McKinsey partner: Generative AI helps solve various challenges in cloud migration

McKinsey & Company partner Bhargs Srivathsan said at a recent conference in Singapore that as long as it is used properly, generative AI technology is expected to reduce cloud migration workload by 30% to 50%.

Srivathsan believes that “the current progress can only be said to have just taken the first step. As the large language model (LLM) matures, the timetable for migrating workloads to the public cloud will continue to shorten, The efficiency of the migration process can also be improved accordingly."

She suggested that organizations first use a large language model to understand the infrastructure in the system, analyze the shortcomings and advantages, and then continue to apply it after the workload transfer is completed. AI tools see if the migration is actually effective.

In addition, you can also use large language models to complete more related work, such as writing explanatory materials such as architectural review committee guidelines.

The partner said that although many companies have just begun to consider adopting AI technology, 40% of the companies invested by McKinsey are already updating their IT investments.

Srivathsan believes that the relationship between generative AI and the cloud is "symbiotic".

“It must be admitted that without the popularization of public cloud, it would be impossible to truly bring generative AI into life. Correspondingly, generative AI can also effectively accelerate public cloud migration and help users migrate from the original There is a public cloud to unlock the separation.”

In Srivathsan’s view, the four core use cases of generative AI are content generation, customer engagement, creating synthetic data, and writing code. Of course, writing code here is not about completing software development from scratch. The coding ability of generative AI is mainly reflected in taking over legacy codes that no one is familiar with after employees leave, or converting original codes into new language forms.

She also emphasized that the reason why public cloud is more reliable than trying to build an internal model is because enterprise users often do not have sufficient GPU reserves. Moreover, the cost of ready-made commercial models on the market is also cheaper than self-training.

Srivathsan pointed out that corresponding guardrails can also be set up for users who are in regulated industries, have large amounts of proprietary data, or are worried about intellectual property rights being infringed.

In her opinion, large language models will mainly run in ultra-large-scale infrastructure environments in the next five or six years until the models mature. And unlike what many people imagine, the implementation of generative AI does not necessarily require such exaggerated computing power reserves. After all, there are few use cases that place such stringent requirements on latency.

In other words, unless it is the autopilot function running on Tesla, or the software responsible for directing the real-time operation of the manufacturing workshop, there is really no need to pile up the hardware too much.

Also, in most cases there is no need to use custom or large-scale models.

The McKinsey partner commented, “Many companies think they need to buy a supercar to deliver pizza. Of course, they don’t need to. Models that really meet the needs are often less complex and not that big. For example For example, there is definitely no need to use a large model with 65 billion parameters to generate customer service support scripts."

But she also gave suggestions that if developers are accessing non-proprietary models or models that they should not have access to, data, it is necessary to add an API gateway between inside and outside the organization to establish a "real-time alert" mechanism.

The above is the detailed content of McKinsey partner: Generative AI helps solve various challenges in cloud migration. For more information, please follow other related articles on the PHP Chinese website!