Ekaterina Almasque
Artificial Intelligence is no longer a tech buzzword. It is here and making real-world differences. If you asked any prominent organization leader in 2018 what their thoughts on AI were, you'd almost always get the same response: we need AI, but we don't know where to start.
Today, we are witnessing real-world problems that were once impossible to solve being remedied by a new generation of AI models. However, one issue with these models is that they are becoming extravagantly large, with millions, billions, and, sometimes, trillions of parameters. This either leads to unsustainable, resource-hungry AI or the need to create smaller models that produce the same results while keeping costs under control. However, the latter requires a true breakthrough in our approach to AI.
Aside from edge use cases, why aren't we seeing these smaller AI models?
There needs to be heavy investment in data and infrastructure to create functional AI models. Only major corporations, such as Google, Microsoft, and OpenAI, have the scale to invest in this infrastructure, while most other enterprises simply can't. This creates a bottleneck for further AI progression while creating huge gaps between the industry's "haves" and "have nots." Semi-solutions are provided by large players with pretrained foundation models. However, while these models are useful for some basic use cases, they afford little control and flexibility, hampering their ability to solve real-world problems at the hands of an enterprise. Moreover, there is also growing concern about escalating the bias between those who are already players in the "AI economy" and those who are exceedingly excluded from the future data economy.
“Within this layer, emerging, innovative start-ups will play integral roles in filling various gaps or even becoming the platform that connects many dots.”
While creating smaller AI models is a solution to the above problem, a fundamental issue hinders it. Current AI algorithms need a vast dataset to have reasonable accuracy, which if you aren't a major player like Google, you just won't have.
Disrupting AI's infrastructure
Overcoming the obstacles mentioned above requires the building of an entirely new infrastructure layer with a significant breakthrough. Within this layer, emerging, innovative start-ups will play integral roles in filling various gaps or even becoming the platform that connects many dots. However, these same start-ups will face spiraling competition for talent from well-resourced large-scale enterprises, as well as recurring issues with failing to commercialize their product effectively. Unless there is a mind-blowing breakthrough tomorrow, no start-up will be able to provide this for some time.
The task of creating an alternative infrastructure to Google or OpenAI's massive datacenter structure of GPUs dedicated to single AI workloads could come with support and significant funding from governments. However, for this to occur, we will need to see an influx of government investment, either directly or via specialized funds, as it is currently siloed and needs to be collated. We are excited to welcome a disruptive player who will play a critical role in the construction of this infrastructure, rising to the challenge that AI presents.
While this new infrastructure layer will be crucial in building a better future with AI, major focus must be placed on spearheading inclusivity within AI. The pretrained models in use are mainly trained using US or Chinese datasets. Unfortunately, these datasets are limited to single-geography datasets,, leading to immense distortion within the industry. If we want to build an inclusive AI society moving forward, we must disrupt this model entirely by championing the use of datasets from across the globe. By doing so, we will ensure that AI innovation can continue, but from the ground up and anywhere on the planet.
A better future with AI
Unfortunately, the current dynamic within the AI industry is becoming more and more prohibitive to bottom-up innovation and large-scale enterprise adoption. With exclusive datasets driving large-scale, single-use pre-trained models, significant disruption is needed to drive broader use cases and create AI for people. Governments are too siloed in their approach to the problem, and startups need to accomplish "mission impossible" to spearhead this disruption. We need to rethink the status quo. Perhaps we need to rethink AI as we know it.
