Once upon a time, the tech clarion call was “cellphones for everyone” – and indeed mobile communications have revolutionized business (and the world). Today, the equivalent of that call is to give everyone access to AI applications. But the real power of AI is in harnessing it for the specific needs of businesses and organizations. The path blazed by Chinese startup DeepSeek demonstrates how AI can indeed be harnessed by everyone, especially those with limited budgets, in order to meet their specific needs. Indeed the advent of lower-cost AI promises to change the deeply-entrenched pattern of AI solutions often remaining out of sight for many small businesses and organizations due to cost requirements.
LLMs are – or were – a pricey endeavor, requiring access to massive amounts of data, large numbers of powerful computers to process the data, and time and resources invested in training the model. But those rules are changing. Operating on a shoestring budget, DeepSeek developed its own LLM, and a ChatGPT-type application for queries – with a far smaller investment than those for similar systems built by American and European companies. The approach of DeepSeek opens up a window into LLM development for smaller organizations that don’t have billions to spend. In fact, the day may not be far off when most small organizations can develop their own LLMs to serve their own specific purposes, usually providing a more effective solution than general LLMs like ChatGPT.
While debate remains over the true cost of DeepSeek, it’s not simply the cost that sets it and similar models apart: It’s the fact that it relied on less-advanced chips and a more focused approach to training. As a Chinese company subject to U.S. export restrictions, DeepSeek was unable to access the advanced Nvidia chips that are generally used for the heavy-duty computing required for LLM development, and was therefore forced to use less-powerful Nvidia H-800 chips, which cannot process data as quickly or efficiently.
To compensate for that lack of power, DeepSeek took a different, more focused and direct approach to its LLM development. Instead of throwing mountains of data at a model and relying on computing strength to label and apply the data, DeepSeek narrowed down the training, utilizing a small amount of high-quality “cold-start” data and applying IRL (iterative reinforcement learning, with the algorithm applying data to different scenarios and learning from it). This focused approach allows the model to learn faster, with fewer mistakes and less wasted computing power.
Similar to how parents may guide a baby’s specific movements, helping her successfully roll over for the first time – rather than leaving the baby to figure it out alone, or teaching the baby a wider variety of movement that could in theory help with rolling over – the data scientists training these more focused AI models zoom in on what is most-needed for certain tasks and outcomes. Such models likely do not have as wide of a reliable application as larger LLMs like ChatGPT, but they can be relied upon for specific applications, and carrying those out with precision and efficiency. Even DeepSeek’s critics admit that its streamlined approach to development significantly increased efficiency, enabling it to do more with far less.
This approach is about giving AI the best inputs so it can reach its milestones in the smartest, most efficient way possible, and can be valuable for any organization that wants to develop an LLM for its specific needs and tasks. Such an approach is increasingly valuable for small businesses and organizations. The first step is starting with the right data. For example, a company that wants to use AI to help its sales and marketing teams should train its model on a carefully selected dataset that hones in on sales conversations, strategies, and metrics. This keeps the model from wasting time and computing power on irrelevant information. In addition, training needs to be structured in stages, ensuring the model masters each task or concept before moving onto the next one.
This, too, has parallels in raising a baby, as I have learned myself since becoming a mother a few months ago. In both scenarios, a guided, step-by-step approach avoids wasting resources and reduces friction. Finally, such an approach with both baby humans and AI models results in iterative improvement. As the baby grows, or the model learns more, its abilities improve. This means models can be refined and improved to better handle real-world situations.
This approach keeps costs down, preventing AI projects from becoming a resource drain, making them more accessible to smaller teams and organizations. It also leads to better performance of AI models more quickly; and, because the models are not overloaded with extraneous data, they can also be adjusted to adapt to new information and changing business needs – key in competitive markets.
The arrival of DeepSeek and the world of lower-cost, more efficient AI – although it initially spread panic throughout the AI world and stock markets – is overall a positive development for the AI sector. The greater efficiency and lower costs of AI, at least for certain focused applications, will ultimately result in more use of AI in general, which drives growth for everyone, from developers to chipmakers to end-users. In fact, DeepSeek illustrates Jevons Paradox – where more efficiency will likely result in more use of a resource, not less. As this trend looks set to continue, small businesses that focus on using AI to meet their specific needs will also be better set for growth and success.
Credit: Source link
