Metaverse Drives Edge Computing

As video content distribution has shaped global demand for inter-continental data transport and high-speed connections between major data centers, the metaverse will shape data center and connectivity network requirements. And the key words are “more” and “less.”

More bandwidth; less latency. More artificial intelligence; less centralized computing. More at the edtge; less in the core. More symmetrical bandwidth, less copper.

“Making the metaverse a reality will require significant advancements in network latency, symmetrical bandwidth and overall speed of networks,” says Dan Rabinovitsj, Meta VP for connectivity. 

The metaverse “will require innovations in fields like hybrid local and remote real-time rendering, video compression, edge computing, and cross-layer visibility, as well as spectrum advocacy, work on metaverse readiness of future connectivity and cellular standards, network optimizations, improved latency between devices and within radio access networks (RANs), and more,” he says. 

Already, experts predict Metaverse environments will require more data centers, more edge computing, more distributed computing, more collocation, more content distribution mechanisms, will require more power consumption and more cooling.

Eventually, fully-developed metaverses will require advances in chip technology as well. Beyond all that, blockchain will probably be necessary to support highly-decentralized value exchanges. And it is impossible to separate metaverse platforms and experiences from use of artificial intelligence, for business or consumer uses. 

source: iCapital Network 

If metaverses are built on persistent and immersive computing and tightly-integrated software stacks, platforms will be necessary. New developments in chip manufacturing also will be needed. 

For connectivity providers--especially internet service providers--far lower latency will be key. Today’s latency-sensitive applications such as video calling and cloud-based games have to meet a round-trip time latency of 75 milliseconds to 150 ms. Multi-player, complex games might require 30 ms latency. 

“A head-mounted mixed reality display, where graphics will have to be rendered on screen in response to where someone is focusing their eyes, things will need to move an order of magnitude faster: from single to low double digit ms,” says Rabinovitsj. 

Image rendering will require edge computing. “We envision a future where remote rendering over edge cloud, or some form of hybrid between local and remote rendering, plays a greater role,” he adds. “Enabling remote rendering will require both fixed and mobile networks to be rearchitected to create compute resources at a continuum of distances to end users.”

Bandwidth could increase by orders of magnitude over what is required to view 720p video on a standard smartphone screen. That might work with just 1.3 Mbps to 1.6 Mbps of downlink throughput. 

But a head-mounted display sitting just centimeters from the eyes required to display images at retina grade resolution will need to be many orders of magnitude larger, he notes. 

To be sure, most of what happens that is part of metaverse experiences rests on things that happen up the stack from computing and communications. 

source: Constellation Research

But we already can see how metaverse support will require changes in computing architecture and network capabilities.

Digital Twin Applications Likely to be Commercial Before Many Metaverse Apps

Though consumer applications for metaverse seem to get more attention, digital twins are likely to produce more value, near term. 

A digital twin is a virtual representation of an object or system and is a type of “metaverse.”  It is updated from real-time data and uses simulation, machine learning and reasoning to help decision making, often to model processes and flows. 

That might be especially useful for complex systems, as when operators of such systems want to look at possible changes; predict breakdowns or predict maintenance activities. 

source: Leeway Hertz

When used to monitor and possibly avoid traffic congestion, digital twins might be useful to optimize traffic flows. Rerouting traffic around accident delays is one possible advantage, assuming the ability to manage traffic lights, open or close routes. 

In manufacturing, digital twins might be used to rapidly create prototypes or optimize supply chain support. In health care, assuming the knowledge base is refined enough, digital twins might help doctors evaluate outcomes from various treatments.  

Will Current AI Interest Produce Spring, or Another Winter?

Does our experience with mobile computing have relevance for our predictions about useful and commercial use cases for artificial intelligence? Possibly. 

Initially, mobile internet, for example, was sort of a “mobile version” of the desktop experience, allowing users to keep doing things they were doing at their desks. But the mobile experiences sometimes were "de-featured" versions of what could be done on desktop devices. Mobile email was a salient exception: mobile utility was higher and ease of use about the same.

The analogy is mobile voice, which allowed communications away from the desk or cordless phone. Lots of value was created.

Then we moved into an era where the mobile device was the preferred device for many use cases, including photo sharing and social media. 

The proliferation of mobile apps was a characteristic of that era, where hotel check-in, for example, was expected to be invoked on a mobile device. 

Now there are many use cases where mobile is itself the native environment, preferred to desktop use cases which do not work as well, if at all.  Turn-by-turn driving instructions or ride-hailing services provide examples. 

source: Medium 

So are there implications for adoption of practical AI? Maybe. 

The difference in the case of AI is that development has not been strictly linear. At least in terms of expectations, we have moved through alternating periods where expectations were greater or lesser. 

It might be quite fair to say that interest in the near-term value of artificial intelligence has moved through periods of inflated expectations before, even if we are at present in a period of high anticipation. 

source: Saegus 

Historically, interest in the immediate value of artificial intelligence has waxed and waned. 

source: Medium

AI Investment Hit an Inflection Point About 2020

Any way you look at it, private investment in artificial intelligence has blossomed. 

source: AI Index Report

AI Will Mostly be Encountered as a Feature of a Product, Not a Product as Such

Whether we encounter artificial intelligence as a user of apps, services or products, we increasingly will be using AI in the future. Very rarely, if ever, will we actually be buying “AI” as a product, though. It will simply be part of some other things we do buy. 

Which is to say we will normally encounter AI as “weak AI,”  some form of machine learning that is optimized for a particular purpose, such as speech recognition, search results, recommendations,  facial recognition or some customer service app. 

This form of AI will be wildly valuable and ubiquitous, early on. 

source: Toolbox 

What we will not often actually use or see is “strong AI” (“ deep AI”),  when machines apply their intelligence to solve complex problems. Still less are we at a stage when “artificial superintelligence” is possible, allowing machines to surpass human intelligence in terms of completing tasks. 

Weak AI will prove very useful as a way of detecting patterns in human or machine behavior that allow decisions to be made. When people use voice assistants such as Siri, they already are engaging in a process supported by weak AI. 

Almost no consumer or business user will ever buy “AI” in a direct sense, as a product. It will be a feature of a product. 

Verizon, Atos Partner on Computer Vision at the Edge

Verizon Business is adding the Atos computer vision to its multi-access edge computing fabric, aiming to support industrial internet of things applications that include the ability to do predictive analysis of industrial process disruptions. 

By 2025, by some estimates, computer vision will be the second-biggest use of artificial intelligence in the manufacturing industries. 

source: Fractovia 

The Atos computer vision platform will analyze 180 billion data points every hour. Using this system, the engineers and operators will be able to pinpoint exactly when and where operation downtime is predicted, up to 30 days in advance, the partners say. 

The Atos platform brings AI-powered video analytics to mission critical environments,  Verizon says. Verizon hopes the Atos BullSequana Edge servers, deployed in Verizon MEC facilities, will  strengthen its 5G edge offers and unlock new use cases. 

The solution is available to customers across a variety of industries including transport, industrials, logistics and manufacturing, Verizon says.  

The Atos Computer Vision Platform enables organizations to process and analyze massive amounts of complex video and image data in real-time to automatically monitor, manage and improve working practices, security and surveillance.

The Brain and Implications for AI

21st century artificial intelligence is dominated by deep learning. 

But one of the isues with deep learning is that it’s often completely data-driven. Prior knowledge is not incorporated. Adding prior knowlege reduces computation, but adds an element of judgment or subjectivity. 

And that will be an issue. We can go faster when we incorporate what we believe we already know. That's why we have "rules of thumb."

But humans process and perceive in ways that are not strictly based on "what exists in reality" but how we conclude things are in reality. We sometimes construct it, in other words. And humans are able to understand and sort through lots of abstractions we might just call "common sense."

Algorithms cannot do this unless they are taught. The "data" does not necessarily help. It's kind of how we had issues with machine vision. "Objects" humans easily recognized often were difficult for AI to perceive. We had to embed that knowledge in the algorithms. 

It all matters for applied AI since AI is about inferences. And inferences made by humans often involve all sorts of embedded rules that make inference generation easier and more accurate.