8 min

We need to keep talking about Artificial Intelligence (AI) on mobile devices and this conversation will continue throughout this year and beyond. Despite the fact that Mobile World Congress 2024 is behind us now, this is not a debate that stops once the crowds leave Barcelona. Perhaps now is a good time to welcome a deeper dive into how mobile AI works, where it will impact enterprise (and indeed consumer) application usage and how it might actually change the way we live our lives.

Andrea Mirabile, global director of AI research at Zebra Technologies thinks that enabling intelligent interactions and enhancing the user experience on mobile devices necessitates significant advancements in both hardware and software components of the ecosystem.

“One crucial aspect [in this space] is the integration point of AI with mobile Operating Systems (OS). Deeply embedding AI capabilities within a mobile OS like iOS and Android enables interaction between AI agents and system-level services. Additionally, standardising protocols and interfaces for AI agents to communicate with third-party services and apps is vital for smooth interactions. This standardisation may entail adopting industry-wide norms for data exchange, communication protocols and security mechanisms,” said Mirabile.

He reminds us that because AI agents have the potential to enjoy interactions with diverse services and apps, safeguarding user data privacy is paramount. Robust measures such as data encryption, anonymisation techniques and user consent mechanisms are crucial for protecting sensitive information. Furthermore, he says, mobile AI systems must adhere to stringent security protocols to prevent unauthorised access and malicious attacks. This includes measures such as secure authentication, secure communication channels and regular security updates.

The evolutionary path of devices

“In the past, mobile phones relied on physical keyboards for user input. Examples include early Nokia and Motorola phones, but Blackberry’s larger keyboards improved typing and quickly became popular among professionals for email communication and productivity tasks. Then the iPhone popularised touchscreen interfaces, leading to more immersive experiences with features like multi-touch and virtual keyboards,” explained Mirabile, pondering the changing role of what we used to simply call our ‘phones’ in the past.

“Building upon the advancements in touch screen interfaces, future mobile devices may focus on voice interactions, featuring sleek designs and navigation primarily through voice commands. These devices would be contextually aware, understanding the user’s preferences, habits and surroundings to provide personalised and relevant responses and rely heavily on virtual assistants for assisting users with tasks, providing information and coordinating interactions with other devices and services” he added.

Thankfully (or at least there are industry signals suggesting that we give thanks) the AI community and industry is increasingly prioritising the development of safety features in virtual assistants and ensuring adherence to ethical guidelines while respecting user privacy and mitigating potential risks such as misinformation or harmful suggestions.

Model architecture matters

Mobile devices have of course evolved significantly since the first-generation iPhone’s debut, but they still lack the computational prowess needed to fully harness today’s Large Language Models (LLMs). The key to unlocking mobile and edge AI’s potential lies not in sheer computational power, but in a strategic approach to model architecture, data management and leveraging a device’s native computing capabilities. This is the opinion of Rahul Pradhan, VP of product and strategy for AI, ML & data at Couchbase.

“True mobile AI cannot depend solely on cloud-based solutions. This isn’t just a matter of connectivity but of efficiency, speed and data privacy. AI that relies on transmitting data to a central server cannot achieve real-time responsiveness. Latency introduces delays that compromise AI-generated insights’ reliability, not to mention the bandwidth costs associated with constant data transmission,” said Pradhan. “Cloud servers are best suited to high-computational tasks, such as training deep learning models and serving LLMs. Conversely, tasks requiring immediate interaction between the AI and users, along with other machine learning processes, are more efficiently handled on the device, at the edge of the network. This approach enhances performance while ensuring user privacy by minimising data transmission.”

The Couchbase data guru further explains that reducing a device’s computational burden is another critical step. Techniques like model quantization, which simplifies AI models by reducing the precision of their calculations, are essential for maintaining performance without compromising functionality. Innovations such as GPTQ (a post-training quantization (PTQ) method for 4-bit quantization that focuses primarily on GPU inference and performance), which compresses models post-training; LoRA, which fine-tunes smaller matrices within a pre-trained model; and QLoRA, which optimises GPU memory usage for greater efficiency, represent options tailored to specific application needs.

Also read: Apple shows how LLMs can be run on smartphones

Privacy, security & data synchronisation

“Other key drivers and considerations for mobile AI are data privacy and security and data synchronisation. Implementing strong data encryption and privacy-preserving techniques ensures user data is protected, strengthening one of the primary advantages of processing data locally. Meanwhile, mechanisms to synchronize data between edge devices and the cloud or central servers will ensure data integrity and consistency across the network,” clarified Pradhan. 

He advocates a unified data platform capable of managing diverse data types and enabling AI models to access and interact with local data stores, both online and offline as a significant advantage. Spoiler alert: Couchbase likes to talk about unified data platform technology in relation to its ability to provide unified data platform technology, but hey, we all knew that right?

Pradhan evangelises around this approach and says that it not only improves performance but also enhances the user experience by ensuring that AI applications are responsive, reliable and capable of operating in a variety of environments.

Finally, the best architecture for mobile AI – and ultimately any AI – minimises complexity. The simpler the architecture, the more power it can dedicate to AI itself – especially important in a mobile environment.

Mobile AI safety concerns?

When considering mobile AI safety, the big area of concern is safety from hackers. Will we we get to the point of trusting the virtual assistants with our data, our movements, or even our deepest darkest thoughts. It’s much easier now to know when a scam caller is talking to you – you can tell by their tone of voice and mannerisms. In the future, when we are conversing with machines, they will literally all sound the same.

“People will assume they’re talking with a voice assistant, but it could be a hacker controlling it,” warns Hakim Akayour, head of revenue assurance at Sagacity. “This could have huge commercial impacts on an organisation if such a breach occurrs, so we are seeing companies pumping billions into security, as if that trust is broken it will be very hard to recover. However, nothing is bullet-proof. The various touchpoints and APIs that a virtual will need to interact with will create many loopholes that it will be near impossible to avoid.”

Following rising customer expectations

The Sagacity team suggest that that we have seen time and again that once a standard is set, customer expectations rise and everyone must follow. So, it stands to reason that voice and conversational style interactions will start to become commonplace in apps, with search and voice merging into one.

“Take Alexa as an example. If I need to get more batteries, I can just say ‘Alex order batteries’ today. I don’t specify what brand, which is why we have recently seen Amazon creating its own range of batteries which it will default to. Voice assistants use machine learning to identify what ‘brandless’ products – i.e. products where people just search for the item – are popular and can then ensure they are selling products through their own supply chains and add a load of markup,” added Akayour.

AWS: the view from the (mobile) cloud

For the telco industry, AWS views generative AI and machine learning playing a pivotal role in increasing operational efficiencies and business performance, enhancing the customer experience and driving innovative new offerings. Sameer Vuyyuru, director, business development for telecom industry at AWS says that like other industries, generative AI can help telcos enhance efficiencies in many functions: for example, populating RFPs, deploying chatbots to aid in sales and personalising marketing to individuals at scale. 

In fact, the AWS telco team expect to see huge growth in this area. They project that by 2026, 95% of telcos will deploy data, analytics and AI initiatives to enhance their customer experience and improve their product planning, up from 50% in 2022, according to magical analyst house Gartner. 

However, there are a few industry-specific applications AWS believes to be truly transformative. 

“Already, many telcos leverage AI to augment human interactions and improve the consistency of experience and resolution speed. Generative AI can take these activities one step further, with interactive voice response – an evolution of early chatbots deployments to help customers resolve issues or get answers to questions. In addition, generative AI can help analyse real-time call discussion to provide prompts and resources to agents to help resolve customer inquiries. Customer service agents will play a key role in the process, but we do believe generative AI can reinvent and improve every customer experience and application,” said Vuyyuru.

Continuing his strategic thread here, the AWS telco director suggests that generative AI can also play a key role in all the aspects of the network lifecycle. 

Mobile AI @ infrastructure level

“During installation of network elements, engineers rely on manuals and documented processes. Generative AI can ingest this data and provide interactive guidance and prompts to speed up and simplify installation tasks. Foundation models can also be trained on network topology and configuration data to suggest the configuration of network elements. When there is a network failure, generative AI-based applications can recommend troubleshooting actions and procedures to networking operating engineers,” notes Vuyyuru.

Lastly from AWS, the team say that generative AI can help telcos more easily identify areas where they are losing revenue or incurring revenue leakage. Deployed across business processes, generative AI can examine profits, revenues, various consumer plans, expenses and customer charges to provide recommendations on how to evolve offerings to optimise profits.

It’s immediately clear that if we ask: how does mobile AI work? Then the answer depends upon whether we’re asking the question from an infrastructure perspective, from a networking and integration perspective, from an application functionality angle or indeed from a user experience viewpoint. In most cases, we should be asking about all these factors if we want a truly robust and rounded answer. If we want to know which innovations are going to be impacting mobile AI by the end of the decade, then we won’t have to wait long either – just hold on please caller.

Image credit: Adrian Bridgwater