Midjourney, for instance, likely works by using a combination of a language model, which interprets the text prompts users write -
AI is here
But how do we make AI as trustworthy as it is transformative?
How do we apply AI at every scale?
And how do we make AI as personal as it is powerful?
Smarter AI for All
From individual to enterprise, here’s how to make AI work for you
There’s been an explosion of potential use cases for AI in recent years, down largely to the development of the large language model (LLM) in 2017.
Based on research that tries to model the human brain, the LLM is able to consider not just individual words, like predictive text does on your phone; but whole sentences, comparing the use of every word and phrase in a passage to other examples across the data it’s been trained on.
- large
- language
- models
- building
- sentences
The result is the new field of “generative AI”, with both startup and large organizations racing to build apps that access foundation models like GPT-4 or LaMDA via an API.
Apps like ChatGPT, or the image generating platform Midjourney, can now respond to prompts with text and images so sophisticated they’ve been compared to human answers.
trigger: 1
…and a diffusion model, a type of foundation model that creates a random noise – think of it like a paint splatter – and gradually transforms it into a coherent image.
First, the language model identifies key features or themes in the prompt from the concepts it’s been trained on.
So, for a prompt like “Imagine a portrait of a labradoodle, painted in the style of Van Gogh”, those features and themes would be “labradoodle” or “Van Gogh”.
These are essentially the “meaning” of that text, which the language model captures in a fixed-size vector called an “embedding”.
The diffusion model then evaluates the current state of the “paint splatter” (or “initial tensor”, to give it its technical term),
…then considers what the text embedding says, against the patterns or colors associated with “labradoodles” or “Van Gogh” in its training set.
Based on that information, the model then decides whether to add further layers of random dots, keep those that already develop the pattern, or discard those that don’t…
…until finally, after multiple iterations of this process, a likeness starts to emerge. An image that’s never been seen before - and certainly not painted by Van Gogh.
With so many iterations and processes working in parallel, generative AI technology such as this is computationally intensive, requiring a level of processing power that’s generally only possible on the public cloud. That, in turn, raises troubling ethical questions.
These models are often based on troves of public data, meaning it’s hard to know for sure that their results are fair and unbiased - or that our own personal data hasn’t made its way into the model. And because they’re hosted far away from their source, latency is increased, making it prohibitively expensive to process the amount of data AI requires at the speed with which users want it. So expensive, only large corporations can afford to develop and host LLMs on the cloud.
Clearly, a one-size-fits-all approach isn’t enough. If AI is going to be as secure, trustworthy and accessible as it’s transformative, we need to move beyond the cloud.
Hosting AI on the edge
For AI to be effective in many instances, it needs to process data closer to where the action actually happens - the mobile devices or workstations that create the data in the first place.
That's where edge computing comes in.
trigger: 1
An edge server is the same as a traditional server, but instead of being located in a traditional data center, it sits on the “edge” of a network.
That means it can perform compute, networking, storage and security functions close to where users need them
– for example, in a healthcare setting or manufacturing site.
Data is created on devices that are nearby, whether that’s something small like a glucose monitor, or something much larger, like an autonomous vehicle, or a stock management tool. It then has a much shorter distance to travel, meaning AI models can work faster, make faster decisions, and use less power.
trigger: 1
It’s one of only three islands in the world where a rare seabird called the pink-footed shearwater nests, but since humans introduced the racoon-like coati to the island years ago, they’ve been under threat.
The charity works with local people to monitor coatis, and remove them from the shearwaters’ nests.
Recently they installed 70 camera traps at nesting sites around the island to detect any new coatis.
The cameras generated thousands of photos per day, but the island’s satellite internet connection was too slow to send them over the network to the mainland to be analyzed.
Instead, volunteers had to collect the hard drives of each camera, wait for a plane that only arrived twice a month, and send them over 400 miles to Santiago.
As a result, the time between a camera detecting a coati, and the team finding out, could be as long as three months – more than enough for the coatis to re-establish themselves and damage the shearwaters’ nesting sites.
The charity worked with Lenovo to create a technology hub on the island, upgrading the internet connection and providing compact, rugged edge servers with Neptune cooling technology, so they can work in a remote, hot environment.
By being onsite, the servers can process data closer to where the action actually happens - the cameras and devices that create the data in the first place.
trigger: 1
So now when a monitor detects something at a nesting site, it triggers a camera, which starts taking shots
…and uploads the images to a powerful ThinkEdge SE450 edge server, with integrated GPUs for rapid AI inferencing
…where a machine learning model capable of processing 4.8 images per second, or 415,000 per day, does the first layer of AI analysis.
The model quickly identifies the images that are useful and actionable, filtering out all the ones that don’t contain a coati. This dramatically reduces the number of images needing further analysis by the team in Santiago.
In fact, the data payload is so small it can be sent via the internet, removing the need to ship hard drives out by plane.
The edge server also relays data back to the camera devices, alerting the conservation team on the island.
So if a coati is detected, they can spring into action fast and protect the nesting birds from predation.
The time it takes to discover new coatis has gone from three months to a matter of weeks – enough time for the team to take the steps that protect the birds and help increase their numbers.
Edge computing is a way of customizing AI, so a small enterprise such as Island Conservation can get the most out of its processing power.
trigger: 1
Not only can they conduct initial analysis quickly and cost-effectively
…but they can do it securely and privately too, because the data doesn’t enter the cloud.
Edge AI also transforms manufacturing and industry-scale operations, where rapid data collection and analysis can optimize logistics. With Lenovo’s Supply Chain Intelligence, for example, AI-driven analysis across 200 countries, 2,000 international suppliers and 300,000 materials sourced, yielded 60% faster decision-making and 20% cost reduction across Lenovo.
Some potential use cases of generative AI, however, need a level of localization that’s even higher.
Getting personal with on-device AI
AI can also be processed directly on a personal device, removing the need to connect to a server or cloud. Data doesn’t leave the device it’s created on. It’s processed there, meaning it can understand and adapt to the individual using it.
The result is a highly personalized AI, like the one helping Erin Taylor. The 24-year-old has recently been diagnosed with ALS, a neurodegenerative disorder that typically leads to full-body paralysis and loss of speech.
trigger: 1
When Erin moves her eyes in a specific direction,
…the eye-tracking hardware in front of her registers the movement,
…which controls a circular keyboard on the screen of a Lenovo ThinkPad facing her,
…leveraging a personalized, compressed large language model (LLM) hosted on the device to make suggestions for each word.
Erin then chooses the correct suggestion with another flick of her eye,
…until a whole sentence is composed.
Because on-device AI operates entirely offline, it can respond fast and accurately to Erin. This speed and efficiency means it’s perfect for powering wearable tech like fitness trackers or VR headsets. But for assistive technology, this kind of offline reliability is essential.
Computing on-device like this tackles issues like cost, energy use, reliability, latency and privacy – everything that makes scaling and growing a new technology difficult.
For best results, the devices themselves should be optimized for AI. Lenovo’s new range of AI PCs, for example, are powered by dedicated neural processors (NPUs) as well as GPUs. This makes them particularly adept at handling AI-related tasks like speech recognition, blurring the background in video calls or detecting objects while editing photos and videos.
But what on-device AI offers in speed, security and personalization, it lacks in space and storage, which limits its power and model training capabilities.
What’s needed is a more flexible approach.
Hybrid AI offers the best of all worlds
At Lenovo, we believe AI needs to be processed across the public cloud, on the edge and personal devices to achieve its full potential.
That’s why the Edge Company approached us recently with their anti-bird strike solution BCMS© VENTUR.
It’s estimated birds get caught in an airplane’s engines every 15 minutes around the world, 80% of which take place during landing and takeoff.
But with BCMS© VENTUR, airports could effectively deter birds from the vicinity of their runways and significantly reduce the number of collisions, thanks to AI-based data processing that detects birds in the area. Having been trained on more than 100,000 images of various bird species, the model can recognize birds with greater than 95% accuracy. The result is an innovative bird scaring solution able to detect five times as many birds as human observers.
AI analysis and operation at the edge is essential to such a time-sensitive task; but the deeper analysis and model evolution happens in the cloud with greater computing resources – this is hybrid AI in action.
But, just as importantly, thanks to the distributed IT infrastructure it uses, it’s also reliable, cost-efficient and manageable – meaning it’s easy to replicate in airports around the globe.
This hybrid infrastructure is a more efficient way to distribute generative AI workloads, saving time and energy. It also offers a seamless personalized experience to users, while at the same time protecting their privacy and data security.
trigger: 1
So while devices take care of smaller, easier processing,
…servers on the edge are freed up to manage the bigger, more powerful tasks like model training and inference.
And when access to public data sets is valuable – querying the internet for the latest news, say, or a particular product’s reviews –
…public models can supplement on-device and edge processing power.
Our AI-ready devices, infrastructure, solutions and service will help you implement and manage responsible AI solutions, from the individual to the enterprise. Our innovative Neptune data center cooling technology also cuts power consumption by 40%, helping businesses unlock the power they need for AI without forfeiting their sustainability goals.
It’s part of our commitment to responsible AI, or an AI that’s as inclusive, secure and sustainable as it’s powerful. We’ve also joined global agreements to develop AI the right way, and regularly evaluate training datasets for diversity, privacy, accountability, transparency and environmental impact.
