tech and business discussions.

A post shared on Blind, the platform for anonymous gossip in Silicon Valley, has been circulating, suggesting that Meta is in turmoil over the success of DeepSeek due to how swiftly it has overtaken Meta’s own endeavors to dominate open-source AI with its Llama models.

‘This changes the whole game’

X user @tphuang articulated compellingly: “DeepSeek has commoditized AI outside of the highest tier. A lightbulb moment for me in the first image. R1 is so much more affordable than U.S. labor costs that numerous jobs will become automated within the next five years,” later elaborating on why DeepSeek’s R1 is more appealing to users than even OpenAI’s o1:

“Three major drawbacks with o1: 1) too slow 2) too costly 3) lack of user control/reliance on OpenAI R1 addresses all these issues. A company can procure its own Nvidia GPUs, operate these models, and no longer have to concern itself with extra expenses or sluggish/unresponsive OpenAI servers.”

@tphuang also posed a captivating analogy in the form of a question: “Will DeepSeek be to LLM what Android became to the OS landscape?”

Web entrepreneur Arnaud Bertrand did not hold back regarding the shocking ramifications of DeepSeek’s triumph, writing on X: “There’s no exaggeration in stating how thoroughly this transforms the entire game. And not only concerning AI; it’s also a significant criticism of the U.S.’s misguided attempts to halt China’s technological advancements, without which DeepSeek might not have been achievable (as the saying goes, necessity is the mother of invention).”

The censorship issue

However, some have raised warning flags regarding DeepSeek’s swift ascent, asserting that as a startup operating out of China, it naturally adheres to that nation’s laws and content censorship mandates.

Indeed, my personal experience utilizing DeepSeek on the iOS app here in the U.S. found that it would not respond to questions about Tiananmen Square, the location of the 1989 pro-democracy student protests and uprising, followed by the subsequent violent crackdown by the Chinese military, resulting in at least 200, possibly thousands of fatalities, earning it the title “Tiananmen Square Massacre” in Western media outlets.

Ben Hylak, a former human interface designer at Apple and co-founder of AI product analytics platform Dawn, communicated on X how inquiring about this topic caused DeepSeek R1 to enter a convoluted loop.

As a member of the press myself, I naturally take the matter of freedom of speech and expression extremely seriously and it is arguably one of the most vital, undeniable causes I advocate for.

Yet I would be remiss not to mention that OpenAI’s models and products including ChatGPT also hesitate to reply to a broad array of inquiries concerning even harmless subjects — particularly those involving human sexuality and erotic/adult, NSFW material.

It’s not a direct comparison, of course. And there will be some who, due to hesitance in relying on foreign technology, may question DeepSeek’s ultimate worth and functionality.

Yet, one cannot dispute its efficiency and affordability.

In a period when 16.5% of all goods in the U.S. are imported from China, it’s challenging for me to advise against utilizing DeepSeek R1 based on apprehensions regarding censorship or security issues — particularly when the model code is readily accessible for download, can be taken offline, operated on devices in secure settings, and adaptively fine-tuned as desired.

I certainly perceive an underlying existential concern about the “decline of the West” and the “rise of China,” fueling some of the vigorous debates surrounding DeepSeek. Others have already linked it to how U.S. users migrated to the app Xiaohongshu (also known as “Little Red Book”) when TikTok faced a brief ban in this nation, only to be astonished by the quality of life in China portrayed in the videos shared on that platform. The introduction of DeepSeek R1 takes place within this narrative backdrop — one where China seems (and by numerous measures is evidently) advancing, while the U.S. appears (and by various metrics is also) in retreat.

The initial but certainly not the final Chinese AI model to disrupt the world

It will not be the final Chinese AI model to challenge the supremacy of Silicon Valley titans — even as they, like OpenAI, secure more funding than ever for their aspirations to create artificial general intelligence (AGI), systems outperforming humans in most economically advantageous tasks.

Just yesterday, a new Chinese model from TikTok’s parent company Bytedance — named Doubao-1.5-pro — was launched, delivering performance that rivals OpenAI’s non-reasoning GPT-4o model on third-party evaluations, yet at merely 1/50th of the price.

Chinese models have advanced remarkably quickly, drawing attention even from those outside the tech sector: The Economist magazine recently published an article about DeepSeek’s achievements and other Chinese AI initiatives, and political commentator Matt Bruenig shared on X that: “I have been extensively using Gemini, ChatGPT, and Claude for NLRB document summarization for nearly a year. DeepSeek is superior to all of them at this task. The chatbot version is available for free. The cost to utilize its API is 99.5% less than OpenAI’s API. [shrug emoji]”

What is OpenAI’s response?

It’s no surprise that OpenAI co-founder and CEO Sam Altman stated today that the organization would be integrating its forthcoming second reasoning model series, o3, into ChatGPT even for users without a subscription. OpenAI still seems to be forging its unique path with more proprietary and sophisticated models — establishing the industry benchmark.

However, the pressing question is: with DeepSeek, ByteDance, and other Chinese AI firms closely tailing it, how long can OpenAI sustain its lead in developing and launching innovative AI models? And if it does falter, how severe and rapid will its decline be?

OpenAI does possess a historical precedence favoring it. If DeepSeek and Chinese AI models indeed become analogous to LLMs as Google’s open-source Android did for mobile — capturing a significant market share for a time — one only needs to observe how the Apple iPhone, with its locked-down, proprietary, in-house approach, managed to capture the premium segment of the market and gradually expand into lower tiers, particularly in the U.S., to the point of commanding nearly 60% of the national smartphone sector.

Nonetheless, despite the numerous entities investing heavily in AI models from top laboratories, DeepSeek illustrates that equivalent capabilities might be accessible for considerably less and with considerably more control. In a corporate environment, this could suffice to win the competitive battle.