Are you ready to bring more awareness to your brand? Consider becoming a sponsor for The AI Impact Tour. Learn more about the opportunities here.
Absolutely without a doubt, hands down, 2023 was the year of AI.
And, no surprise: “Next year, just like this year, is going to be all about AI,” John Roese, global CTO for Dell, told VentureBeat in a year-end forecast.
While so far the AI story has been experimental, inspirational, “largely just ideas,” the speed of its evolution is sevenfold that of traditional technology. Very quickly, enterprises will move from theory to practice and everything in tech will be focused on AI’s “aggressive accelerated adoption.”
“Next year is year two of the AI era,” Roese said. “The first wave of practical, in-production AI systems will start to take place in enterprise.”
The AI Impact Tour
Getting to an AI Governance Blueprint – Request an invite for the Jan 10 event.
Identifying the ‘heavy lift’ of AI
In 2024, as enterprises begin to put AI into production, they must implement a top-down strategy, Roese says.
“You’re going to have to decide which areas are your real core,” he advised. “What makes you, you — that’s the place where you want to apply the heavy lift of AI.”
Dell, for instance, has roughly 380 AI-related ideas in the pipeline, he noted. But even as a large enterprise, the company probably only can handle just a handful of those. As he put it, enterprises might rush to do the first four projects on their lists — ultimately outpricing the fifth, which could have been the truly transformative one.
“You have to learn to prioritize,” said Roese. “You might have several good ideas, but which are most important to your company?”
Shift to inferencing, cost of operation
As they shift to inferencing in 2024, enterprises will need to determine the best ways to design and place infrastructure, Roese pointed out.
“People are going to have to start thinking about the actual topology,” he said. “The world of technology is distributed, AI is likely going to be distributed.”
Security is just as critical, as bad actors will begin to directly target inference. Enterprises must consider: “What’s the security wrapper around this?”
Furthermore, the economic discussion around AI will shift in 2024 from the cost of training to the cost of operation, Roese said.
While the cost to fine-tune a model can be high and infrastructure demands are significant, that is just a small part of the AI investment, he pointed out. The training cost is tied to one-time model size and data set use, while the price tag for inferencing is based on utilization, data type, user base size and ongoing maintenance and fine-tuning.
“The meta theme is: AI is going to become a lot more real, and that has consequences,” said Roese.
Gen AI supply chain will improve
There’s no doubt that gen AI systems are “enormous,” and that we need “more tools, more tech and a bigger ecosystem” to put AI to work, said Roese.
While there has been much discussion and concern around availability and sourcing, he predicts that 2024 will bring an “abundance” of tools and models.
“Our ecosystem of AI tools and services is expanding, diversifying and scaling,” he said.
Tools for building systems are getting better all the time, and he expects a diversification of AI frameworks — such as the new Linux Foundation UXL project — and increased availability of both closed and open-source models and tools.
Developers will also be able to easily use and create interfaces to “multiple types of accelerated compute and integrated frameworks” such as PyTorch on the client side and ONYX on the infrastructure side.
“Next year we will have more options at every layer,” said Roese.
Zero trust finally becomes real
Cybersecurity is broken — breaches continue to accelerate even as enterprises incorporate the latest security methods and tools.
The real way forward is through a different architecture, Roese said: Zero trust.
“Everything is authenticated and authorized,” he said. “Everything is tightly coupled in real-time.”
Still, to this point, zero trust has largely been confined to a buzzword, as it’s difficult to put into practice.
“The reason it hasn’t taken off is it’s actually quite hard to do,” said Roese. “It is almost impossible to take an existing brownfield enterprise and make it a zero-trust environment. You would have to unwind every security decision you ever made.”
But now, since AI is essentially brand new, zero trust can be built in from the ground up in truly greenfield environments.
Roese pointed to Dell’s in-the-works zero trust tool Project Fort Zero, which is expected to be validated by the U.S. Department of Defense and made available on the market in 2024.
“We really are losing the cyber war right now,” said Roese. “We need to get out of the hole we’re in, in cyber. The answer is right in front of us. It’s zero trust.”
The ‘common edge’ emerges
To get the most value out of their data, enterprises should be as close to the source as possible.
Going forward, “we are going to do more processing of data out in the real world than in data centers,” said Roese.
This will give rise to what Dell calls “modern edge” multi-cloud platforms.
As he explained, the default “cloud extension” point tools deliver edge for specific workloads. This means that, as enterprises use more clouds and cloud services, edge systems overpopulate — that is, there’s one for every cloud, workload and system.
Enterprises may have hundreds of workloads at the edge, and if they all need their own architecture, it would be “untenable” and “unbearably complex,” Roese contends.
To address this, Dell recently introduced NativeEdge, a common edge platform that supports software-defined edge workloads from any IT, cloud or IoT system. Roese expects this approach to become more prevalent in 2024 as enterprises see the disadvantage of “mono-edges.”
As he put it, “Now, almost all edge service providers have decided they don’t want to build hardware, they want to deliver edge services as containerized code.”
Looking further afield: Quantum will power AI
Large-scale AI presents what Roese calls a “massive parallel problem.”
“Transformers, diffusion models and other new techniques under gen AI are extremely resource-intensive probabilistic functions,” he said.
While it likely won’t be realized for a few years to come — scientists need to get beyond the current 1,000 qubit range to allow for a viable, commercial-grade system — “the workload that quantum will unlock is AI,” said Roese.
The AI of the future, he said, will be spread across a diverse hybrid compute architecture, including quantum.
“The problems of gen AI mathematically are really well solved by quantum computing,” he said. Quantum is “exceptionally good” at highly-scaled optimization problems where the goal is to find the best answers to questions within an “almost infinite set of options.”
“Quantum computers are basically probabilistic computers, they’re really good at things with a billion permutations,” said Roese.
Quantum has been teased for some time now, but Roese affirms that there will come a day — soon — when sufficiently mature quantum systems are available.
“That will have an amplifying effect on wherever we are with AI,” he said. “It will be a bigger disruption than ChatGPT.”
VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.