Artificial intelligence compute products are in short supply and facing hot demand.
Generative AI, which has moved at a rapid clip from corporate research laboratories and academic nonprofits to the forefront of the commercial landscape, has subsequently seen demand for AI application acceleration rise.
“The ChatGPT light bulb went off in everybody’s head, and it brought artificial intelligence and state-of-the-art deep learning into the public discourse,” Andy Hock, senior vice president of product and strategy at Cerebras, told PYMNTS during a conversation for the “AI Effect” series.
“And from an enterprise standpoint, a light bulb went off in the heads of many Fortune 1000 CIOs and CTOs, too,” Hock said. “These generative models do things like simulate time series data. They can classify the languages and documents for applications, say, in finance and legal. They can also be used in broad domains to do things like help researchers develop new pharmaceutical therapies or better understand electronic health records and predict health outcomes from particular treatments.”
Hock explained that Cerebras has built a new class of processor and computer system specifically tailored for AI work. Its processor, which is the largest computer chip ever built — around the size of a dinner plate — provides the necessary compute, memory and communication requirements for large-scale AI. The system allows developers to research, develop and bring to life the next generation of AI models and applications for large enterprise work.
The transformative potential of generative models, after all, requires an equally transformative compute engine to power it.
Read also: UAE Launches Bilingual Gen AI Named Jais
Cerebras’ systems reduce the training time of large-scale AI systems to days or weeks, enabling faster iterations and quicker market introduction of new capabilities.
And while the readiness of companies, particularly in security and accuracy-critical industries like finance, healthcare and manufacturing, to train and build AI systems responsibly may be, as Hock explained, “relatively middle of the road” in terms of their ability to train and build AI models, organizations are evolving rapidly — as is the availability of engineering resources.
“Many big enterprises have extraordinary data assets, but data that’s ready to be used to train one of these models — things like whether it’s clean, de-duplicated, and do they know how to tokenize it and get it ready to be fed into one of these AI models — that’s a different matter,” said Hock, noting that the percentage of people around the world who know how to build an AI system is small.
Still, he added that bigger corporate enterprises “have been thinking for years” about how to maintain data security and protect their critical intellectual property, which are two of the biggest foundational elements behind the responsible development of AI.
Hock also emphasized the importance of defining evaluation metrics and engaging domain experts to ensure models perform as desired and comply with regulations.
The process of training an AI model involves datasets, a serious amount of compute power and domain expertise, he said.
Certain industries, such as finance, healthcare and manufacturing, are well-suited for domain-specific models due to their access to relevant data and familiarity with using compute to accelerate their business. Hock also highlighted the potential of AI models in scientific research, including fundamental biology, energy, climate science and material science.
“Being better suited for the development of domain-specific models boils down to the availability of data,” he added. “Commercial domains that have access to data and are familiar with using compute to accelerate their business are really well positioned [to capture AI’s benefits].”
Data readiness, domain expertise and deep collaboration are crucial for organizations looking to most effectively use AI systems to streamline their own workflows and transform their legacy processes, he said.
Speaking about Cerebras’ Jan. 15 collaboration with the Mayo Clinic to advance AI in the healthcare space, Hock explained that, “there’s massive datasets on one side, and a massive set of problems on the other side … and a massive opportunity to improve healthcare outcomes across the board.”
“Collaboration and a partnership model between computational and AI experts on one side, and domain experts in a particular enterprise industry on the other side, is really, really important as well as having the right compute engine and datasets under the hood,” he added.
Looking ahead, Hock predicted that two innovations will streamline AI development and accelerate generative AI. First, purpose-built AI compute platforms, like Cerebras’, will play a transformative role in the field. Second, the emergence of multilingual models and the convergence of language and multimodal data, such as imagery and video, will expand the applications of AI beyond traditional language-based models.
“I’m declaring my bias here, but I think there are tremendously positive outcomes [AI will bring to the business landscape], and in order to push forward on those, we also should push forward quickly with the right tools on how to do this work safely and responsibly,” Hock said.
“AI will help enterprises move with greater efficiency and make better products, and it’s also going to help our human workforce and bright analytical minds actually do the kind of work that human minds are really well suited for,” he added.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.