The breakout news of DeepSeek’s smaller, more nimble language model has sparked meaningful conversations around the size and scope required for effective AI implementation. Until DeepSeek, building a proprietary AI model was out of reach for all but the biggest players with the deepest pockets and the newest toys. Now, the scaffolding from the larger open-source models can be used to craft smaller, more specialized models, so more industry players can get in on the AI action.
Think Before Automating
With this new avenue, small and medium-size players have more questions to ask when it comes to building their own (smaller) large language models. A company hoping to get in on the action needs a clear understanding of its business model, the pain point it hopes to address via automation and the balance of human input versus AI tools required.
Leaders also need to understand how workflows might be redesigned so humans and AI can work in tandem. Finally, they’ll need to think through how to best leverage staff strengths and skillsets against company goals, timelines and finances. All that information will help them decide if creating a proprietary AI model makes sense and how best to deploy it. With so many factors to consider, it seems like you almost need AI to decide if you need AI!
At the moment, using AI is a developing technique, but soon it could become a necessity as companies armed with proprietary AI tools will be able to move faster. AI-enabled businesses will be able to analyze deals more quickly, read markets more effectively and be better equipped to capitalize on volatile markets. Those without proprietary AI tools might find that they can’t get to the deal table or the stock market quickly enough to beat out competitors.
In the near future, even the most expedient processes will require human oversight or input in some form. For example, an AI model built to help evaluate deals might be able to quickly pull data for all producing wells from a specific formation, then suggest the best-suited graphs or maps to display that data, but it will take a well-trained geoscientist to put all the pieces together and evaluate if it makes sense to sign a contract.
In this case, geoscientists can spend their time using their brain for what they love: deeply thinking about all the available data, understanding business needs and putting the pieces together in a comprehensive analysis. They might still perform some data cleanup along the way, but 80 percent of the work can be done by a machine, while the most valuable 20 percent is done by a person.
Any given project will require a human-to-AI balance: it might be 80-20, 50-50 or 100-0 for each task within the larger project. The real question to ask is what work is best suited to the machine and what is best suited to a human, and the answer to that will vary based on the purpose of the model. The term “fit for purpose” comes into play here, which means building an AI product that is designed specifically for the task to which it is dedicated.
By focusing product development on a single task, costs are reduced, workflows concretely outlined and efficiencies improved. Instead of having one model that might perform many tasks decently, you could create many, smaller models that do their specific tasks very well.
Until now, having multiple models for multiple purposes was a laughable resource vacuum. But DeepSeek has demonstrated that a more nuanced, fiscally conservative option is (or soon will be) available, and there might be an option to have several smaller, cheaper and more task-focused models that don’t need data repositories requiring their own nuclear power plants to function. Essentially, we might use five or six different specialized AI models within a single project versus one behemoth to address all the project’s needs.
There might be another five to ten years before the industry figures out how to best match models to tasks, create all the models it needs and so on, but that’s not very far away. Given the speed with which the tech sector is moving, it’s worth taking the time to run these thought experiments now and consider the possibilities so that if they come to fruition sooner rather than later, we’re prepared.