The initial panic surrounding DeepSeek, a Chinese AI company, and its seemingly revolutionary, cost-effective AI model has largely subsided, revealing the overblown nature of the initial market reaction. The dramatic drop in Nvidia’s valuation and the frenzy within the AI community appear to have been driven more by political anxieties than a genuine technological threat. Experts like George Morgan, CEO of Symbolica, dismiss the market’s response as misinformed, suggesting a bias against a Chinese company achieving such a feat. The core innovation of building cost-efficient foundational models, as DeepSeek claims to have done, is not novel; it’s been an ongoing area of research for years.
The central claim of DeepSeek, training a large language model for a mere $5.6 million, is misleading. This figure represents the cost of a single training run, not the entire development process. Building a large language model typically requires numerous, often thousands, of such training runs. DeepSeek achieved its reported cost reduction by leveraging pre-existing open-source models, like Meta’s Llama, thus significantly lowering the financial barrier to entry. Their own technical paper acknowledges that the publicized figure excludes the substantial cost of prior research, implicitly admitting to much higher overall development expenditures.
Industry figures like Writer CEO May Habib and Qodo CEO Itamar Friedman echo this skepticism. They point out that DeepSeek’s publicized cost likely represents only the final stage of training and omits the cumulative expenses incurred throughout the development process. This practice of focusing on a single, isolated cost figure obscures the true resources required to build such a model. While DeepSeek’s cost efficiency claims are questionable, its contribution to the field shouldn’t be entirely dismissed. The company employed reinforcement learning, a well-established technique, to achieve impressive results and, crucially, made their technology open-source, enabling widespread access and replication.
DeepSeek’s emergence has sparked a much-needed discussion about resource efficiency in AI development, particularly in light of OpenAI’s significant fundraising efforts for expansive data centers. Timnit Gebru, founder of the Distributed Artificial Intelligence Research Institute, highlights DeepSeek’s challenge to the prevailing narrative that massive resources are essential for building advanced AI models. Their accomplishment compels a reevaluation of current investment strategies within the AI industry and demonstrates that significant progress can be achieved without exorbitant expenditure. This challenge to the status quo has inevitably led to a war of words mirroring the ongoing competition in cost reduction.
OpenAI has accused DeepSeek of violating its terms of service by utilizing outputs from its proprietary models, a practice known as distillation, to train their own AI systems. This accusation is ironic, given OpenAI’s own controversial practice of scraping publicly available data, including copyrighted material, to train its models, a practice that has resulted in lawsuits. Gebru finds OpenAI’s accusation “laughable,” given their own ongoing legal battles defending their use of public data. The irony of OpenAI accusing another company of exploiting data underscores the complex and often contentious issues surrounding data usage in AI development.
Ultimately, DeepSeek’s approach is not groundbreaking. Microsoft adopted a similar strategy with its Phi models, training them on outputs from more advanced models like OpenAI’s GPT-4. Experts agree that DeepSeek’s work, while noteworthy, isn’t the revolutionary “Sputnik moment” some have claimed. The company’s contribution lies not in a novel technological breakthrough, but in demonstrating a more accessible and cost-effective approach to AI development, challenging the prevailing narrative of resource-intensive model training and sparking a critical conversation about efficiency and accessibility in the AI landscape.
