According to S&P Global, 2024 will be the year of AI ‘app-makers.’ Foundation models like large language models (LLMs) have dominated recent discussions. But, now investors are increasingly focused on companies developing AI applications that deliver tangible benefits for specific use cases. In fact, according to data from S&P Global Market Intelligence and 451 Research, AI companies without their own foundation models attracted more than twice as much investment in the first quarter of 2024, compared to the same period last year.
One of the most exciting promises of AI is it’s ability to save workers time. But, for AI to make a meaningful impact, businesses need AI tools that are tailored to specific industries or job roles. At the same time, these tools must be trustworthy and reliable. Yet, while AI chatbots built on LLMs can communicate well and offer general advice, they often lack the specialized knowledge or tools required. This makes them susceptible to inaccuracies or hallucinations due to their broad range of training data. This is where more targeted tools, fine-tuned for their specific use cases, are more likely to provide reliable and accurate outputs.
Senior Director, New Product Solutions at Dropbox.
To illustrate this point, consider the upcoming Olympics. Foundation models are like the core traits of a good Olympian, representing fitness, dedication, and an unwavering pursuit of excellence. However, the Olympics features 32 sports with over 400 different events, each requiring different skills and experience—much like the diverse industries and job roles in society. And, while AI provides the core technology that will power various products and services, each of these individual products needs to be specialized with the appropriate skills to provide value for their specific use case.
It’s rare for one athlete to compete in multiple different sports or disciplines at the Olympics. Each athlete is highly specialized for their specific sport. A sprinter for example optimizes their strength and physique to be powerful and fast over short distances. This however means they are not suited to other disciplines, such as long distance running. The most prominent AI chatbots today are all-rounders. They are designed to have general world knowledge across a broad range of topics. A given chatbot may be able to provide surface level information on a broad range of topics, but it may not excel at more specific tasks.
Take an AI-powered universal search tool for example. It needs to be able to find and retrieve the correct information, fast. Like a sprinter running the 100m Dash, it is optimized to save crucial seconds each time it performs. However, there are other tasks that may require an AI designed for sustained performance over a longer period of time, more like the long distance runner. For example, predictive AI models in business forecasting need to learn the patterns of activities of each business by analyzing historical data, building up this knowledge with use over time.
By becoming specialized in the business’ operations it can provide forecasts on the future trajectory of the company based on previous outcomes. Predictive AI models also need to constantly adjust forecasts based on the continuous changes in operations and external business factors. But with recent research from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrating that multiple large language models working together provides a more accurate outcome, perhaps a new type of AI ecosystem will emerge.
Is the future of AI a decathlete or a team of specialised athletes?
As we look at the trajectory of the AI ecosystem, we can see two distinct paths the industry can move in. The first is a race to create the best general-purpose AI model. This AI system would perform to a high-level across a variety of tasks, like a decathlete is able to compete across various events, from sprinting to long jump and pole vault. The advantage of this path would be a seamless employee experience which streamlines workflow. However, like the decathlete, who may not match the specialist’s performance in any single event, a general AI model might struggle to achieve the same level of excellence as more focused tools.
The alternative path sees the future AI ecosystem as a network of specialist AI products, more like a team of specialist athletes. In this model, each AI specializes in a particular domain, much like how individual athletes focus on specific sports. This approach mirrors how an Olympic team combines the talents of sprinters, swimmers, and gymnasts to maximize their collective medal potential for their country. The specialization ensures that each AI performs optimally within its domain, often surpassing the capabilities of a general-purpose system. However, the success of this networked approach will require sophisticated coordination and interoperability to create a seamless experience for users.
As we try to predict how future AI ecosystems will evolve, we might look to the Olympics in Paris this summer for a glimpse at two potential paths. Whether we end up with a decathlete style general-purpose AI tool or a network of tools that resembles a team of specialised athletes will depend on the objectives and decisions of the businesses in the collective technology industry. Much like how each country will have different objectives going into the Olympics.
From strategically focusing on a specialism, to optimizing the likelihood of a win, to a broader approach to win as many gold medals across as many disciplines as possible, the type of AI ecosystem each business will implement will heavily depend on their own unique objectives. For some businesses, growth through acquiring market share in a fluid market will require speed and agility, whereas, customer retention in a stagnant market will require a more strategic long term plan.